Keep similar-ai from tripping the global circuit on a lone URL 502, clamp Qdrant search to 100, and add Server-Timing plus slow-request logging. Studio shared props, Academy S3 exists caching, heat chunking, and Redis/scheduler hygiene stay in this rollout.
6.2 KiB
M8 — Redis Orphan Cleanup & Presence Compaction
STATUS: COMPLETE (tooling + tests; no production UNLINK/SREM executed)
WHEN: 2026-08-23 ~18:55 Europe/Ljubljana
1. Production before snapshot (read-only)
| Metric | Value |
|---|---|
| used_memory | 883.41 MB (926,320,256) |
| used_memory_rss | 176.88 MB |
| used_memory_peak | 917.11 MB |
| maxmemory | 2.00 GB |
| fragmentation_ratio | 0.20 (RSS ≪ used → swapped) |
| evicted_keys | 0 |
| connected_clients | 28 |
| blocked_clients | 0 |
| ops/sec | 92 |
| keys | 4838 (4100 with TTL) |
| REDIS_PREFIX | skinbase-database- |
| HORIZON_PREFIX | skinbase_horizon: |
| Queue | llen | delayed | reserved |
|---|---|---|---|
| default | 0 | 0 | 0 |
| search | 0 | 0 | 0 |
| 0 | 0 | 0 | |
| collections | 32632 | 0 | 0 |
| forum-moderation | 553090 | 0 | 0 |
| forum-security | 5538 | 0 | 0 |
Presence SCARD 2,480,634. Sample 200 members: 200 stale / 0 live (ratio 1.0).
Legacy SCAN: skinbasenova_horizon: 662 keys ~7 MB; skinbasenova-database- 11 keys ~23 MB.
2. Producer containment
45s later: collections 32632, forum-moderation 553090, forum-security 5538 — unchanged.
Config: collections/forum_ai/forum_sec dispatch_enabled=false.
Newest forum payloads ~18:47 then no growth. Proceed.
3. Queue class samples (LINDEX/LRANGE ends only, not full 553k)
| Queue | Sample | Classes | Oldest | Newest | avg bytes |
|---|---|---|---|---|---|
| forum-moderation | 204 | 100% AnalyzeForumPostJob | 2026-04-30 | 2026-08-23 18:47 | 987 |
| forum-security | 204 | FirewallActivityMonitor + BotActivityMonitor | 2026-04-30 | 2026-08-23 18:47 | 993 |
| collections | 204 | Duplicate / Health / Recommendation (~equal) | 2026-04-30 | 2026-08-23 17:43 | 1054 |
No reserved/delayed.
4. forum-moderation
UNLINK waiting + notify after dry-run.
Months-old AnalyzeForumPostJob has no moderation value. Command samples classes via LINDEX stride (no 553k LRANGE). Refuses if dispatch_enabled, reserved/delayed, or unexpected class (unless --force).
5. forum-security
UNLINK. Windowed monitors from April–August are expired. Same guards.
6. collections
A. SAFE TO DISCARD AND REBUILD. Jobs hold collectionId only; handle() loads current DB. Old queue is a 4-month backlog of the same refresh. Do not auto-delete; operator --execute collections when ready. Re-enable COLLECTIONS_V5_DISPATCH_ENABLED only after a Horizon consumer exists.
7. Presence stale ratio
Sample 100% stale (200/200). Estimated ~2.48M ghost members, ~272 MB.
8–9. Presence pruning
PresenceIndexPruner: SSCAN, pipeline EXISTS, Lua EXISTS record==0 then SREM. Never SMEMBERS, never DEL index.
Race: Lua is atomic vs SETEX. If a visitor is rewritten after SREM, they SADD again on next track.
Schedule: hourlyAt(33) if ONLINE_VISITOR_INDEX_PRUNE_ENABLED=true (default false). Avoids :02 snapshot, :15 rank, :21 leaderboards, :25 nova, nightly rec 02:00. First catch-up: raise --max-batches manually (200 batches × 500 ≈ 100k members/run → ~25 hours at hourly default).
10–11. Legacy prefix
Current process uses skinbase-database- / skinbase_horizon:. Obsolete skinbasenova-* ~30 MB. Command uses an unprefixed Predis client, UNLINKs only skinbasenova-database- and skinbasenova_horizon: keys, never current prefixes.
12. Expected reclaim (used_memory, RSS later)
| Stage | ~used_memory |
|---|---|
| forum-security | 5.5 MB |
| forum-moderation | 557 MB |
| legacy prefix | 25–30 MB |
| presence (after full prune) | ~272 MB |
| collections (if later) | 35 MB |
| Total if all stages | ~870 MB toward a small Redis |
RSS may stay ~177 MB until OS reclaims. No MEMORY PURGE.
13. Operator commands (not run here)
# A audit
php artisan tinker --execute="echo json_encode(['used'=>Redis::info('memory')['used_memory_human'],'mod'=>Redis::llen('queues:forum-moderation'),'sec'=>Redis::llen('queues:forum-security'),'col'=>Redis::llen('queues:collections'),'scard'=>Redis::scard('skinbase:presence:online:index')]);"
# B forum-security
php artisan skinbase:redis-cleanup-orphans forum-security --dry-run
php artisan skinbase:redis-cleanup-orphans forum-security --execute
# C forum-moderation
php artisan skinbase:redis-cleanup-orphans forum-moderation --dry-run
php artisan skinbase:redis-cleanup-orphans forum-moderation --execute
# D legacy
php artisan skinbase:redis-cleanup-legacy-prefix --dry-run
php artisan skinbase:redis-cleanup-legacy-prefix --execute --max-keys=500
# E presence (repeat)
php artisan skinbase:prune-online-visitor-index --dry-run --max-batches=5
php artisan skinbase:prune-online-visitor-index --execute --max-batches=200
# then set ONLINE_VISITOR_INDEX_PRUNE_ENABLED=true and config:cache
# F collections (optional later)
php artisan skinbase:redis-cleanup-orphans collections --dry-run
php artisan skinbase:redis-cleanup-orphans collections --execute
Abort: skip --execute. UNLINK is async; cannot restore payloads. Presence SREM is idempotent; live visitors re-SADD on next request.
14. Files
app/Support/Redis/OrphanQueueCleanup.php
app/Services/Traffic/PresenceIndexPruner.php
app/Console/Commands/RedisCleanupOrphansCommand.php
app/Console/Commands/PruneOnlineVisitorIndexCommand.php
app/Console/Commands/RedisCleanupLegacyPrefixCommand.php
config/traffic.php
routes/console.php
.env.example
tests/Feature/Redis/OrphanQueueCleanupTest.php
tests/Unit/Traffic/PresenceIndexPrunerTest.php
tests/Unit/Redis/LegacyPrefixFilterTest.php
docs/optimization-m8-redis-orphan-cleanup.md
15. Risks
- Sampled classes (400 LINDEX) might miss a rare job type;
--forcerequired then. queues:*:notifyUNLINK with the list is required for Laravel blocking pop leftovers.- Presence first pass is slow; enable scheduler only after a few manual batches.
- Predis fatals: do not LRANGE forum-moderation in tinker/Horizon UI until UNLINK.
- Collections UNLINK loses queued refresh intent; DB is source of truth.
No FLUSH, no DEL of huge keys from HTTP, no Horizon workers, no Redis restart.