Paolo ArdoinoVerified account

1/7 Problem: How to scale trading/liquidation engines up to 500k/1M/10M users? 100k users are manageable, but eventually single threaded architecture will bottleneck growing number of users.https://twitter.com/paoloardoino/status/1238965074427088903 …

2/7 * The order book (OB) for each pair is run on a single thread, with its own I/O queues. * User data UD is deterministically sharded among different threads * Liquidation engine (LE) is a set of threads with direct mem access to user data threads (positions, balances, ..)

3/7 * LE interacts with OB via in-mem priority queue. * LE adds to a liquidation-order (LEO) enough meta-info for the order book thread to evaluate if liquidation should still happen, bankruptcy price, ...

4/7 * Each LE sorts users by risk (probability of getting liquidated). * Improvement: a single shared sorted map can be generated across the threads

5/7 * Depending on the size of the position to be liquidated, LEO can be run into the book or positions can be acquired by the system and sold via more complex algos. We call it liquidation stages (LS).

6/7 * When a LEO gets processed, the OB thread evaluates (querying UD) if the user should still be liquidated and using which LS. * We tested this system with 100sK users + apocalyptic cascading liquidations and ended up always in ensuring the 0-sum game