You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
bug(reporting): a Core seller cannot carry conditional metrics under full coverage #1239
On 8.0.0b18, two rules combine so that a seller cannot report a metric that applies to some inventory and not other inventory -- the case #1170 was written for ("GAM may provide impressions, spend, and clicks while video-completion or viewability metrics are unsupported ... for the same media buy").
Conformance admits only exact metrics._validate_request_against_capabilities (adcp/reporting/conformance.py) raises CAPABILITY_MISMATCH: requested metric 'clicks' is not exact in the offering for any requested metric whose offering support is partial.
An unsupported cell demotes the constituent. Per InlineFetchResult.cell_availability, "mixed availability promotes the constituent to partial". A configuration with coverage_requirement: full then fails the slice with PARTIAL_RESULT.
So a seller that requests a conditional metric either fails conformance (rule 1) or, if it requests it anyway, fails every slice for inventory where the metric does not apply (rule 2). A display buy has no completed views, so on a typical mixed book almost every slice fails.
Runtime and conformance disagree
The runtime producer does not enforce rule 1: ReliableReportingService built with ReportingAccountContext(requested_metrics=(... "clicks", "completed_views" ...)) happily requests partial metrics. Only the conformance harness rejects them. An adopter who tests with the service and not the harness ships something the source contract says is invalid.
Reproduction (ESA, adcp 8.0.0b18)
Offering declares impressions/spendexact, four counters partial.
Context requests all six; configuration coverage_requirement: full.
Fetch returns rows for a display buy with completed_views absent and cell_availability={buy: {"completed_views": MetricEvidence.unavailable(...)}}.
Worker turn: reporting slice failed ... code=PARTIAL_RESULT retry=retryable, no revision.
What we did
Request only the exact pair and carry the conditional counters as extra row columns when the ad server measured them -- no control total, no availability claim. That works, but it means per-metric evidence (#1170) is never used, and buyers get the counters without any statement of whether a missing one means "not applicable" or "missing".
Ask
Decide the intended model and make runtime and conformance agree. One coherent option: a metric whose cell is unsupported (inapplicable to that inventory) does not demote coverage -- only missing/delayed do -- and partial offering metrics are requestable, with their per-cell evidence mandatory. That is what makes #1170 usable under full coverage.
Classification: Bug Bucket(s): reporting, validation (neither label exists in this repo — label gap flagged in run summary) Status: ready-for-human
What the experts said:
ad-tech-protocol-expert: Non-breaking fix — partial offering support is unreachable under the exact-only conformance gate; unsupported cells are structural "inapplicable" signals, not missing-data signals, and should not demote coverage.
code-reviewer: Core two fixes are safe to automate; the "per-cell evidence mandatory" extension is underspecified. Edge case: a constituent whose all cells are unsupported has a zero-denominator in the corrected roll-up and needs explicit handling.
My take: Both root-cause bugs are confirmed — the conformance.py:224 exact-only gate and the inline_source.py:1140–1147 constituent demotion logic are client-side policy mistakes. Before implementing, two design questions need resolution from @bokelley: (1) what does "per-cell evidence mandatory" mean concretely for partial metrics — which fields, which error code, enforced in conformance or runtime or both? (2) what coverage status should an all-unsupported constituent carry?
Once the model is decided, the implementation scope is:
conformance.py:224–229: extend admission to partial metrics (and add evidence check when specified)
inline_source.py:1140–1147: treat a constituent as present when all cells are in _AVAILABLE ∪ {unsupported}; add zero-denominator guard for all-unsupported constituents
New test in tests/conformance/reporting/test_inline_cell_availability.py covering full-coverage slice with mixed exact/partial metrics and unsupported cells
Activity
bokelley commentedon Sep 28, 2026
Triage
Classification: Bug
Bucket(s): reporting, validation (neither label exists in this repo — label gap flagged in run summary)
Status: ready-for-human
What the experts said:
partialoffering support is unreachable under theexact-only conformance gate;unsupportedcells are structural "inapplicable" signals, not missing-data signals, and should not demote coverage.unsupportedhas a zero-denominator in the corrected roll-up and needs explicit handling.My take: Both root-cause bugs are confirmed — the
conformance.py:224exact-only gate and theinline_source.py:1140–1147constituent demotion logic are client-side policy mistakes. Before implementing, two design questions need resolution from @bokelley: (1) what does "per-cell evidence mandatory" mean concretely for partial metrics — which fields, which error code, enforced in conformance or runtime or both? (2) what coverage status should an all-unsupportedconstituent carry?Once the model is decided, the implementation scope is:
conformance.py:224–229: extend admission topartialmetrics (and add evidence check when specified)inline_source.py:1140–1147: treat a constituent aspresentwhen all cells are in_AVAILABLE ∪ {unsupported}; add zero-denominator guard for all-unsupportedconstituentstests/conformance/reporting/test_inline_cell_availability.pycovering full-coverage slice with mixed exact/partial metrics andunsupportedcellsTriaged by Claude Code. Session: https://claude.ai/code/session_01Uknm2rn2T5eDk11oP73VZa
Generated by Claude Code