For those interested, I gave a talk on Athena at PyData UK 2018 called "Python at Massive Scale". 4500 developers making 20k commits a week. Codebase with 35m LOC.
The video is here: https://www.youtube.com/watch?v=ZYD9yyMh9Hk
It covers Athena's origins, what it is used for, application architecture, infrastructure, dev tooling and culture.
I understand the value of it, but as an experienced scientific developer and python dev, the culture shock was huge. I don't have fond memory working on it. It is very different from traditional programming environment, closer to a kind of reimplementation of smalltalk env w/ python. I believe an influence was actually an old system that used smalltalk in the 90ies at JPM.
One of the fancy thing was integrating reactive programming, which worked through ugly hacks, at least at that time, by parsing code to detect dependencies. IIRC, it could manage list comprehension but not loops. They also had their own python binary w/ both python2 and python3 in one process.
I remember my first few months unlearning normal Python and figuring out how to build the 'pixie graph', a lazily-evaluated Python DAG suited for calculating financial instruments. It took a while to get your head around this, but when you did it was a very powerful and productive way to building trading and risk management applications.
To get some sense in how this worked, here are two public projects on github with good introductions:
JPM's actual pixie code is proprietary, extremely performant after 12 years of pushing into bigger and bigger scale problems, and is definitely not on GitHub!
I've not heard the term "twiddle scopes" before; in Athena it is "tweaking".
For finance applications (risk) the whole concept works quite well, doesn’t it - combining the natural advantages of code and Excel.
[Disclosure: I was part of the Quartz Core team in 2011/2012.]
I suspect the Common Lisp influence was beneficial.
The weirdest thing, language-wise, that I noticed was scoping. It's neither dynamic nor static scoping, but weird scoping. (But there are work-around to get something like static scoping.)
Outside of the language, the whole CVS-based version control and review process was weird. But understandable as a product of the late 1990s, when review-before-going-into-permanent-history must have been way ahead of its time.
Not the spaces in variable names?
But I was used to different and exotic conventions from the obscure languages I played with over the years.
It was ubiquitous in GS by the mid-2000s, and was rather instrumental to GS navigating the crisis in 2008.
Being able to accurately and quickly compute risks across the entire firm's books, rather than manually merging across separate systems was key advantage, which encouraged JP, Citi and BofA to build a SecDB clone themselves.
https://www.goldmansachs.com/our-firm/history/moments/1993-s...
This would have had me running screaming from the building
AAAAaaaaaa
Imagine the benefit to the business had they just stuck with Smalltalk the whole time! It is remarkable how many businesses could have been quietly making money for years using a high-quality dynamic language such as Smalltalk or Lisp rather than going through the whole C/C++/Java/Python/Ruby/JavaScript/Go treadmill.
But I guess if they had done that then their programmers might not have gotten to type as many {}! Nor gotten to reinvent the wheel so often.
And Kapital was an interesting experiment but rightfully died. Smalltalk just wasn't the right environment for something that complex and scale. I'd expect most lisps to suffer in the same way. And TBH it was more the image based model rather than the dynamic language per se though that was an issue.
There's a point where dynamism becomes a liability and Kapital went way beyond that point. Python fares a bit better but still, a huge effort is needed to control its failings at scale. This isn't helped at all by sticking a DAG in the middle of it as was done with Athena and its ilk. This all can be made to work but in the "cos its Turing complete" sense rather than the language helping in any meaningful way.
Source: I suffered through it.
I had to LOL. You have very nicely put into words an experience I bet many have had.
Why? Eg Scheme or even the uglier Common Lisp are quite disciplined. And Clojure ain't too bad either.
As far as I can tell, Lispers don't usually go around monkey-patching things.
(I realize Smalltalk is great, but my point is it has its issues as well otherwise I recon sooner or later the advantages would be so clear that Amazon or Google or someone would be all over it since it would give them a competitive edge ...)
That was the official story anyway. A lot of things were claimed when I worked there, a lot of it BS.
The culture in banking was definitely not for me: generally fairly smart people, but very perverse culture. This was the only environment I ever worked where people who actually lie to my face to make sure I made mistake when coding and fail. Everything was custom, and sometimes that meant athena, sometimes that meant terrible systems such as a "distributed database" that was a single threaded wrapper on top of an C++ STL hashmap running on NFS. The guy behind it was reworking its own protocol based on UDP instead of TCP because said system was too slow...
There was 0 abstraction, so if you wanted something such as EUR/USD pair, you had to ask the one guy who knew which id it was, so that you could get the data from in code. To this day I am convinced this was done on purpose for job security.
That system you're thinking of was Kapital.
http://www.cincomsmalltalk.com/main/successes/financial-serv...
It's the only Smalltalk usage at JPM I believe, they could have doubled-down on it but instead went to Python and as you say, tried reinventing the wheel in it.
That's not reinventing the wheel, that's building a new wheel you hope performs better/different with different materials and maybe some different techniques, and has a long and storied history of both successes and failures.
Reinventing the wheel would be if they published this system and some other bank recreated it without using what they published, thus "inventing" something that already existed and was available.
If it was like the predecessor system I worked with, it could handle either. But in a list comprehension, it could figure out the individual dependencies of the individual list items.
So, suppose you had a a function that took in a list, and returned a list with some computations performed on each element. If done with list comprehension, and one input element changed, only the corresponding output element would need to be recomputed, and only that would be recomputed.
If done with a loop, and one input element changed, the entire output array would be recomputed (even if all but one element remained unchanged - it would be hard to guarantee that property by code inspection at the time the dependency graph is built).
My first thought was Athena as well, but it could be Quartz (it started in 08-09 at BaML, vs Athena starting in 2006).
Both Citi and Barcap also tried (albeit to less success than JPM/BaML) to do similar things.
With that said, For those unfamiliar - at least two major banks run insanely large production python installs across thousands of developers. Core skills (nowadays) are python and react - and yes, we’re all actively hiring :)
All of the large banks also have pretty good open source initiatives - just check out github repos for JPMorgan, BaML, or GS. We have also been active participants in Pycon (key sponsors and doing sessions, like Steve’s) since at least 2009.
I joined JPMorgan in 2010, as the first bank in London to offer me a Python role. I ended up staying for 8 years, almost all of it working in Athena: Commodities and FX trading; a bit of Equities; then with the Athena Core team working on the machine learning environment.
I learned a ton, worked with many hugely impressive people - on both the trading and technology sides of the business - and left with lots of good memories.
It is not a far-fetched argument that institutes like that are only a burden to society, existing only to widen socio-economical gaps. The point of view of an insider would be very interesting I think.
[0] "There was one remarkable legacy system when I was working in a bank sometime back."
https://www.techrepublic.com/article/jpmorgans-athena-has-35...
The update isn't a job for one person, and I don't think it is finished yet, unless it is ahead of this alleged schedule...
Thankfully, both JPM and BAML seem to have started to ditch the UI frameworks, at least on the web side (in favor of React mostly). I don't know what the situation is on the desktop UI side. But I know what they had before was loathed by many.
Developers hated it because it was like working with one hand tied behind your back. Traders hated it because developers couldn't deliver the best product because they were working with one hand tied behind their back.
I presume the core teams responsible for building the framework hated it because trying to get support for the proprietary UI tools was painful.
I imagine hiring managers and HR hated it because they would blatantly lie about the job - that the role would be WinForms/WPF (desktop) or JavaScript (web) UI development, without mentioning that all of that was really wrapped up by the proprietary Python framework, and very rarely would you get to touch the underlying industry standard tech stacks.
The Athena team ran out of names and one major component of it is called Bob.
c::a+b
total::select sum(balance) from accounts
I'm not sure if the K4 engine is smart enough to avoid recompute on updates to other columns.
If you want more fine grained control, you can set up triggers - e.g. "execute this code on update to variable a" - where your code gets the changed indices of a list/vector (or changed tuples of a table) as arguments; this is useful if a plain complete recompute is too costly, or you need to do something at the time of change, rather than at the next evaluation.
K2 and A+ had a similar system, each with different syntax.
Same team subsequently moved onto Bank of America to create Quartz, which is BoA's version of the same concept.
Then subsequently moved on to found Washington Square Technologies, where I think they're doing something similar, but not attached to a specific bank? Not 100% sure.
I've heard the concept was shopped around to a few other banks as well, but did not take root.