OSS-Singularity/zcode-dispatch-observatory
OSS-Singularity/zcode-dispatch-observatory
Watch
1
0
Fork
You've already forked zcode-dispatch-observatory
 0
Code Issues Pull requests Projects Releases 137 Packages Wiki Activity Actions

Pinky W: cheap poll cycles — response cache, delta memo, visibility pause (v1.27.0) #53

Merged
ClaudiuSchuster merged 1 commit from oss-oo/pinky-w-cpu into main 2026-09-26 08:29:19 +00:00
Conversation 0 Commits 1 Files changed 4 +1297 -147
ClaudiuSchuster commented 2026-09-26 08:28:06 +00:00
Owner

Owner top blocker: one open dashboard pegged a core (~170 % CPU, station 90–95 °C). Three layers, semantics untouched.

R1 — response cache: both poll endpoints serve cached bodies behind a change key over EVERY source (DB stat incl. -wal/-shm + data_version probe, tasks index, ledgers, contract, questions side-channel, state file, provider config, live /proc digests). Any source mutation rebuilds on the NEXT poll; TTL 8 s refreshes time-derived fields; ONE fetch per cycle shared by both endpoints; ZDO_CACHE=0 opts out.

R2 — hot-path memo: per-session facts keyed on session.time_updated (flushes with parts/turns — verified live 2026-09-26), re-derived only for changed/new/hot sessions; doorbell scan rowid-incremental. Measured on the station snapshot (340 sessions, 164k parts): idle rebuild 45 ms vs 2189 ms before; old-vs-new payloads IDENTICAL (normalized diff).

R3 — visibility governor: hidden tab pauses main/drawer/modal/dock timers; visible re-arms + immediate refresh. Default cadence unchanged (3000 ms).

Evidence: 67/67 unittests (15 new, one per source class incl. second-writer inserts, running→terminal + webhook-once, tasks flip, ledger rewrites, TTL, dedupe, busy-raises-not-stale, doorbell memo pickup); Super-Linter green (BASH/BASH_EXEC/MARKDOWN/PYTHON_PYLINT/PYTHON_RUFF/YAML); fixture rig E2E PASS: new session + running→terminal each visible in ~1.03 s (≤ 2 cycles budget 8 s), cache-hit bytes identical, CDP probes green (table, usage SVG, drawer + steer console, R3 pause/resume in the live page). Live G1/G2 measurements run after merge+install (dossier follows).

Owner top blocker: one open dashboard pegged a core (~170 % CPU, station 90–95 °C). Three layers, semantics untouched. **R1 — response cache**: both poll endpoints serve cached bodies behind a change key over EVERY source (DB stat incl. -wal/-shm + data_version probe, tasks index, ledgers, contract, questions side-channel, state file, provider config, live /proc digests). Any source mutation rebuilds on the NEXT poll; TTL 8 s refreshes time-derived fields; ONE fetch per cycle shared by both endpoints; ZDO_CACHE=0 opts out. **R2 — hot-path memo**: per-session facts keyed on session.time_updated (flushes with parts/turns — verified live 2026-09-26), re-derived only for changed/new/hot sessions; doorbell scan rowid-incremental. Measured on the station snapshot (340 sessions, 164k parts): idle rebuild 45 ms vs 2189 ms before; old-vs-new payloads IDENTICAL (normalized diff). **R3 — visibility governor**: hidden tab pauses main/drawer/modal/dock timers; visible re-arms + immediate refresh. Default cadence unchanged (3000 ms). Evidence: 67/67 unittests (15 new, one per source class incl. second-writer inserts, running→terminal + webhook-once, tasks flip, ledger rewrites, TTL, dedupe, busy-raises-not-stale, doorbell memo pickup); Super-Linter green (BASH/BASH_EXEC/MARKDOWN/PYTHON_PYLINT/PYTHON_RUFF/YAML); fixture rig E2E PASS: new session + running→terminal each visible in ~1.03 s (≤ 2 cycles budget 8 s), cache-hit bytes identical, CDP probes green (table, usage SVG, drawer + steer console, R3 pause/resume in the live page). Live G1/G2 measurements run after merge+install (dossier follows).
Pinky W: cheap poll cycles — response cache, delta memo, visibility pause (v1.27.0)
All checks were successful
Checks / Functional checks (pull_request) Successful in 18s
Details Explain this check
8e5564feb8 
The dashboard's 3 s poll recomputed the full station scan twice per
cycle (~4.4 s CPU per 3 s ≈ 170 % of one core with one tab open).
Three layers, semantics untouched:

- Response cache (R1): /api/sessions and /api/usage serve byte-identical
  cached bodies while a change key over EVERY source is stable — session
  DB + tasks index (stat incl. -wal/-shm + PRAGMA data_version probe),
  dispatcher ledgers (bells/queue/dispatches/contract), questions
  side-channel, observatory state, provider config, live /proc digests
  (spawn linkage, cwd locks, in-flight steers). Any mutation rebuilds on
  the next poll; a TTL (ZDO_CACHE_TTL_S, 8 s) refreshes time-derived
  fields; both endpoints share ONE fetch per cycle; ZDO_CACHE=0 opts
  out. Invalidated per source class in tests/test_poll_cache.py (15
  tests: second-writer inserts, running→terminal + webhook-once, tasks
  archive flip, ledger rewrites, contract, state, questions dir, live
  digest, TTL, dedupe, busy-raises-not-stale, doorbell memo pickup).
- Hot-path memo (R2): per-session facts (context, pending questions,
  activity, hydration, models) keyed on session.time_updated (flushes
  together with parts/turns — verified live), re-derived only for
  changed/new/hot sessions; doorbell lineage scan rowid-incremental.
  Idle rebuild 45 ms (was 2189 ms measured on the station snapshot);
  old-vs-new payloads proven identical.
- Visibility governor (R3): a hidden tab pauses the main/drawer/modal/
  dock timers, visible re-arms + one immediate refresh (CDP-verified in
  the live page). Default cadence stays 3000 ms.

Also: SERVER_VERSION v1.27.0 marker in the startup line and the
sessions payload; README section + env knobs documented.
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
OSS-Singularity/zcode-dispatch-observatory!53
No description provided.