Rotorsoft / act-root / 32586650779

100%

push

github

feat(act): one worker reads the event log for all of them (#1537)

* test(act-pg): measure correlate with workers as real processes

Correlate had never been benchmarked multi-process. Every earlier run put
several Acts in one Node process with notify off, so only the process doing the
committing was ever armed and only it ever scanned, which hid the cost
entirely.

With W real worker processes and notify on, every worker reads every event and
writes every mark: reads per committed event come out at exactly 1.00, 2.00,
4.00 and 8.00 for one, two, four and eight workers, and subscribe entries scale
identically. The checkpoint is per-instance, seeded from the durable value once
at init and advanced locally after that, so the shared checkpoint is a
cold-start floor and nothing more.

Delivery is untouched. Handled and fetched counts track commits exactly at
every worker count, so the duplication is entirely in discovery and never in
delivery.

This is the measurement RFC 1484 deferred single-writer coordination on, and it
argues for serializing correlation rather than for caching what the duplication
produces: a window cache removes one read out of W+1, so half at one worker and
an eleventh at eight, while the cost that grows with scale is untouched.

Also corrects the drain-fetch numbers from the earlier correlate-reads run,
which counted aggregate loads as fetches. The drain is exactly one fetch per
event with no empty windows, and removable round trips are 31 percent rather
than the 62 percent first published.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* feat(act): lease correlation so one worker scans for all of them

Every worker read the whole event log looking for work and wrote the same
bookkeeping about what it found. Measured with real worker processes, reads
and mark-writes per committed event came out at exactly one, two, four and
eight for one, two, four and eight workers.

The writes are the expensive half. They la... (continued)

4237 of 4237 branches covered (100.0%)

Branch coverage included in aggregate %.

152 of 152 new or added lines in 8 files covered. (100.0%)

9191 of 9191 relevant lines covered (100.0%)

1331.71 hits per line