The Framework for ML Governance

The Framework for ML Governance by Kyle Gallatin

ML is becoming commoditized, but there hasn’t been much in the way of scalable frameworks to support the delivery and operation of models in production (of course, Algorithmia has one, hence their sponsoring of this report).

ML needs to follow the same governance standards for software & data as other more traditional software engineering does in your organization. In the development phase, this includes validation and reproducibility of your model, and documentation of your methods and rationale. The delivery and operational phases are much more complex, including things like observability, auditing, cost visibility, versioning, alerting, model cataloging, security, compliance, and much more.

MLOps is the set of best practices and tools that allow you to deliver ML at scale. ML governance is how you manage those practices and tools, democratizing ML across your organization through nonfunctional requirements like observability, auditability, and security. As ML companies mature, neither of these are “features” or “nice-to-haves”—they are hard requirements that are critical to an ML strategy.

Data scientists can’t do this all by themselves. They need support from the larger organization. The value of ML governance needs to be understood at the highest levels. They should involve infrastructure, security, and other domain experts early on, and be sure to be open in their communication with the rest of the company.

Advent of Code 2020: Swift

Advent of Code was a fun diversion this year. It seemed a bit easier than some previous years. Perhaps it helped that I picked a language that I had some experience with, Swift. Here are some quick thoughts on Swift as an AoC language, some general AoC tips, and notes on some of the more interesting puzzles.

You can find all my 2020 code here.

Swift for AoC

First, the complaints:

Swift is too pedantic about strings. AoC usually involves loads of string manipulation and parsing, so being able to easily index into strings is super helpful. Swift does not make this easy out-of-the-box. I understand why. Strings are complicated in a language like Swift that supports Unicode well. Indexing into an array is expected to be a constant-time operation, but that’s not possible when you have to look at the preceding bytes in order to figure out what the _i_th character is. Still. You can at least make it easier and less ugly than myString[myString.index(myString.startIndex, offsetBy: i)]. Thankfully Swift is easy to monkey-patch, so I added some useful extensions to make it tolerable.

Regexes are similarly annoying in Swift. Here it just feels like they never properly ported them over from Objective-C. Here’s an example:

// print the animals found in a string
let myString = "The quick brown fox jumped over the lazy dogs."
let animalsRegex = try NSRegularExpression(pattern: #"(fox|dog)"#, options: [])
let matches = animalsRegex.matches(in: myString, options: [], range: NSRange(myString.startIndex..<myString.endIndex, in: myString))
for match in matches {
    for i in 1..<match.numberOfRanges {
        print(NSString(string: myString).substring(with: match.range(at: i)))

My colleague Mike Simons gave me a RegEx helper he uses, which reduces the above to:

let animalsRegex = try RegEx(pattern: #"(fox|dog)"#)
let matches = animalsRegex.matchGroups(in: myString)
for match in matches {
    for group in match {

Tuples not being hashable or even extendable is a pain for a lot of AoC problems. I resorted to using a custom struct for a bunch of problems, but at some point discovered SIMD vectors, which have the added benefit of having arithmetic operators defined for them.

What I loved about Swift:

Working with a statically-typed language in a proper IDE (Xcode) is such a huge time-saver. Everything is checked as you type (which can be a little annoying when Xcode starts scolding you for not using variables that you’ve defined before you get a chance to write the code that uses them; but you learn to ignore it). You have a full-featured debugger.

Playgrounds are really useful for trying out quick snippets of code.

Swift is generally a pleasant, straight-forward language to write. There’s not all the overhead of thinking about memory and ownership like there is in Rust, for example.

It’s actually pretty performant—at least when not running the debug build (I usually got a 10x speed-up when building for release).

General AoC Tips

  • Hash maps, hash maps, hash maps. Even if the problem seems array-shaped, there’s a good chance a hash map will be more efficient.
  • Many of the problems are grounded in some basic CS principle (day 5 this year being a classic example). Think about what that might be before diving in.
  • Oftentimes part 2 will be part 1 but MORE. Days 15 (Rambunctious Recitation) and 23 (Crab Cups) this year were good examples of that. When coding part 1, think about how your solution might scale if you were doing a million times more, or with a million times more data.

The Puzzles

Ok, on to specific puzzles. I’m just going to talk about some of the more interesting ones. Spoilers follow:

Day 5: Binary Boarding (code)

This is a classic AoC problem in that it disguises basic CS principles. If you were to solve this naively, it might be kind of annoying. but when you realize that the seating strings (e.g. FBFBBFFRLR) are just binary numbers, with (F,L) -> 0 and (B,R) -> 1, it becomes trivial. Particularly helpful here was Swift’s ability to convert a binary string to decimal with Int(binary, radix: 2).

Part 1 just required sorting the list of converted seat numbers and taking the last (highest) one. Part 2, finding the first non-consecutive id, is probably faster just iterating through the sorted list and stopping on the first id that is not one more than the last id (which is what I did originally), but it was more fun and concise to use Gauss’ formula to calculate what the total should be if all the consecutive numbers were present and subtract the actual sum of all numbers.

For kicks, I did a “no code” version in a spreadsheet.

Day 7: Handy Haversacks (code)

This was the first day that required some real thought. I started down the path of having a bidirectional graph with each bag having parents and children, but I got bogged down. I think sometimes creating objects complicates things vs. just slinging around string hash maps.

Day 9: Encoding Error (code)

This is totally an interview question I would have failed. It’s easy enough doing it the “dumb” (O(n^2)) way that I did it, constructing a lookup table of all sums in the preamble, or, for part 2, iterating through every possible sequence of two consecutive numbers, then three, etc.—also O(n^2). After I finished, a friend shared the “inchworm” technique, which is O(n). Neat! I totally failed that interview.

Day 10: Adapter Array (code)

Part 2 was fun to figure out. There were very different ways to approach this. I went with recognizing that we were dealing with power sets. I wrote up my thinking in my code, so I won’t duplicate it here. Some friends solved this using dynamic programming, which I definitely need to level up on.

Day 11: Seating System (code)

Nothing particularly interesting here—the first of three Game of Life-type problems. I just had to share this, though.

Day 12: Rain Risk (code)

Pretty straightforward. Part 2 was fun because I [re-]learned how to rotate a vector.

Day 13: Shuttle Search (code)

This was a fun one. I solved part 1 in bed on my phone with a calculator, and then spent the entire next day working on part 2. I learned about the Extended Euclidean algorithm and the Chinese remainder theorem. I used the existence construction to solve it, but the simpler solution that many used was sieving.

Day 17: Conway Cubes (code)

Game of Life again! Doing it in 3D and then 4D was a fun twist. Nothing particularly tricky. I just have a soft spot for GoL.

Day 18: Operation Order (code)

This was a fun one. I spent a lot of time scratching my head and drawing diagrams, until I remembered it’s a lot easier to parse expressions in postfix, and you can use the Shunting yard algorithm to convert infix to postfix.

Day 19: Monster Messages (code)

Woo, boy! This was a good one. I’d never written a parser before, so this was a learning experience. My part 1 solution involved creating a matcher from the grammar, and it was fast and elegant. Part 2 threw in the wrench of recursive rules, which absolutely do not work when building a matcher. I tried a whole bunch of things like adding recursion limits, but I could not get it to work. I ended up moving on and came back to it only after I’d finished all the other days. By then I had heard the term CYK algorithm bandied about, and so I did a complete rewrite using that. I spent a bunch of time reading about the algorithm and about Chomsky normal form. I “cheated” a bit and didn’t write code to convert the provided rules to Chomsky normal form, as there were only a few rules that needed to be tweaked and it was easy enough to do it by hand:

Original Chomsky normal form
107: 18 | 47 107: "b" | "a"
11: 42 31 | 42 11 31 11: 42 31 | 42 133
133: 11 31
8: 42 | 42 8 8: 47 50 | 18 4 | 42 8

Day 20: Jurassic Jigsaw (code)

This was a favorite. Part 1 was easy once I figured out a way to generate a hash code that uniquely identified a side, irrespective of its orientation:

static func sideToInt(_ side:String) -> Int {
    let binary = side.replacingOccurrences(of: ".", with: "0")
                     .replacingOccurrences(of: "#", with: "1")
    let num = Int(binary, radix: 2)!
    let numReversed = Int(String(binary.reversed()), radix: 2)!
    return num * numReversed * (num ^ numReversed)

Once you have that, your corners are the tiles that have two “singleton” sides (i.e. unique to that tile). No assembling necessary…for part 1. Part 2 did require assembling, and that was a bit tedious. It was a cool problem, though.

Day 23: Crab Cups (code)

Classic AoC problem where the simple solution for part 1 utterly fails in part 2. I actually got the answer to part 2 after letting my dumb-but-optimized part 1 solution run overnight, but while it was running, mulling it over in bed, I realized that this was a linked list problem. I fixed it in the morning and the runtime went from ~ 7 hours to 0.2 sec.

Day 24: Lobby Layout (code)

This was surprisingly easy for this late in the calendar, although my first naive attempt trying to map it to a 2D coordinate system failed. But I googled “hexagonal grid coordinate system” and found this cool page. I used the cube coordinate system and it was easy-peasy.

Part 2: Yay, more Game of Life! I’m sorely tempted to write a visualization for this.


I scratched an itch last week and actually wrote some code (these days my primary IDE is Google Docs). Check out humanize, which replaces any large numbers in text piped into it with shortened versions using SI prefix symbols (K, M, G, T, etc.). So “File size: 972917401236” becomes “File size: 973G”.

This came about because I was debugging an issue with a Parquet file using parquet-tools, whose output looks like:

% parquet-tools meta part-01373-r-01373.gz.parquet | grep "row group"
row group 1:           RC:100 TS:121513011 OFFSET:4 
row group 2:           RC:100 TS:44877183 OFFSET:14293218 
row group 3:           RC:111 TS:28645460 OFFSET:19704281 
row group 4:           RC:168 TS:31794536 OFFSET:23387243 
row group 5:           RC:363 TS:28078938 OFFSET:27531686 
row group 6:           RC:15547 TS:899825873 OFFSET:31242035 
row group 7:           RC:9963 TS:47930700 OFFSET:142353955 
row group 8:           RC:20100 TS:911417774 OFFSET:155302131 
row group 9:           RC:20100 TS:889810662 OFFSET:269749697 
row group 10:          RC:10100 TS:855908709 OFFSET:382090935 
row group 11:          RC:7282 TS:1532744480 OFFSET:486019001 

Tired of squinting and counting digits to see how big some of these numbers were, I figured there’d be some easy bash/sed way to convert these numbers into something more readable. It is pretty easy to do this for isolated numbers using numfmt:

% echo "1234567890" | numfmt --to=si --round=nearest

But that doesn’t work on numbers embedded in text:

% echo "This is a big number: 1234567890" | numfmt --to=si --round=nearest
numfmt: invalid number: ‘This’

GNU sed can do a replacement with the result of a function, but it’s rather ugly:

% echo "This is a big number: 1234567890" | sed -re 's/[0-9]+/`numfmt --to=si --round=nearest &`/g;s/.*/echo &/;e'
This is a big number: 1.2G

But I’m doing this work on a Mac, which does not come with GNU sed (or numfmt, for that matter, although you can brew install coreutils to get gnumfmt). I started working on a perl one-liner to do it, but decided this was something that needed to exist on its own, and should be easy for anyone to install.

I chose to write it in Go because it has an easy way for people to install binaries with go get. (Rust has something similar with cargo install, but Go is more common in my workplace.) The problem with go get is that you have to run it again to upgrade, so for my fellow Mac users I added a Homebrew tap.

Long story short:

% parquet-tools meta part-01373-r-01373.gz.parquet | grep "row group" | humanize -binary
row group 1:           RC:100 TS:116Mi OFFSET:4 
row group 2:           RC:100 TS:43Mi OFFSET:14Mi 
row group 3:           RC:111 TS:27Mi OFFSET:19Mi 
row group 4:           RC:168 TS:30Mi OFFSET:22Mi 
row group 5:           RC:363 TS:27Mi OFFSET:26Mi 
row group 6:           RC:15Ki TS:858Mi OFFSET:30Mi 
row group 7:           RC:10Ki TS:46Mi OFFSET:136Mi 
row group 8:           RC:20Ki TS:869Mi OFFSET:148Mi 
row group 9:           RC:20Ki TS:849Mi OFFSET:257Mi 
row group 10:          RC:10Ki TS:816Mi OFFSET:364Mi 
row group 11:          RC:7Ki TS:1Gi OFFSET:464Mi


I wrote a Slack bot, ReadBot, to both scratch an itch and check out the current state of Slack integration. As a bonus, I got to spend some time getting familiar with Glitch, a pretty cool live coding platform.

ReadBot lets you connect your Goodreads and Pocket accounts to Slack, and then adding a bookmark (🔖) reactji to a message with an Amazon book link in it will add that book to your Goodreads “to-read” shelf, or any other URL to your Pocket account.

The Slack API and documentation has improved immensely since I last checked it out a couple years ago. I was clearly not the only one to get confused by profusion of integration options (bot? slash command? webhook?) , so they’ve improved their API landing pages to help guide you. There are also plenty of tutorials to get you started.

One of the fears when diving into an API integration is whether that API—or even integrations in general—are going to be supported in the future. Many developers have been burned by neglected, deprecated or even shutdown APIs. Slack, however, seems to be encouraging integrations whole-heartedly with its app directory.

One challenge of creating a new Slack integration has been Slack’s requirement that it be served over SSL. While Let’s Encrypt has made this a lot easier, setting up an SSL-enabled server is still a significant hurdle to clear just to start developing your integration. This is where Glitch comes in.

Glitch is a collaborative live coding environment made by Fog Creek Software. It’s super easy to get started, especially since you can find existing public projects on the site and “remix” them to make them your own. And every project has its own SSL-enabled hostname, right out of the box.

Glitch apps are Node.js-based (they seem to allow Python apps as well, but support is currently unofficial). The app restarts itself on the fly as you edit (although this can be disabled), so your app is always live. Autocomplete package search makes installing new packages a breeze. You can do pretty much everything you need to through the UI, but if you need, you can drop into a shell and do whatever you like to your instance.

There’s no built-in revision/commit history, but you can export your project to GitHub (as well as import from GitHub). Since your app is just a regular Node app, it’s perfectly portable to any other environment.

One general issue with writing Slack bots is that in order to continue working on your app after it is in production, you need to have separate development apps and endpoints. It would be nice if Slack made this a little easier, perhaps by allowing you to specify separate production and development endpoints, slash commands, etc.

I have no plans on publishing ReadBot to the Slack App Directory because I don’t want to have to maintain another production service. If you’re interested in doing so, go for it.

Rescuing Windows with Hammerspoon

I recently got a new monitor and it came with a problem. For some reason when I attached or detached my laptop, many of my windows would end up off the edge of the screen, and I’d have to drag them back over one by one. It was annoying enough that I decided I’d write an app to “rescue” them.

I figured it wouldn’t be too hard. Get a list of active windows, see which are off-screen, and move them over. But it was hard. My code only half-worked. And for some reason I couldn’t get Chrome windows to move.

I started looking around and found Swindler. It seemed to be just what I needed and confirmed that I wasn’t completely incompetent:

Writing window managers for macOS is hard. There are a lot of systemic challenges, including limited and poorly-documented APIs. All window managers on macOS must use the C-based accessibility APIs, which are difficult to use and are surprisingly buggy themselves.

But then I saw Hammerspoon listed at the bottom of the Swindler README, under Related Projects. Intrigued, I checked it out. It was just what I needed.

Hammerspoon lets you write Lua scripts to interact with the MacOS system, and it’s pretty powerful. The Getting Started Guide gives a good sense of some of the things you can do with it.

I’d never written Lua before, but with plenty of examples to crib from and good API documentation, it didn’t take me long to come up with a little script that did just what I needed:

I have that file saved at ~/.hammerspoon/rescuewindows.lua, and in ~/.hammerspoon/init.lua I have:

local rescueWindows = require "rescuewindows"
hs.hotkey.bind({"cmd", "alt", "ctrl"}, "R", rescueWindows)

So when it hit ⌃⌥⌘R, all my off-screen windows slide over to my main screen. Nice!