Skip to content

v1.134.0

Choose a tag to compare

@github-actions github-actions released this 22 Sep 09:12
d9d7824

Added

  • A run can publish its own frozen report. BENCHER_REPORT_DIR let a launcher
    outside the process ask for a complete report, but nothing let it ask for that report
    to be published: publish=True calls the in-process Publisher protocol, which only
    a benchmark's own source can wire up, so every out-of-process launcher had to find the
    frozen directory afterwards and run bencher publish against it. BENCHER_PUBLISH_STORE,
    BENCHER_PUBLISH_PREFIX, BENCHER_PUBLISH_HTTP_BASE and the optional
    BENCHER_PUBLISH_RECEIPT, BENCHER_PUBLISH_EXPIRY_DAYS and
    BENCHER_PUBLISH_MINIMUM_REMAINING_DAYS configure the same target the CLI takes, and
    bn.run(..., publication=PublicationTarget(...)) is the in-process form. Publication
    enables the complete export by itself, freezing to a temporary directory that is removed
    on success and kept, named in the raised PublicationFailed, on failure. A partially
    configured environment is an error rather than a silent skip.
  • Execution provenance can come from the environment. Execution has carried
    source_revision, workflow, workflow_run, lane and attempt since complete
    reports existed, and Execution.start documents that the launcher is the thing that
    resolves them — but the only way to supply them was execution_context(...), inside
    the benchmark process. A launcher that is a different process (a CI job, or a
    wrapper that runs python bench.py) knows the revision and had no way to say so, so
    every out-of-process run froze a report.json whose provenance was entirely null.
    BENCHER_SOURCE_REVISION, BENCHER_WORKFLOW, BENCHER_WORKFLOW_RUN,
    BENCHER_LANE, BENCHER_ATTEMPT and BENCHER_DISPLAY_LABEL fill exactly those
    fields. An explicit argument still wins, an empty variable is an absent one, and
    bencher still runs no subprocess and infers no revision of its own.

Changed

  • bencher.publishing.Publisher is now CompleteReportPublisher. Two public
    classes were called Publisher: the protocol in bencher.bench_report, exported as
    bn.Publisher, which a run calls in-process with a live report, and the concrete
    class that commits an already-frozen report directory to an object store. The two do
    different jobs at different seams, and sharing a name made every discussion of "the
    publisher" ambiguous and every grep useless. The protocol keeps its name, since it
    is the one downstream projects implement; the concrete class is renamed, with no
    alias left behind — a name that still resolves is a collision that still exists.

Fixed

  • GcloudStore authenticated under whatever interpreter the caller had. env
    defaulted to a copy of os.environ, which _auth passed to
    gcloud auth print-access-token. gcloud ships its own interpreter, so an inherited
    PYTHONPATH — a sourced ROS 2 workspace, a venv — put a foreign standard library
    ahead of its own and it failed to start, and the store reported "gcloud
    authentication failed", pointing at credentials that were fine. PYTHONHOME is
    dropped alongside it: gcloud's POSIX wrapper unsets that one itself, but neither
    variable is ever right for a subprocess that is not this Python.