Ron Toland - Notes from Clojure/Conj 2017

It was a good Conj. Saw several friends and former co-workers of mine, heard some great talks about machine learning, and got some good ideas to take back to my current gig.

There were some dud talks, too, and others that promised one (awesome) thing and delivered another (boring) thing, but overall it’s inspiring to see how far Clojure has come in just ten years.

My notes from the conference:

DAY ONE

KEYNOTE FROM Rich Hickey: Effective Programs, or: 10 years of Clojure

clojure released 10 years ago
never thought more than 100 people would use it
clojure is opinionated
- few idioms, strongly supported
- born out of the pain of experience
had been programming for 18 years when wrote clojure, mostly in c++, java, and common lisp
almost every project used a database
was struck by two language designers that talked disparagingly of databases, said they’d never used them
situated programs
- run for long time
- data-driven
- have memory that usually resides in db
- have to handle the weirdness of reality (ex: “two-fer tuesday” for radio station scheduling)
- interact with other systems and humans
- leverage code written by others (libraries)
effective: producing the intended result
- prefers above the word “correctness”, none of his programs ever cared about correctness
but: what is computing about?
- making computers effective in the world
- computers are effective in the same way people are:
  - generate predictions from experience
  - enable good decisions
- experience => information => facts
programming is NOT about itself, or just algorithms
programs are dominated by information processing
but that’s not all: when we start talking to the database or other libraries, we need different protocols to talk to them
but there’s more! everything continues to mutate over time (db changes, requirements change, libraries change, etc)
we aspire to write general purpose languages, but will get very different results depending on your target (phone switches, device drivers, etc)
clojure written for information-driven situated programs
clojure design objectives
- create programs out of simpler stuff
- want a low cognitive load for the language
- a lisp you can use instead of java/c# (his common lisp programs were never allowed to run in production)
says classes and algebraic types are terrible for the information programming problem, claims there are no first-class names, and nothing is composable
in contrast to java’s multitude of classes and haskell’s multitude of types, clojure says “just use maps”
says pattern matching doesn’t scale, flowing type information between programs is a major source of coupling and brittleness
positional semantics (arg-lists) don’t scale, eventually you get a function with 17 args, and no one wants to use it
sees passing maps as args as a way to attach names to things, thinks it’s superior to positional args or typed args
“types are an anti-pattern for program maintenance”
using maps means you can deal with them on a need-to-know basis
things he left out deliberately:
- parochialism: data types
- “rdf got it right”, allows merging data from different sources, without regard for how the schemas differ
- “more elaborate your type system, more parochial the types”
- in clojure, namespace-qualified keys allow data merging without worrying about colliding schemas (should use the reverse-domain scheme, same as java, but not all clojure libraries do)
- another point: when data goes out over the wire, it’s simple: strings, vectors, maps. clojure aims to have you program the same inside as outside
smalltalk and common lisp: both languages that were designed by people for working programmers, and it shows
- surprisingly, the jvm has a similar sensibility (though java itself doesn’t)
also wanted to nail concurrency
- functional gets you 90% of the way there
pulled out the good parts of lisp
fixed the bad parts: not everything is a list, packages are bad, cons cell is mutable, lists were kind of functional, but not really
edn data model: values only, the heart of clojure, compatible with other languages, too
static types: basically disagrees with everything from the Joy of Types talk
spec: clojure is dynamic for good reasons, it’s not arbitrary, but if you want checking, it should be your choice, both to use it at all and where to use it

Learning Clojure and Clojurescript by Playing a Game

inspired by the gin rummy card game example in dr scheme for the scheme programming language
found the java.awt.Robot class, which can take screenshots and move the mouse, click things
decided to combine the two, build a robot that could play gin rummy
robot takes a screenshot, finds the cards, their edges, and which ones they are, then plays the game
lessons learned:
- java interop was great
when clojurescript came out, decided to rebuild it, but in the browser
robot still functions independently, but now takes screenshot of the browser-based game
built a third version with datomic as a db to store state, allowing two clients to play against each other
absolutely loves the “time travel” aspects of datomic
also loves pedestal

Bayesian Data Analysis in Clojure

using clojure for about two years
developed toolkit for doing bayesian statistics in clojure
why clojure?
- not as many existing tools ass julia or R
- but: easier to develop new libraries than in julia (and most certainly R)
- other stats languages like matlab and R don’t require much programming knowledge to get started, but harder to dev new tools in them
michaellindon/distributions
- open-source clojure lib for generating and working with probability distributions in clojure
- can also provide data and prior to get posterior distribution
- and do posterior-predictive distributions
wrote a way to generate random functions over a set of points (for trying to match noisy non-linear data)
was easy in clojure, because of lazy evaluation (can treat the function as defined over an infinite vector, and only pull out the values we need, without blowing up)
…insert lots of math that i couldn’t follow…

Building Machine Learning Models with Clojure and Cortex

came from a python background for machine learning
thinks there’s a good intersection between functional programming and machine learning
will show how to build a baby classification model in clojure
expert systems: dominant theory for AI through 2010s
- limitations: sometimes we don’t know the rules, and sometimes we don’t know how to teach a computer the rules (even if we can articulate them)
can think of the goal of machine learning as being to learn the function F that, when applied to a set of inputs, will produce the correct outputs (labels) for the data
power of neural nets: assume that they can make accurate approximations for a function of any dimensionality (using simpler pieces)
goal of neural nets is to learn the right coefficients or weights for each of the factors that affect the final label
deep learning: a “fat” neural net…basically a series of layers of perceptrons between the input and output
why clojure? we already have a lot of good tools in other languages for doing machine learning: tensorflow, caffe, theano, torch, deeplearning4j
functional composition: maps well to neural nets and perceptrons
also maps well to basically any ML pipeline: data loading, feature extraction, data shuffling, sampling, recursive feedback loop for building the model, etc
clojure really strong for data processing, which is a large part of each step of the ML pipeline
- ex: lazy sequences really help when processing batches of data multiple times
- can also do everything we need with just the built-in data structures
cortex: meant to be the theano of clojure
- basically: import everything from it, let it do the heavy lifting
backend: compute lib executes on both cpu and gpu
implements as much of neural nets as possible in pure clojure
meant to be highly transparent and highly customizable
cortex represents neural nets as DAG, just like tensorflow
- nodes, edges, buffers, streams
basically, a map of maps
- can go in at any time and see exactly what all the parameters are, for everything
basic steps to building model:
- load and process data (most time consuming step until you get to tuning the model)
- define minimal description of the model
- build full network from that description and train it on the model
for example: chose a credit card fraud dataset

Clojure: Scaling the Event Stream

director, programmer of his own company
recommends ccm-clj for cassandra-clojure interaction
expertise: high-availability streaming systems (for smallish clients)
systems he builds deal with “inconvenient” sized data for non-apple-sized clients
has own company: troy west, founded three years ago
one client: processes billions of emails, logs 10–100 events per email, multiple systems log in different formats, 5K–50K event/s
10–100 TB of data
originally, everything logged on disk for analysis after the fact
requirements: convert events into meaning, support ad-hoc querying, generate reports, do real-time analysis and alerting, and do it all without falling over at scale or losing uptime
early observations:
- each stream is a seq of immutable facts
- want to index the stream
- want to keep the original events
- want idempotent writes
- just transforming data
originally reached for java, since that’s the language he’s used to using
data
- in-flight: kafka
- compute over the data: storm (very powerful, might move in the direction of onyx later on)
- at-rest: cassandra (drives more business to his company than anything else)
kafka: partitioning really powerful tool for converting one large problem into many smaller problems
storm: makes it easy to spin up more workers to process individual pieces of your computation
cassandra: their source of truth
query planner, query optimizer: services written in clojure, instead of throwing elasticsearch at the problem
recommends: Designing Data-Intensive Applications, by Martin Kleppmann
thinks these applications are clojure’s killer app
core.async gave them fine-grained control of parallelism
recommends using pipeline-async as add-on tool
composeable channels are really powerful, lets you set off several parallel ops at once, as they return have another process combine their results and produce another channel
but: go easy on the hot sauce, gets very tempting to put it everywhere
instaparse lib critical to handling verification of email addresses
REPL DEMO
some numbers: 0 times static types would have saved the day, 27 repos, 2 team members

DAY TWO

The Power of Lacinia and Hystrix in Production

few questions:
- anyone tried to combine lysinia and hystrix?
- anyone played with lacinia?
- anyone used graphql?
- anyone used hystrix?
hystrix : circuit-breaker implementation
lacinia: walmart-labs’ graphql
why both?
simple example: ecommerce site, aldo shoes, came to his company wanting to rennovate the whole website
likes starting his implementations by designing the model/schema
in this case, products have categories, and categories have parent/child categories, etc
uses graphvis to write up his model designs
initial diagram renders it all into a clojure map
they have a tool called umlaut that they used to write a schema in a single language, then generate via instaparse representations in graphql, or clojure schema, etc
lacinia resolver: takes a graphql query and returns json result
lacinia ships with a react application called GraphiQL, that allows you to through the browser explore your db (via live queries, etc)
gives a lot of power to the front-end when you do this, lets them change their queries on the fly, without having to redo anything on the backend
problem: the images are huge, 3200x3200 px
need something smaller to send to users
add a new param to the schema: image-obj, holds width and height of the image
leave the old image attribute in place, so don’t break old queries
can then write new queries on the front-end for the new attribute, fetch only the size of image that you want
one thing he’s learned from marathon running (and stolen from the navy seals): “embrace the suck.” translation: the situation is bad. deal with it.
his suck: ran into problem where front-end engineers were sending queries that timed out against the back-end
root cause: front-end queries hitting backend that used 3rd-party services that took too long and broke
wrote a tiny latency simulator: added random extra time to round-trip against db
even with 100ms max, latency diagram showed ~6% of the requests (top-to-bottom) took over 500ms to finish
now tweak it a bit more: have two dependencies, and one of them has a severe slowdown
now latency could go up to MINUTES
initial response: reach for bumping the timeouts
time for hystrix: introduce a circuit breaker into the system, to protect the system as a whole when an individual piece goes down
hystrix has an official cloure wrapper (!)
provides a macro: defcommand, wrap it around functions that will call out to dependencies
if it detects a long timeout, in the future, it will fail immediately, rather than waiting
as part of the macro, can also specify a fallback-fn, to be called when the circuit breaker is tripped
adding that in, the latency diagram is completely different. performance stays fast under much higher load
failback strategies:
- fail fast
- fail silently
- send static content
- use cached content
- use stubbed content: infer the proper response, and send it back
- chained fallbacks: a little more advanced, like connecting multiple circuit breakers in a row, in case one fails, the other can take over
hystrix dashboard: displays info on every defcommand you’ve got, tracks health, etc
seven takeaways
- MUST embrace change in prod
- MUST embrace failure: things are going to break, you might as well prepare for it
- graphql is just part of the equation, if your resolvers get too complex, can introduce hystrix and push the complexity into other service
- monitor at the function level (via hystrix dashboard)
- adopt a consumer-driven mindset: the users have the money, don’t leave their money on the table by giving them a bad experience
- force yourself to think about fallbacks
- start thinking about the whole product: production issues LOOK to users like production features
question: do circuit-breakers introduce latency?
- answer: a little bit at the upper end, once it’s been tripped

The Tensors Must Flow

works at magento, lives in philly
really wants to be sure our future robot masters are written in clojure, not python
guildsman: tensorflow library for clojure
tensorflow: ML lib from google, recently got a c api so other languages can call into it
spoiler alert: don’t get TOO excited. everything’s still a bit of a mess
but it DOES work, promise
note on architecture: the python client (from google) has access to a “cheater” api that isn’t part of the open c api. thus, there’s some things it can do that guildsman can’t because the api isn’t there
also: ye gods, there’s a lot of python in the python client. harder to port everything over to guildsman than he thought
very recently, tensorflow started shipping with a java layer built on top of a c++ lib (via jni), which itself sits on top of the tensorflow c api, some people have started building on top of that
but not guildsman: it sits diretly on the c api
in guildsman: put together a plan, then build it, and execute it
functions like guildsman/add produce plan maps, instead of executing things themselves
simple example: adding two numbers: just one line in guildsman
another simple example: have machine learn to solve | x - 2.0 | by finding the value of x that minimizes it
tensorflow gives you the tools to find minima/maxima: gradient descent, etc
gradient gap: guildsman can use either the clojure gradients, or the c++ ones, but not both at once
- needs help to port the c++ ones over to clojure (please!)
“python occupies the 9th circle of dependency hell”: been using python lightly for years, and still has problems getting dependencies resolved (took a left turn at the virtual environment, started looking for my oculus rift)
demo: using mnist dataset, try to learn to recognize handwritten characters

The Dawn of Lisp: How to Write Eval and Apply in Clojure

educator, started using scheme in 1994, picked up clojure in 2009
origins of lisp: john mccarthy’s paper: recursive functions of symbolic expressions and their computation by machine, part i
implementation of the ideas of alonzo church, from his book “the calculi of lambda-conversion”
“you can always tell the lisp programmers, they have pockets full of punch cards with lots of closing parenthses on them”
steve russel (one of the creators of spaceware) was the first to actually implement the description from mccarthy’s paper
1962: lisp 1.5 programmer’s manual, included section on how to define lisp in terms of itself (section 1.6: a universal lisp function)
alan kay described this definition (of lisp in terms of lisp) as the maxwell equations of software
how eval and apply work in clojure:
- eval: send it a quoted list (data structure, which is also lisp), eval will produce the result from evaluating that list
  - ex: (eval '(+ 2 2)) => 4
- apply: takes a function and a quoted list, applies that function to the list, then returns the result
  - ex: (apply + '(2 2)) => 4
rules for converting the lisp 1.5 spec to clojure
- convert all m-expression to s-expressions
- keep the definitions as close to original as possible
- drop the use of dotted pairs
- give all global identifiers a ‘$’ prefix (not really the way clojure says it should be used, but helps the conversion)
- add whitespace for legibility
m-expressions vs s-expressions:
- F[1;2;3] becomes (F 1 2 3)
- [X < 0 -> -X; T -> X] becomes (COND ((< X 0) (- X)) (T X))
dotted pairs
- basically (CONS (QUOTE A) (QUOTE B))
definitions: $T -> true, $F -> false, $NIL, $cons, $atom, $eq, $null, $car, $cdr, $caar, $cdar, $caddr, $cadar
- note: anything that cannot be divided is an atom, no relation to clojure atoms
- last few: various combos of car and cdr for convenience
elaborate definitions:
- $cond: own version of cond to keep syntax close to the original
- $pairlis: accepts three args: two lists, and a list of existing pairs, combines the first two lists pairwise, and combines with the existing paired list
- $assoc: lets you pull key-value pair out of an association list (list of paired lists)
- $evcon: takes list of paired conditions and expressions, plus a context, will return the result of the expression for the first condition that evaluates to true
- $evlist: takes list of expressions, with a condition, and a context, and then evalutes the result of the condition + the expression in a single list
- $apply
- $eval
live code demo

INVITED TALK FROM GUY STEELE: It’s time for a New Old Language

“the most popular programming language in computer science”
no compiler, but lots of cross-translations
would say the name of the language, but doesn’t seem to have one
so: CSM (computer science metanotation)
has built-in datatypes, expressions, etc
it’s beautiful, but it’s getting messed up!
walk-throughs of examples, how to read it (drawn from recent ACM papers)
“isn’t it odd that language theorists wanting to talk about types, do it in an untyped language?”
wrote toy compiler to turn latex expressions of CSM from emacs buffer into prolog code, proved it can run (to do type checking)
inference rules: Gentzen Notation (notation for “natural deduction”)
BNF: can trace it all the way back to 4th century BCE, with Panini’s sanskrit grammar
regexes: took thirty years to settle on a notation (51–81), but basically hasn’t changed since 1981!
final form of BNF: not set down till 1996, though based on a paper from 1977
but even then, variants persist and continue to be used (especially in books)
variants haven’t been a problem, because they common pieces are easy enough to identify and read
modern BNF in current papers is very similar to classic BNF, but with 2 changes to make it more concise:
- use single letters instead of meaningful phrases
- use bar to indicate repetition instead of … or *
substitution notation: started with Church, has evolved and diversified over time
current favorite: e[v/x] to represent “replace x with v in e”
number in live use has continued to increase over time, instead of variants rising and falling (!)
bigger problem: some sub variants are also being used to mean function/map update, which is a completely different thing
theory: these changes are being driven by the page limits for computer science journals (all papers must fit within 10 years)
overline notation (dots and parentheses, used for grouping): can go back to 1484, when chuquet used underline to indicate grouping
- 1702: leibnitz switched from overlines to parentheses for grouping, to help typesetters publishing his books
three notations duking it out for 300 years!
vectors: notation -> goes back to 1813, and jean-robert argand (for graphing complex numbers)
nested overline notation leads to confusion: how do we know how to expand the expressions that are nested?
one solution: use an escape from the defaults, when needed, like backtick and tilde notation in clojure
conclusions:
- CMS is a completely valid language
- should be a subject of study
- has issues, but those can be fixed
- would like to see a formal theory of the language, along with tooling for developing in it, checking it, etc
- thinks there are opportunities for expressing parallelism in it

Day Three

Declarative Deep Learning in Clojure

starts with explanation of human cognition and memory
at-your-desk memory vs in-the-hammock memory
limitation of neural networks: once trained for a task, it can’t be retrained to another without losing the first
- if you train a NN to recognize cats in photos, you can’t then ask it to analyze a time series
ART architecture: uses two layers, F1 and F2, the first to handle data that has been seen before, the second to “learn” on data that hasn’t been encountered before
LSTM-cell processing:
- what should we forget?
- what’s new that we care about?
- what part of our updated state should we pass on?
dealing with the builder pattern in java: more declarative than sending a set of ordered args to a constructor
his lib allows keyword args to be passed in to the builder function, don’t have to worry about ordering or anything
by default, all functions produce a data structure that evaluates to a d4j object
live demos (but using pre-trained models, no live training, just evaluation)
what’s left?
- graphs
- front end
- kafka support
- reinforcement learning

Learning Clojure Through Logo

disclaimer: personal views, not views of employers
logo: language to control a turtle, with a pen that can be up (no lines) or down (draws lines as it moves)
…technical difficulties, please stand by…
live demo of clojure/script version in the browser
turns out the logo is a lisp (!): function call is always in first position, give it all args, etc
even scratch is basically lisp-like
irony: we’re using lisp to teach kids how to program, but then they go off to work in the world of curly braces and semicolons
clojure-turtle lib: open-source implementation of the logo commands in clojure
more live demos
recommends reading seymour papert’s book: “Mindstorms: Children, Computers, and Powerful Ideas”
think clojure (with the power of clojurescript) is the best learning language
have a tutorial that introduces the turtle, logo syntax, moving the turtle, etc
slowly introduces more and more clojure-like syntax, function definitions, etc
fairly powerful environment: can add own buttons for repeatable steps, can add animations, etc
everything’s in the browser, so no tools to download, nothing to mess with
“explaining too early can hurt”: want to start with as few primitives as possible, make the intro slow
can create your own lessons in markdown files, can just append the url to the markdown file and it’ll load (!)
prefer that you send in the lessons to them, so they can put them in the lessons index for everyone to benefit
have even translated the commands over to multiple languages, so you don’t have to learn english before learning programming (!)
lib: cban, which has translations of clojure core, can be used to offer translations of your lib code into something other than english
clojurescript repls: Klipse (replaces all clojure code in your page with an interactive repl)
comments/suggestions/contributions welcome