PERF: use one resettable flush task per collaboration room (#663) #778

New issue

Open

opened 2026-10-02 13:10:32 +00:00 by kayg · 2 comments

kayg commented

2026-10-02 13:10:32 +00:00

Owner

Context: #663 coalescing/backpressure audit (rules 5 and 8). Source evidence at c4a61e8cf and job/merge-round-7a 2f4482ded.

Evidence at round-7a crates/calternal-collab/src/session.rs:

:1176 schedule_flush starts a new Tokio task for each dirty update (:1184). Generation invalidates old tasks after their sleep, but does not cancel them before they run. At the three-second cap, overdue tasks can all attempt the same flush lock.
:2881 calls schedule_flush after each changed WebSocket frame. An authenticated client has a message byte cap; that cap does not bound the count of pending watcher hints or timer tasks.
The last-client retry path already uses one retrying flag per room (:2890 onward). Reuse that ownership discipline.

Reasoned impact: high edit rates allocate many timers holding Hub/Room references. Slow filesystem or projection publication increases pending work. Generation checks skip stale work only after the timers wake. This is unbounded per-room scheduling; this audit did not run a hostile load or demonstrate process exhaustion, so this is not labelled BLOCKER.

Concrete fix: one resettable flush task per room with a first-dirty deadline and bounded retries. Preserve the 750 ms quiet/3 s max-save guarantees and unload/shutdown invariants.

Tests: use a controlled slow flush, then send a bounded burst. Assert a fixed task count per room and convergence to the last source revision. Continuous edits must still save by the maximum wait; shutdown and last-client leave must flush; duplicate events and rename/delete must converge. Measure pending tasks high-water and CPU/RSS with the existing collaboration fixture.

Duplicate search: #265 large-Note sync, #270 public frame limit, #668 shared app delta. None replaces per-frame flush timers. Concurrent audit #761 owns the watcher queue; reuse its ingress and overflow contract, but keep this issue scoped to room persistence scheduling.

Context: #663 coalescing/backpressure audit (rules 5 and 8). Source evidence at c4a61e8cf and job/merge-round-7a 2f4482ded. Evidence at round-7a crates/calternal-collab/src/session.rs: - :1176 schedule_flush starts a new Tokio task for each dirty update (:1184). Generation invalidates old tasks after their sleep, but does not cancel them before they run. At the three-second cap, overdue tasks can all attempt the same flush lock. - :2881 calls schedule_flush after each changed WebSocket frame. An authenticated client has a message byte cap; that cap does not bound the count of pending watcher hints or timer tasks. - The last-client retry path already uses one retrying flag per room (:2890 onward). Reuse that ownership discipline. Reasoned impact: high edit rates allocate many timers holding Hub/Room references. Slow filesystem or projection publication increases pending work. Generation checks skip stale work only after the timers wake. This is unbounded per-room scheduling; this audit did not run a hostile load or demonstrate process exhaustion, so this is not labelled BLOCKER. Concrete fix: one resettable flush task per room with a first-dirty deadline and bounded retries. Preserve the 750 ms quiet/3 s max-save guarantees and unload/shutdown invariants. Tests: use a controlled slow flush, then send a bounded burst. Assert a fixed task count per room and convergence to the last source revision. Continuous edits must still save by the maximum wait; shutdown and last-client leave must flush; duplicate events and rename/delete must converge. Measure pending tasks high-water and CPU/RSS with the existing collaboration fixture. Duplicate search: #265 large-Note sync, #270 public frame limit, #668 shared app delta. None replaces per-frame flush timers. Concurrent audit #761 owns the watcher queue; reuse its ingress and overflow contract, but keep this issue scoped to room persistence scheduling.

kayg changed title from ~~PERF: bound collaboration watcher hints and flush scheduling (#663)~~ to PERF: use one resettable flush task per collaboration room (#663)

2026-10-02 13:11:31 +00:00

kayg referenced this issue

2026-10-02 13:11:35 +00:00

PERF: bound the collaboration filesystem event queue (#663) #761

kayg referenced this issue

2026-10-02 13:16:50 +00:00

PERF: instant everywhere — app-wide speed architecture rules, audit and shared primitives #663

kayg referenced this issue

2026-10-02 13:19:27 +00:00

PERF: instant everywhere — app-wide speed architecture rules, audit and shared primitives #663

kayg commented

2026-10-02 13:25:39 +00:00

Author

Owner

Starting work on job/notesperf, based on job/merge-round-7a at 2f4482ded066d9c5d9c59130377907f7fd2916c9. I will address the issue with focused changes and regression coverage, then report the final head SHA and verbatim gate output here.

Starting work on `job/notesperf`, based on `job/merge-round-7a` at `2f4482ded066d9c5d9c59130377907f7fd2916c9`. I will address the issue with focused changes and regression coverage, then report the final head SHA and verbatim gate output here.

kayg commented

2026-10-02 19:10:47 +00:00

Author

Owner

Each dirty room now owns one resettable debounce task with a three-second maximum dirty deadline. A 100-call burst regression test passed: cargo test -p calternal-collab --lib flush_burst_starts_one_room_timer_task (1 passed). Commit: 750b36129. The full crate gate and performance profile are pending.

Each dirty room now owns one resettable debounce task with a three-second maximum dirty deadline. A 100-call burst regression test passed: `cargo test -p calternal-collab --lib flush_burst_starts_one_room_timer_task` (1 passed). Commit: 750b36129. The full crate gate and performance profile are pending.

kayg referenced this issue

2026-10-02 21:09:16 +00:00

PERF: mount Notes before optional syntax grammars load (#663) #747

kayg referenced this issue

2026-10-03 02:29:47 +00:00

PERF: mount Notes before optional syntax grammars load (#663) #747

kayg referenced this issue from a commit

2026-10-03 02:35:28 +00:00

fix(collab): save edits accepted during the single room timer flush