## Summary
Follow-up to #1215 and #1220. Both of those addressed *individual* slow
queries; this PR addresses *concurrency* — under today's scraper load
the cursor query is fast on average (mean 42ms) but the connection pool
saturates and queue depth blows up at the Go HTTP layer.
## Diagnosis
`pg_stat_statements` (added in #1215) made this possible to see:
```
max_ms mean_ms calls template
10,755 41.8 94,009 plain cursor pagination, no filter
7,277 214.7 14,770 ILIKE substring filter
7,638 330.8 7,656 ILIKE substring + is_latest filter
```
Mean times are healthy. But `max_exec_time` of 7–10s on the cursor
query, combined with sustained ~15 req/s from scrapers (ServiceNow's
148.139.x.x range, anonymous `node` user-agent, etc) saturates
`MaxConns=30 × 2 pods = 60`. New requests queue at the Go HTTP layer;
nginx-side p99 hits 35s; eventually scrapers time out at 60s, retry, and
amplify the queue.
Today's two ongoing alerts are both this pattern:
- 17:35–17:40 UTC `Availability dropped below 95%` — 4,526 GET requests
in 5 min, hundreds of 504s
- 18:17 UTC `Publish Endpoint Latency` re-fire — same scraper
concurrency dragging the publish path
Critically, the symptoms today were also visible during yesterday's
incident, but yesterday's broken cursor (#1215) was the dominant cause.
After #1215 the cursor is fast individually; concurrency now becomes the
next bottleneck.
## Changes
### pgxpool (`internal/database/postgres.go`)
| Setting | Before | After |
|---------|-------:|------:|
| `MaxConns` | 30 | **60** per pod |
| `MinConns` | 5 | **10** per pod |
| `MaxConnIdleTime` | 30 min | (unchanged) |
| `MaxConnLifetime` | 2 h | (unchanged) |
2 pods × 60 = 120 total app connections. Cuts queue depth roughly in
half at current scraper load.
### PG cluster (`deploy/pkg/k8s/postgres.go`)
- `max_connections: 100 → 200` — required to support the larger pool.
120 (app) + ~10 (PG internals: autovacuum, replication, admin) + 70
headroom.
- Added explicit `resources:` block — previously the pod had no resource
limits, making node-level OOM behaviour unpredictable.
| | Request | Limit |
|---|--------:|------:|
| memory | 512Mi | **4Gi** |
| cpu | 200m | 1500m |
## Resource budget
| Node | Now | After |
|------|----:|------:|
| dy89 (PG node) | 39% mem (~2.4 GiB / 6 GiB) | ~65% mem worst-case
(~3.9 GiB) |
| 2yxm | 36% mem | unchanged |
Both nodes well within capacity. CPU usage <50% on both, plenty of
headroom.
PG worst-case memory math:
- 200 conns × ~15 MB per backend = 3 GiB
- + `shared_buffers` 128 MiB
- + `maintenance_work_mem`, OS overhead, etc.
- ≈ 3–4 GiB total
## Deployment caveat
`max_connections` is a postmaster-level setting → CNPG triggers a PG
restart on the change. Same in-pod restart shape as the
pg_stat_statements deploy yesterday — registry pods see ~30s of DB
unavailability, covered by v1.7.1's retry-with-backoff (8 attempts, 1→8s
capped). One registry pod may bounce once before recovering, like
yesterday.
**Time the merge for a low-traffic UTC window.**
Order of operations matters within the deploy itself:
1. Pulumi applies the new CNPG spec → PG restarts with
`max_connections=200`
2. Rolling deploy of registry pods picks up `MaxConns=60` config
3. New conns are accepted up to 200 limit
Pulumi's standard ordering does step 1 before step 2 in this scenario
(Pulumi resource graph: CNPG cluster precedes Deployment). If for any
reason it doesn't, the worst case is `too many connections` errors
during a small window — Self-correcting once the rollout completes.
## Test plan
- [x] `go build ./...` clean for app + deploy
- [x] `make lint` clean
- [x] `go test -race ./internal/database/...` green
- [ ] On merge: staging deploy applies the spec change cleanly; PG
restarts; pool reaches 60 max
- [ ] On prod deploy: same; verify no `too many connections` errors
during the rollout window
- [ ] After deploy: query `SHOW max_connections` returns 200; query
pg_stat_statements after a scraper burst, confirm `max_exec_time` for
cursor query no longer hits 10s
## Out of scope
- **ILIKE substring search** (`server_name ILIKE '%foo%'`) is
unindexable. Three pg_stat_statements variants run with means 141–333ms
and max 4–7s. Worth replacing with a `pg_trgm` GIN index or full-text
search in a separate PR.
- **Per-IP nginx rate limiting** — defends against scraper retry storms
regardless of pool size.
- **`Cache-Control` on `/v0/servers`** — let nginx absorb
scraper-repeated cursors.
- **Pre-existing `superfluous WriteHeader` warnings** — separate Huma
framework issue.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
736d376858