feat: log convert metrics to benchmark local runs

[wietzesuijker]

Add basic observability to compare runs and debug slowdowns.

What changed

• Feat (CLI): Local Dask controls
  • --dask-mode {threads,processes,single-threaded}
  • --dask-workers, --dask-threads-per-worker
  • Optional --dask-perf-html <path> for Dask HTML performance report
• Metrics: Always write to <output>/debug/ (on success and failure)
  • dask_run_summary.json (params + wall time)
  • dask_metrics.json and timestamped dask_metrics_<run_id>_attemptN.json
  • Fields: wall_clock_s, workers, threads_total, tasks_observed, tasks_per_sec, compute_time_s_sum, transfer_time_s_sum, memory_used_bytes, memory_limit_bytes, optional spilled_nbytes, dashboard_link.
  • Task timing parsed defensively across Dask versions.

Dependencies

• Add: ipykernel, jupyter, bokeh (for testing/diagnostics/UI. I should probably ditch .ipynb support, it was useful for testing).

Compat

• Defaults mirror prior behavior (threads, 4×1)

Example

uv run eopf-geozarr convert --dask-cluster --dask-perf-html out/debug/dask-report-threads.html \  
 "https://objectstore.eodc.eu:2222/e05ab01a9d56408d82ac32d69a5aae2a:202505-s02msil2a/18/products/cpm_v256/S2B_MSIL2A_20250518T112119_N0511_R037_T29RLL_20250518T140519.zarr" \
  ./S2B_MSIL2A_20250518_T29RLL_geozarr.zarr --verbose \
  --groups measurements/reflectance/r10m measurements/reflectance/r20m measurements/reflectance/r60m
# Inspect OUT_DIR/debug/{dask_run_summary.json,dask_metrics.json}

Somewhat worryingly, this example took four attempts the last time I tried (it skips data already processed), the last attempt produced this dask_metrics.json

{
  "status": "ok",
  "run_id": "20250820-145816",
  "attempt": 4,
  "dask_enabled": true,
  "mode": "threads",
  "wall_clock_s": 416.0143037919988,
  "workers": 0,
  "threads_total": 0,
  "tasks_observed": 260,
  "tasks_per_sec": 0.6249785106667781,
  "compute_time_s_sum": 1444.2037508487701,
  "transfer_time_s_sum": 0.023849010467529297,
  "memory_used_bytes": 0,
  "memory_limit_bytes": 0,
  "dashboard_link": "http://192.168.0.12:8787/status"
}

the uv.lock diff looks kinda gross, I haven't used uv much, sorry for that.

[wietzesuijker]

Closing as stale