Loading
Suggested changes for the subscription coalescer
Suggested changes for !266624 (merged).
This MR targets your branch 630854-subscription-coalescer, so you can merge it there. The commits carry my name. Rebase and reset the author if you want them under yours.
Two commits fix bugs:
- A key can stop flushing. If the scheduler raises in
Coalescer#submit, the entry stays and no flush is scheduled. Later updates for that key are never sent. - Old updates carry into the next window.
entry.latestis not reset after a flush, and one blank WAL location then pins every later merged update for that key to the primary.
The other commits move code out of ActionCableWithLoadBalancing and follow existing patterns. Each commit message says "Behavior change" or "Structure only", so you can review the structure commits quickly.
| Commit | Type |
|---|---|
Use Concurrent::Map in the subscription coalescer |
behavior |
Extract Envelope from ActionCableWithLoadBalancing |
structure |
Compare WAL locations with an Lsn value object |
behavior |
Add CoalescePolicy for subscriptions that coalesce |
structure |
| Resolve the coalesce policy once for each stream | performance |
Extract subscribers_of and rename defer_update |
structure |
Inject the subscriptions into ActionCableSampler |
structure |
How I checked it
- 21 fast examples and 228 integration examples pass, and RuboCop reports no offenses.
- A deferred submit takes 1.26 µs, against 0.38 µs for the
Mutexversion. With the real WAL merge it takes 5.05 µs, against 2.16 µs. A deferred update replaces a full subscription run that takes milliseconds. - Before the "once for each stream" commit, every message on every subscriber stream paid about 1.6 µs for a schema lookup. After it, subscriptions without
coalescepay nothing. - The cap check and the topic counter are separate steps, so the key caps can go over by about the size of the worker pool.
- The block given to
@bursts.computemust not call@burstsagain, because the map's lock cannot be taken twice by one thread.