Ask HN: How do HFT systems store position and order data?

[robinrheem]

HFT systems are really sensitive in latency. Which means it can't actually use traditional databases like MySQL, PostgreSQL, or etc. But then, how do we actually persist order and position data that the algorithms make? Would we just print it in the screen like application logs? I'm not sure that will be viable since logs can actually get lost. Maybe write on local files as CSV or HDF5? And perhaps make a cron or something to read those files and persist it into somewhere else? Or is there a solution that runs in memory but persists to disk?

[pnako]

The hot path involves the network card, the CPU, RAM and nothing else. The trading algorithm will keep various data structures in memory to keep track of current position and position limits.

Outside of the hot path/thread, you can do pretty much whatever you want. You typically broadcast your trades on the network for other systems to pick them (and put that data into SQL databases, typically). The databases used for risk and compliance, etc. will most likely get their data directly from the exchange.

The idea is that you don't need to query a central system for every single trade you want to make. HFT systems have some limits set up at the beginning of the trading session, telling they can take a position of up to X in security Y.

[robinrheem]

Thank you very much for your kind reply. So does that mean the algorithm, which is in the hot path will just broadcast via a messaging system like ZeroMQ or Kafka to another subscriber that’ll just save the order and position data to the database right? Plus the messaging system will be nonblocking and just do a fire and forget style.

[pnako]

The hot path is the time between receiving market data from the network, and reacting to it by potentially sending an order. It's typically a thread constantly polling events from the network card.

In that path the rest of the world does not exist: you are interacting with an exchange, possibly multiple exchanges, and that's it. The rest of the world doesn't exist. No ZeroMQ or Kafka or whatever.

After you send an event, you can set a flag or copy some data structures for another thread of lower priority to do some house keeping. This might include sending an update on some messaging system, yes. This can be something like TIBCO RV, or a proprietary equivalent, or perhaps the ones you mentioned.

But the bottom line is: when doing low-latency, high-frequency trading you don't want to do anything that is not strictly required in the hot path. You must read the message coming from the market, very quickly decide how you want to react (i.e. send a new order or update an order, at which price point, etc.) and that's it. You will in fact even avoid computing too much stuff here; you would have pre-computed as much stuff as possible to be ready to very quickly do a lookup and maybe some trivial computation before making your decision.