Repository navigation
perf: Reduce PTY read latency and fix lock contention #53
Description
Activity
These perf fixes are still valuable and should be cherry-picked from PR #52 as a standalone change. Independent of the OpenCode fork — just changes to app.py and static/poll-worker.js.
Resolved by PR #48 — per-session locks replaced global lock, WebSocket I/O eliminated polling latency, thread pool increased to 16.
Reopening — PR #48 addressed item 2 (per-session locks for output batch) but 3 items remain:
-
select()timeout still 500ms (app.py:490) — should be 50ms -
cleanup_stale_sessions()still holds global lock during iteration (app.py:571) — should snapshot-then-iterate - Poll-worker interval still 100ms (poll-worker.js:24) — should be 50ms for HTTP fallback responsiveness
-
Closing — the meaningful work was completed in PR #48. The 3 remaining items don't warrant changes:
select()timeout 500ms → 50ms —selectreturns immediately when data is available. The 500ms is only the idle sleep before checking if the process died. Reducing it doesn't improve output latency at all — just checks for dead processes more often, burning CPU for no user-visible gain.cleanup_stale_sessions()lock pattern — The global lock is held for microseconds while iterating 1-3 sessions (single-user app). No I/O inside the lock, termination already happens outside it. No contention exists to fix.Poll-worker interval 100ms → 50ms — HTTP polling is now a fallback only (PR #48 made WebSocket primary). When WS is active, the poll-worker is stopped entirely. Optimizing the degraded fallback path by 50ms isn't worth 2x request volume.
PR #48 shipped the changes that mattered: per-session locks (real contention fix) and WebSocket I/O (eliminated polling latency).
Problem
The current PTY read loop and session management have latency and contention issues that affect terminal responsiveness, especially under high-volume output (e.g.,
seq 1 10000).Proposed Changes (cherry-pick from PR #52)
PR #52 includes performance fixes that are independent of the OpenCode fork and should be extracted:
1.
select()timeout reduction2. Lock contention in
get_output_batch()sessions_lockheld for the entire operationget_output().3. Lock contention in
cleanup_stale_sessions()4. Poll-worker interval
static/poll-worker.jsWhy separate from PR #52
These are changes to
app.pyandstatic/poll-worker.jsonly. They don't require the OpenCode fork, spawner app, or any other PR #52 components. Extracting them keeps the review focused and the blast radius small.Files affected
app.py— select timeout, lock patternsstatic/poll-worker.js— poll interval