On 2015-11-18, at 08:19, Victor Gil wrote: > We have rather high transaction volume, especially around the market > open/close times, so, yes, we did have duplicates before and sometimes had to > retry multiple times [based on the bumped sequence number] to provide for > uniqueness. > > I can't go into specifics, but the whole idea was to spread the input queues, > containing a dynamic mix of transactions, into logical sub-queues related to > their "types", all of which are to be still processed in parallel. However, > within a given "type" each distinct transaction "originator" has to be > processed sequentially, so that a request to "cancel" an order is delayed > until the original order has actually been seen and processed. > Ah, yes. The high-frequency trading wars. 60 Minutes had a story a few years ago mentioning how one institution overcame a disadvantage by introducing a wire(!) delay line to delay its competitor's transactions by a few milliseconds.
-- gil
