dc3ce57887
* docs(perf): preregister plugin-state polling workload * perf(reef): skip unchanged pending-review retries Carry the existing watched review receipt in parked inbox state. A decision mutation, policy or envelope change, or expiry resumes the normal trust and replay path. Avoid repeated replay claims and releases while an owner review is pending. Release note: reduce background SQL and worker requests for Reef messages waiting on owner review. No schema or public Plugin SDK change. * fix(perf): satisfy polling workload lint rules * test(reef): isolate pending-review regression coverage * test(reef): route receipt regressions through the database broker * test(reef): avoid shadowing the observed review receipt Co-authored-by: Peter Steinberger <steipete@gmail.com>
1.6 KiB
Plugin-state polling workload
Acceptance rule (before measurement)
Exercise an isolated Gateway with the bundled Reef channel enabled against a loopback HTTP/WebSocket relay. Seed synthetic identities and parked inbound messages awaiting review. Keep the real replay, review, trust, and plugin-state workers in the path. Do not use operator state or external services.
Count statements and worker requests per completed retry and per steady-state minute, with main-thread SQL reported separately. Separate startup and first delivery from warm polling. Attribute each recurring shape to its owner before changing production code. Also test an empty inbox control.
A candidate must remove recurring work while preserving replay exclusion, pending-review decisions, TTL visibility, quota behavior, and committed updates from other supported writers. No SDK signature or schema change is intended. Warm statements, requests, and main-thread work must not increase. Retain a regression for each removed read/write class and TTL/cross-writer correctness.
For wall-clock or CPU A/B comparisons, interleave at least five samples per arm; report median, p95, and range, keeping cold and warm separate. Every user-facing metric must hold or improve; a 20% adverse cell stops the candidate. Acquire the campaign timing lock and wait for load below 24 before timed comparisons. Counts and CPU profiles remain useful independently of wall-clock qualification.
If current main does not reproduce the live shape, report the workload and rates and stop that item. No optimization is justified by a synthetic direct-store loop that bypasses the channel owner.