Context¶

Use the context API to stamp a run with metadata: git SHA, dataset identifier, hardware tag, anything you’d later want to filter on.

import mew

mew.update_context(
    git_sha="abc123",
    dataset={"name": "uniform-1k", "size": 1000},
)

The snapshot is taken when mew.run() starts, so concurrent mutations don’t affect an in-flight run. Reporters receive it under ctx["custom"]:

{
  "context": {
    "date": "2026-05-19T10:00:00+00:00",
    "host_name": "laptop",
    "session_id": "01975f2e-9c40-7b31-a1d4-8f0e2c5b7a90",
    "session_tag": "v0.3.1-12-gabc123",
    "custom": {
      "git_sha": "abc123",
      "dataset": {"name": "uniform-1k", "size": 1000}
    }
  },
  "benchmarks": [...]
}

Session identity¶

Every mew.run() invocation is one session and gets a generated session_id (a time-ordered UUIDv7), persisted in the context block (JSON/JSONL) or as per-row columns (Parquet). This keeps runs distinguishable when several land in one archive — even two runs in the same wall-clock second.

session_tag is the human label next to it: pass mew run --session-tag before (or session_tag= on mew.run()). When no tag is given, the CLI derives one from git describe --always --dirty; opt out with auto_session_tag = false under [tool.mew]. Note --session-tag labels the run’s output — it is unrelated to -t/--tag, which selects which benchmarks run.

Dotted keys¶

Dotted keys produce nested dicts — handy for SQL drill-downs against the Parquet sink:

mew.set_context("dataset.size", 1000)
mew.set_context("dataset.kind", "uniform")
# -> {"dataset": {"size": 1000, "kind": "uniform"}}

As an example, a DuckDB query on the resulting run data might look like this:

SELECT json_extract(custom, '$.dataset.size') AS size, real_time
FROM 'results.parquet';

API surface¶

mew.set_context() — set one key (dotted keys produce a nested struct).
mew.update_context() — set many keys at once.
mew.get_context() — snapshot the current state (deep-copied).
mew.clear_context() — wipe all entries.

JSON-friendly values are strongly preferred — the JSON and Parquet sinks serialize with str() by default, which loses information.