Description
Problem
With #34, an idle Redis worker waits on a marker with BZPOPMIN instead of polling. When a single Redis core saturates, the PR drains fewer jobs than polling did, because each job costs more round trips.
Measured on a 16-vCPU orb, Redis 7.4 pinned to one dedicated core, workers and producer pinned to other cores, 6 interleaved reps per variant (head 9b55a64):
| Cell | main drained/s |
PR drained/s | Round trips/job |
|---|---|---|---|
| 100 workers, concurrency 1, 10k/s | 8932 | 7145 (−20%) | 4.0 → 5.9 |
| 32 workers, concurrency 10, 10k/s | 9070 | 7459 (−18%) | |
| 100 workers, concurrency 10, 5k/s | 4897 | 4992 |
Redis time per job: 102 µs on main, 138 µs on the PR. Round trips cost Redis about 14 µs each, more than the work inside the scripts.
Where the extra round trips go
A job wakes several workers. Per job, at saturation:
- 1.2 to 1.9
BZPOPMINreturns; - 1.3 to 2.4
popFromcalls that find nothing; - up to 0.34
RELEASE_MARKERcalls (32 workers, concurrency 10).
The push adds a ready marker, and the acquisition adds it again while jobs remain, so both often wake a worker for the same job.
Ideas, untested
- Re-mark only when more jobs remain than workers already woken, or carry a count in the marker.
- Let the acquisition script pop the next job for a worker woken by a marker, so a wake costs one round trip.
- Drop the release round trip for a ready marker popped after the wait ended, if another marker already covers the queue.
Any change must keep the guarantee that a stored job wakes at least one idle worker, or is found at the next stalled check.
Activity