HRT's Python fork: Leveraging PEP 690 for faster imports
hudsonrivertrading.com
hudsonrivertrading.com
For most teams I would be pretty skeptical of a internal Python fork, but the Python devs at HRT really know their stuff.
Sometimes it will be worth the tradeoff to put that person and a programmer together to code up a solution in another language. Sometimes it will be worth it to have the non-programmer write it in Python and then do Herculean things in the background to make it fast enough.
Nim exists, Crystal exists
But it also wouldn't surprise me if a lot of shops land on python because that's what their hiring pool knows.
if you are someone like HRT I presume the bulk of their money comes at very short holding periods so you have e.g. fast signals that work short term and then mid frequency alpha signals that spit out a forecast over a few timeframes i.e. it might not be that they buy (aggressively) really quickly but rather than someone sells to them and then they hold onto the position for longer than they would if they have no opinions.
Similarly this shapes where you post your orders e.g. if you really want it then you want to be top of the book
This article from a16z explains the mechanics of reordering transactions for profit (MEV): https://a16zcrypto.com/posts/article/mev-explained/
Though the encoder runs 64/66bits at a time, so you really get around 8B every 7ns or so.
Most trading firms are past the whole "beat the other guys to buy". Established large investment firms already have all that on lockdown in terms of infrastructure and influence to the extent where they basically just run the stock market at this point (i.e Tesla posts horrible quarter numbers, but stock goes up).
Most of the smaller firms basically try to figure out the patterns of the larger firms and capitalize on that. The timescales have shifted quite a bit.
In some cases the order leaving the card starts to emerge before the packet containing the market data event that they're responding to has even finished arriving.
Waiting for a full microsecond for the packet to arrive before responding means you're already too slow
The speed game is essentially over
There are many more games to play than delta one takeout and the solutions certainly don’t fit on one or a handful of FPGA’s.
Do you mean that because it involves a lot of hardware design now? The days of being able to offer around the inside in C++ on a regulated securities exchange are over, but there's still C++ driving the thing, that 20ns "tick to trade" or however it's being measured in some instance is still pretty basic response stuff, light speed is still a thing. There's a C++ program upstairs running the show, and it's trying to do it's job in under a mike for sure.
The OG talk on this is Carl Cook's: https://www.youtube.com/watch?v=NH1Tta7purM
But there are more recent talks (Optiver is especially transparent about it but other people talk about it too): https://www.youtube.com/watch?v=sX2nF1fW7kI, that's David Gross at CppCon last year, it can't have changed that much since last year.
No matter how fast you process the data, the ping difference of 1ms is going to be an advantage that you can never beat.
There is a reason why firms like HRT trade mostly in derivatives and futures.
This is a non sequitur from who’s winning the HFT game
But yes, like you I had a great experience
Its a finance firm - i.e scam firm. "We have a fancy trading algorithm that statistically is never going to outperform just buying VOO and holding it, but the thing is if you get lucky, it could".
Scammers are not tech people. And its pretty from their post.
> In Python, imports occur at runtime. For each imported name, the interpreter must find, load, and evaluate the contents of a corresponding module. This process gets dramatically slower for large modules, modules on distributed file systems, modules with slow side-effects (code that runs during evaluation), modules with many transitive imports, and C/C++ extension modules with many library dependencies.
As they should.
The idea that when you type something in the code and then the interpreter just doesn't execute it is how you end up with Java like services, where you have dependency injection chains that are so massive that when the first time everything has to get lazily injected the code takes a massive amount of time to run. Then you have to go figure out where is the initialization code that slows everything down, and start figuring out how to modify your code to make that load first, which leads to a mess.
If your python module takes a long time to load, this is a module problem. There is a reason why you can import submodules of modules directly, and overall the __init__.py in the module shouldn't import all the submodules by default. Structure your modules so they don't do massive initialization routines and problem solved.
Furthermore, because of pythons dynamic nature, you can do run time imports, including imports in functions. In use, whether you import something up at the top and it gets lazily loaded or you import something right when you have to use it has absolutely no difference other than code syntax, and the latter is actually better because you can see what is going on rather than the lazy loading being hidden away in the interpreter.
Or if you really care, you can implement lazy work process inside the modules, so when you import them and use them the first time it works exactly like lazy imports.
To basically spend time building a new interpreter with lazy loading just to be able to have all your import statements up at the top just screams that those devs prefer ideology over practicality.
You wish lol. How do you think they pay for all the developers?
Firms like HRT don't even take outsider money, they don't really need to.
And besides, we don't get paid for beating stocks, a lot of funds will do worse than equities in a good year for the latter, the whole point is that you're benchmarked to the risk free rate because your skill is in making money while being overall market neutral. So you rarely take a drawdown anywhere near as badly as equities.
As a service this is often a portfolio diversification tool for large allocators rather than something they put all the money into.
It is true however that some firms are basically just rubbish beta vehicles that probably should in an ideal world shut down.
Good returns - take other peoples money, trade it, take 20% of profits
Excellent returns - trade your own money, make a bit less overall but keep 100% of profits
In the second case, why start a company?
Prop shops usually come about as the partners buy out other investors in a fund
HRT trades their own money so if it didn't beat VOO then they'd just buy VOO. There are no external investors to scam.
Please don't post shallow dismissals...
I agree, which is why you should design your modules correctly and import only the stuff you need.
I was pointing out that lazy imports vs runtime imports in functions are basically the same and lead to the same issues.
How do you achieve this without making gazillions of modules, where each module has just a few stuff?
Are you saying just use local import everywhere?
So when you do this
import bigmodule
Then, you have your big initializer code in bigmodule.financedata. But the stuff you need for running scripts is in bigmodule.scripts.
So when you write
from bigmodule import financedata
But if you write
from bigmodule import scripts
You don't need to have gazillion modules, just good organization. Also, in general, its a good practice to gate intensive compute/network operations behind an explicit function you need to call.
Also thank you for focusing the convo on the tech stuff instead of repeating finance bro myths
And when you have to depend on external libraries beyond your control, how do you typically handle those situations?
As for external libraries, you import them in places where you need to use them only to avoid the same pitfalls. Its also pretty easy to analyze the import process within those libraries, and then again import specific submodules only that limit what actually gets loaded.
I guess you meant "run all imports at startup is desirable to check if they work", but I have a hard time agreeing with that (personally I think having a good test suite is needed, whereas running more code at startup is not wanted).
How they make $8B/y underperforming VOO?
Reference: https://www.businessinsider.com/hudson-river-trading-hrt-8-b...
In case you are unaware - very single trading firm makes money on fees, not by outperforming the market. This goes for firms like Vanguard too.
Just think about it for a little bit - if you could reliably outperform the market by any % with an algorithm why even start a company? Just take out loans, invest, make money, repeat and become rich. No expenses to manage a company.
I really want to know what you think this company does, precisely
If a company has customers, and those customers buy a product, the company charges a price for that product.
They don't...
b) If you don't have customers, why have a company?
There is no way to invest in the company, and the only way of becoming a "customer" is to engage in trading.
A customer might want to offload TSLA and are willing to pay the market rate of $329. This trade might work in HRT's favor if they can sell it for $329.01. It's just 1 cent, but over millions of transactions these small amounts of profit add up.
The value captured by HRT is meaningful in the aggregate, but tiny and irrelevant to the institutional clients, and therefore can't be thought of as a fee. What HRT provides in return for taking clients' trades is liquidity.
So then, you have to offer additional services. HRT has the SDP that they provide, and of course charge fees for. But then the question is why would anyone do this, versus just going through any other financial institution.
The answer is basically all up on their website
"As a liquidity provider, HRT develops automated trading algorithms designed to provide the best prices to our clients".
The question that should be asked is as a user, why would I want to sign up with HRT or any similar financial company? The answer for HRT is because you want to have access to more complicated financial derivatives - you don't need to sign up and pay fees to buy basic stock.
So they promise that their algorithms give the user the best price, which is a legal way of saying that you will pay less for a certain asset and make money on it, and you can't say that because you can't guarantee this.
And its well known that nobody ever gets rich of an algorithm in finance unless you are well established large firm intricately tied with the government that can move so much money as to influence trading.
Moreover, I would be very surprised if the majority of their $8 billion annual profit came from client market making.
The incentive for users to sign up with them is to get access to "better" pricing for whatever commodity they pair the buy/sell orders for - but remember these are futures so its all betting, and so the algorithms don't really mean anything.
Isn’t it actually the opposite? they pay for order flow instead? They should be making money from bid-ask spread, not fees.
They pay fees to exchanges.
As a market maker, some rebates are given back conditional on their activity.
They have no users.
You’re just constantly obliviously asserting falsehoods that betray an almost comical lack of understanding of the reality of these businesses.
“client market making”. That’s very different from “market making”, you two are not in agreement at all here
All your posts here are low-information anti-finance rants.
It would be great if you included any sort of evidence or argument.
Reading on to the other comments, it looks like you're throwing out a lot of accusations and claims. I don't know what you think you know, but from the looks of it, you don't really know HRT's business. I don't really these days, but I knew it years ago, and it's not from taking client money or arbitrage or some weird scam. It's not magic but the world of algo trading isn't a ponzi scheme.
You're confusing prop shops and hedge funds.
I was quite impressed by the interviews, mostly for their pragmatism and skill-fitting. The programming interview wasn’t LC, it was “can you use a language (preferably Python) to parse a CSV and get useful information out of it,” because that’s the skill level the team needs. On the other hand, the Linux and DB interviews were quite in-depth, because again, the team needs those skills.
10/10 would interview again if I’m ever near an office again.
Of course that doesn't solve the overhead of finding the modules, but that could be optimized without lazy import, for example by having a way to pre-compute the module locations at install time.
Exactly this. There must be zero side effects at module import time, not just for load times, but because the order of such effects is 1) undefined, 2) heavily dependent on a import protocol implementation, and 3) poses safety and security nightmares that Python devs don't seem to care much about until bad things happen at the most inconvenient time possible.
> Of course that doesn't solve the overhead of finding the modules, but that could be optimized without lazy import, for example by having a way to pre-compute the module locations at install time.
1) opt for https://docs.python.org/3/reference/import.html#replacing-th...
2) pre-compute everything in CI by using a solution from (1) and doing universal toplevel import of the entire Python monorepo (safe, given no side effects).
3) This step can be used to scan all toplevel definitions too, to gather extra code meta useful for various dynamic dispatch at runtime without complex lookups. See for example: https://docs.pylonsproject.org/projects/venusian/en/latest/i...
3) put the result of (2) and (3) as a machine-readable dump, read by (1) as the alternative optimised loading branch.
4) deploy (3) together with your program.
Oh no. Look I'm not saying you're holding it wrong, it's perfectly valid to host your modules on what is presumably NFS as well as having modules with side effects but what if you didn't.
I've been down this road with NFS (and SMB if it matters) and pain is the only thing that awaits you. It seems like they're feeling it. Storing what is spiritually executable code on shared storage was a never ending source of bugs and mysterious performance issues.
The lazy import approach was pioneered in Mercurial I believe, where it cut down startup times by 3x.
Today, it would be a mistake to use anything other than Rust (hence Jujutsu carrying the flame forward).
It's dirty, but speeds things up vs putting all imports at the top.
import argparse
parser = argparse.ArgumentParser()
parser.parse_args()
import requests
As it is, top-level imports IMHO are only meant to be used for modules required to be used in the startup, everything else should be a local import -- getting everyone convinced of that is the main issue though as it really goes against the regular coding of most Python modules (but the time saved to start up apps I work on does definitely make it worth it).
Import-time side effects are definitely nasty though and I wonder what the implications on all downstream code would be. Perhaps a lazy import keyword is a better way forward.
> monorepo
> vast proliferation of imports
> large modules
> distributed file system
> side-effects
> many transitive imports
This sounds like a very optional problem to have.
Also maybe, if this approach could yield stats on if some import was needed or not ?
I'd say if you see
from typing import Final
[...]
__all__: Final = ("a", "b", "c")
Of course, should one be doing a star import in a proper codebase?
That, or stop helping your coworkers/accommodating them… risky, as a career move. Only seen that work once.
A PEP is very much welcome, but using lazy import libraries is a fairly common, very old, method of speeding things up. My pre PEP 690 code looks like this:
import typing
from lazy import LazyImport
member = LazyImport('my_module.subpackage', 'member')
member1, member2, = LazyImport('my_module', 'member1', 'member2')
if typing.TYPE_CHECKING:
# normal import, for linter/IDE/navigation.
from my_module.subpackage import member
from my_module import member1, member2