prop/docs/runbook_propagation_pipeline.md

Propagation pipeline runbook

The hourly propagation grid flows across two runtimes and a shared NFS volume. This doc names every failure mode at that seam and says how the system recovers.

prop-grid-rs (worker nodes)      Postgres (talos5 — 10.0.15.30)  prop pods (node1/2/3)
 ├─ fetch GRIB2 (NOMADS)         ├─ grid_tasks (SKIP LOCKED)     ├─ ScoreCache (ETS)
 ├─ decode                       ├─ LISTEN/NOTIFY channel        ├─ NotifyListener
 ├─ score 22 bands × 92k         │    "propagation_ready"        ├─ ScoreCacheReconciler (60s)
 ├─ write /data/scores/.../.prop │                               └─ PubSub → LiveView
 └─ NOTIFY propagation_ready ────┘

Primary refresh path (happy case)

1. Rust completes {band, valid_time} compute and writes the .prop file via atomic rename. (Readers also accept legacy .ntms files with NTMS magic during a rolling deploy.)
2. Rust emits NOTIFY propagation_ready '<iso valid_time>'.
3. NotifyListener on each Elixir pod receives it and calls Propagation.warm_cache_and_broadcast/2 which reads the file once and PubSub-broadcasts the payload to every peer.
4. Each pod's ScoreCache receives the message and writes into its local ETS table. /map clients are notified via the "propagation:updated" topic.

Latency budget end-to-end: < 2s.

Failure modes

FM1 — NOTIFY dropped on reconnect

Symptom. {band, valid_time} files land on disk but ScoreCache stays cold until the first lazy miss.

Why. Postgres queues LISTEN notifications per-connection and discards any that are unread on disconnect. A DB restart, a network blip, or the Postgrex.Notifications connection hitting its idle timeout all drop whatever queued notifies were in flight.

Recovery. ScoreCacheReconciler sweeps /data/scores every 60s per-pod and warms missing {band, valid_time} pairs from disk into local ETS. Each pod reconciles independently — no cluster messaging required.

FM2 — NotifyListener not running locally

Symptom. One pod is cold while peers are warm. /map load time on that pod is ~150ms slower than average until the reconciler runs.

Why. The listener isn't currently part of the supervision tree (see lib/microwaveprop/application.ex). When it is re-added, early startup or Postgrex.Notifications boot failure can leave it unsubscribed for several seconds while Rust writes land.

Recovery. Same as FM1 — reconciler catches up within one tick.

FM3 — NFS stale read racing Rust write

Symptom. ScoresFile.read/2 returns {:error, reason} when the reader opens the file exactly between Rust's rename(2) of the .tmp.<uniq> file and the fsync landing on the NFS server.

Why. The rename is atomic on the same NFSv4 filesystem, but the NFS client cache can briefly serve a stale directory listing or a no-longer-existing path.

Recovery. Reconciler calls ScoresFile.read/2 directly and pattern-matches {:error, _} -> :error, logging at debug and skipping the key. The next tick re-reads the file after NFS consistency settles. Lazy-miss path (Propagation.scores_at_fetch/3) has historically surfaced this as a LiveView error — the reconciler warms the cache before that miss happens.

FM4 — Cluster partition splits ScoreCache broadcast

Symptom. Pods on one side of a cluster partition have up-to-date cache; pods on the other side don't, even though both can read NFS and Postgres.

Why. ScoreCache.broadcast_put/3 fans the payload out over PubSub. A libcluster partition stops the message reaching partitioned peers.

Recovery. Each pod's reconciler reads from NFS directly, so it converges regardless of PubSub reachability. Partition does not cause divergent scores — only transient latency until the next sweep.

FM5 — Rust worker dead

Symptom. grid_tasks rows with status = 'queued' accumulate; no new .prop files land; /map serves the last successful run's scores.

Why. Deployment bug, Rust panic, OOM on a worker node, or disk-full on /data.

Recovery. Manual. kubectl -n prop logs deploy/prop-grid-rs for cause. The Elixir side is read-only against the grid — it does not regenerate scores from within the BEAM since the extraction in 65693ed. lib/microwaveprop/workers/propagation_grid_worker.ex still exists as a seed worker: the hourly cron fires it with empty args and it inserts 19 grid_tasks rows (f00 + f01..f18) for Rust to drain. All fetch/decode/score helpers moved to rust/prop_grid_rs/src/pipeline.rs; the Elixir module is retained only as the cron entry point, not as a fallback compute path.

Monitoring signals

• avg_over_time(beam_memory_total_bytes[1h]) on prop pods — steady climb above 1 GiB means the cache isn't pruning (see ScoreCache.prune_older_than/1 in NotifyListener).
• Reconciler log line "ScoreCacheReconciler: warmed N" — if N is consistently non-zero every minute, NOTIFY is dropping and NotifyListener wants investigation.
• grid_tasks queue depth in Prometheus — should drain to 0 between hourly cron ticks.

Test coverage

• test/microwaveprop/propagation/score_cache_reconciler_test.exs covers FM1 and FM3 (missing-file skip).
• Partition recovery (FM4) is untested in CI — requires multi-node cluster and libcluster orchestration.