v0.5.0-beta — the durability release
Pre-release
Pre-release
The durability release. Cluster state stops living only on the box that produced it, and the build that produced it can be named.
Highlights
- Order lifecycle latency 16.4 ms → 757 µs. The OMS egress streams were riding the market-data conflation timer, which exists to conflate a price feed and had no business batching per-order facts. They flush on the Disruptor's
endOfBatchnow: opportunistic batching, no wall clock (#159). - The committed round trip is measurable. CreateOrder offered → first status back, replicated and applied, reported separately from ingress publication, which never leaves the machine. The two differ by roughly 880x on the same run; quoting one as the other is the trap this closes (#161, #173).
- Bundles for the matching engine: a snapshot plus the log behind it, as one durable unit, with a retention watermark (#162).
- The engine jar and the load generator are published to the artifact store, addressed by commit sha and by release tag, with
Build-Shain the manifest (#165, #173). - The cluster machinery comes from
cluster-kitrather than a second copy in this repo, so a fix lands in both engines or neither (#163). - Bare
java -jarworks: shaded jars carryAdd-Opens(#153). The load generator no longer assumes the network interface is calledeth0(#151).
Full detail in CHANGELOG.md.