marin: Skill for Claude Code

Query Finelog

Read lib/finelog/OPS.md for access and query behavior. Read lib/iris/OPS.md under Stats Namespaces for Iris namespace meanings.

Discover before querying; do not assert remembered columns:

uv run finelog namespaces <deployment>
uv run finelog schema <deployment> <namespace>
uv run finelog query <deployment> --format table <<'SQL'
<bounded SQL using schema-confirmed columns>
SQL

finelog query reads SQL from stdin when the positional SQL argument is omitted.

Use marin for the federated view and a regional deployment for peer-local truth or recent rows that may not have forwarded. Preserve cluster and full process/label identity until after per-series delta calculations.

Bound the native time key. Keep telemetry_v1.timestamp_ms predicates numeric. Treat current snapshots as values, imported Prometheus counters as cumulative snapshots with LAG and reset handling, and native Rigging counters as deltas to SUM directly.

Never reset or change a shared namespace during diagnosis. Return the deployment, namespace, time window, query, series semantics, and any forwarding or retention caveat.

Examples

Confirm every schema before adapting an example. Angle-bracket values are placeholders.

Iris task memory by half-hour

Adapted from Echo wiki 230. The dashboard task ID includes the final task index. Select or group by attempt_id after retries.

SELECT date_bin(INTERVAL '30 minutes', ts,
                TIMESTAMP '1970-01-01 00:00:00') AS bucket_start_utc,
       count(*) AS samples,
       round(min(memory_mb) / 1024.0, 1) AS min_gib,
       round(median(memory_mb) / 1024.0, 1) AS median_gib,
       round(max(memory_mb) / 1024.0, 1) AS max_gib,
       round(max(memory_peak_mb) / 1024.0, 1) AS attempt_peak_gib
FROM "iris.task"
WHERE task_id = '/user/job/task'
  AND attempt_id = 0
GROUP BY bucket_start_utc
ORDER BY bucket_start_utc

memory_mb is sampled current memory; memory_peak_mb is the attempt's cumulative peak. Values are MiB despite the names. count(*) exposes partial buckets and gaps. Query the regional deployment if recent hub rows appear incomplete.