Skip to content

k4bench.regression.lineage

k4bench.regression.lineage

How platforms relate to each other over a migration.

Two independent facts, deliberately kept apart: which platform a young one may seed its baseline from, and when a platform stops being expected to run. A predecessor is not necessarily retired — a future migration may keep both compilers running in parallel — so neither map may be inferred from the other.

Regression history is scoped to (detector, platform, sample) and stays that way — a platform is an identity, not a label for "the current one". The cost is a cold start: a new platform's first :data:~k4bench.regression.engine.MIN_BASELINE_RUNS nights have no baseline to be judged against.

A platform may therefore name a predecessor whose measurements it borrows as baseline points only, while it has too few of its own. One-way and one hop; the new platform keeps its own directory, metadata, provenance and history, and no verdict is ever issued for a borrowed point.

BaselineSeed dataclass 

BaselineSeed(platform: str, history: DataFrame)

One series' inherited baseline points, and the platform they came from.

history has the columns :func:~k4bench.regression.engine.evaluate_series walks (run_id, run_date, value, reliable) — the predecessor platform's rows for the same series.

baseline_predecessor 

baseline_predecessor(platform: str) -> str | None

The platform platform may seed its baseline from, or None.

One hop: a predecessor's own predecessor is not reached, since a seed has to be a series measured on comparable software. A chain in the map is legal and expected — gcc17 → gcc16 → gcc14 resolves gcc17 to gcc16 and stops, and the older entry stays for backfills of gcc16's own early nights.

Source code in k4bench/regression/lineage.py 
38
39
40
41
42
43
44
45
46
47
def baseline_predecessor(platform: str) -> str | None:
    """The platform *platform* may seed its baseline from, or ``None``.

    One hop: a predecessor's own predecessor is not reached, since a seed has
    to be a series measured on comparable software. A chain in the map is
    legal and expected — ``gcc17 → gcc16 → gcc14`` resolves gcc17 to gcc16 and
    stops, and the older entry stays for backfills of gcc16's own early nights.
    """
    predecessor = BASELINE_PREDECESSORS.get(platform)
    return None if predecessor == platform else predecessor

platform_retired 

platform_retired(platform: str, night: str) -> bool

Whether platform had stopped being benchmarked by night.

Both dates are ISO YYYY-MM-DD, which compares correctly as text.

Source code in k4bench/regression/lineage.py 
76
77
78
79
80
81
82
def platform_retired(platform: str, night: str) -> bool:
    """Whether *platform* had stopped being benchmarked by *night*.

    Both dates are ISO ``YYYY-MM-DD``, which compares correctly as text.
    """
    retired_on = PLATFORM_RETIREMENTS.get(platform)
    return bool(retired_on and night >= retired_on)