docs: record gitnostr startup sync deployment soak

Record the confirmed 56889bf deployment and fifteen-minute observation window, including rapid relay startup, terminal cleanup, and the remaining filter-size and live-budget issues. Distinguish measured startup behavior from unproven long-term memory stability. Validation used deployed build metrics, journals, service state, and public read probes; this adds no runtime changes. Assisted-by: GPT-6
2026-10-05 15:08:24 +00:00 · 2026-09-23 15:38:09 +00:00
parent 56889bfc1d
commit 2d63c5871c
1 changed files with 47 additions and 0 deletions
@@ -0,0 +1,47 @@
+# gitnostr.com sync startup soak — 2026-09-23
+
+Deployed ngit-grasp `56889bfc1df28282e1edfa229ee0029584fc9353` to
+nixos-fi1's gitnostr.com instance with `deploy . --remote-build --skip-checks`.
+Activation was confirmed at 15:21:50 UTC. The deployment pin is nixos-fi1
+commit `44f3c21`. Other GRASP instances retained their original start times.
+
+## Observations
+
+The final sample at 15:37:24 UTC covered 15 minutes 34 seconds:
+
+- 126 connected relays after 49 seconds; 176 at approximately 14 minutes.
+  Final sample: 168 connected / 305 tracked, with 137 fully caught up,
+  26 syncing history, and 5 connected with historic-sync failures.
+- No service restarts, panics, nonzero persistence-error windows, or
+  `[PARSE_FAIL]` messages. Public NIP-11 and WebSocket repository reads passed.
+- 26 transient watchdog expirations and 66 incomplete-batch retirements.
+  Timeout cleanup appeared promptly in the actor while other relays continued
+  syncing. Not every watchdog needs a batch-retirement log: a batch may already
+  have been removed by another terminal or disconnect.
+- 173,153 hydration deliveries, predominantly duplicates; 85 invalid outcomes
+  and no retryable rejection or persistence-error outcomes in the sampled
+  hydration metrics.
+- Service cgroup memory grew from 1.90 GiB at 49 seconds to 3.03 GiB at the
+  final sample. Discovery was still active. This soak does not establish a
+  long-term memory plateau or demonstrate a leak.
+- Startup integrity reported the same 60 unique PR references requiring manual
+  inspection as the previous deployment; no new references appeared.
+
+## Remaining issues
+
+1. Forty `filter item too large` refusals across nos.lol, relay.damus.io,
+   and strfry.bonsai.com. The refusal classifier does not recognize this
+   deterministic filter incompatibility, so incomplete-batch retirement allows
+   unchanged work to be retried. Add classification/cooldown coverage before
+   considering the soak completely clean.
+2. nos.lol twice failed complete live extension because it needed 20 and then
+   22 slots against 18 usable slots. Investigate budget-aware regrouping or
+   fallback; a connected transport alone does not prove complete coverage.
+3. Continue observing memory beyond startup and the existing integrity findings.
+
+The global startup stall is substantially improved, and timeout/invalid-PR
+behavior matches the intended fixes. The deployment remains active. No further
+application changes were made during this soak.
+
+Raw local evidence is under `work/soak-56889bf/` (gitignored): timestamped
+metrics/state snapshots, current and previous startup journals, and read probes.