Latency
We had been telling ourselves the routing overhead was fine. Then we sat down and profiled it, and it was not. The hot path was doing work on every order that did not need to be there at all. Taking that work out helped. The larger win came from changing how sessions are handed off when several accounts are connected at once. p99 came down 39.28%. More useful than that headline number: the tail got much flatter under concurrent load. Behavior is more predictable when things are busy, rather than only faster on a quiet day.See the current numbers
GET /v1/platform/latency returns platform intent to broker handoff percentiles. Read window_seconds for the span they cover: it is the trailing hour when the platform is busy, and reaches further back when it is quiet.Session recovery
Reconnects after a broker-side drop are fixed. Recovery used to wait on someone noticing and stepping in. The session now comes back on its own, which is how it should have worked from the beginning.Multiple accounts
Running multiple accounts is smoother. Adding a second or third account used to be a path that got less testing than the single account case, and it showed. That is evened out now.Metrics
We had been collecting far more than anyone was reading. We cut it back to the things people check: fill latency, reject rates, and session health. Everything loads quicker, and a problem is easier to spot when there is one.Dashboard
Order flow updates live rather than trailing a few seconds behind, so what you are looking at is what is happening. There are still parts of the interface we are not happy with, and those are next.Tell us what is slow
If you are running on this and something is slow or irritating, say so. That is what ends up at the top of the list.Community
Bring it to the community Slack. This is the fastest way to reach the team.
Roadmap
What we are building next, and why.