Query Finelog logs and telemetry for Iris tasks, workers, and vLLM metrics.
npx skills add https://github.com/marin-community/marin --skill query-finelog
Please help me install this Agent Skill. Skill: query-finelog Source: https://github.com/marin-community/marin/tree/main/.agents/skills/query-finelog Command: npx skills add https://github.com/marin-community/marin --skill query-finelog
Diagnosing distributed training and inference issues requires querying Finelog logs and telemetry across federated and regional deployments, but incorrect schema assumptions, counter semantics, or missing forwarded rows lead to wrong conclusions.
Ask the assistant to query Finelog for the memory usage of a specific Iris task over the last day, discovering the schema first and summarizing results by time bucket.
High-intent search queries and answers about installing and using this skill.
First discover available namespaces with 'uv run finelog namespaces <deployment>' and confirm columns with 'uv run finelog schema'. Then run bounded SQL via 'uv run finelog query', reading SQL from stdin when the positional argument is omitted.
Imported counters carry source_temporality 'cumulative_snapshot', so use LAG partitioned by the full series identity to compute per-interval deltas and discard reset intervals where value decreases. Native Rigging counters are already deltas and can be summed directly.
Missing hub rows usually indicate cross-cluster forwarding delay. Query the regional deployment with the exact attempt-suffixed log key; if regional rows exist but hub rows do not, forwarding has not completed yet.
Use the marin deployment for the federated view across clusters, and a regional deployment for peer-local truth or recent rows that may not have forwarded yet. Preserve the cluster column until after per-series calculations.
You must bound the native time key and keep timestamp_ms predicates numeric, and never assert remembered columns without schema discovery. Never reset or change a shared namespace during diagnosis, and report forwarding or retention caveats with results.
grafanaGenerate RED metrics, RUM data, and LLM telemetry for Grafana Cloud.
openaiCapture and interpret Android CPU, frame, and memory profiling traces via adb.
open-thoughtsManages the full lifecycle of Iris evaluation jobs via the command line.
vibercoder79Generate and update Grafana Cloud dashboards, panels, and PromQL alert rules.
