[jwrosewell]

The Pipeline could not describe the Evidence it accepts, and usage sharing and on-premise Device Detection treated an Evidence key spelled in other letters as a different key. The Pipeline now gives its accepted Evidence, and both read keys in any letters.

Problem

1. Pipeline (pipeline.py at 6ad5ae2d) had no way to give the Evidence keys it can use, though each Flow Element can.
2. Usage sharing hashed the values for its repeat check under the keys as spelled, so the same values under keys in other letters were shared again, and it named each record element by the key's prefix as spelled.
3. On-premise Device Detection gave the native evidence each key as spelled, and the native evidence matches a key's prefix with case, so Header.User-Agent was ignored.

What the specification says

advertize-accepted-evidence.md line 5 at 706de1f0: "All Flow Elements and Pipelines MUST be capable of programmatically describing the Evidence keys that they can make use of."

evidence.md line 10 at 706de1f0: "Keys MUST be case-insensitive and will usually be split into <prefix>.<field>."

What changed

1. c62ef9f9 adds the tests: test_pipeline_evidence_key_filter.py in fiftyone_pipeline_core (the Pipeline accepts what any element accepts, reads its elements' filters when asked, and filters Evidence to what it accepts), two tests in test_share_usage_record_fields.py (a repeat with keys in other letters is shared once, and Evidence is recorded under its prefix in lower case), and test_evidence_key_case.py in on-premise Device Detection.

2. 34dcb7a2 adds Pipeline.get_evidence_key_filter(), whose filter accepts a key when any element's filter does, reading each element's filter when a key is checked, as an element's keys can change once its data or the cloud service's answer arrives. Usage sharing hashes lower case keys for its repeat check and names each record element by the prefix in lower case. On-premise Device Detection gives the native evidence each key with its prefix in lower case.

3. 0c514e76 closes each flow data the Evidence key case test makes.

Before you upgrade

• A usage sharing record names each element by the Evidence key's prefix in lower case, so Header.X is recorded under header, where it was recorded under Header.
• On-premise Device Detection uses Evidence whose key's prefix is in other letters, such as Header.User-Agent, where it ignored it.

How it was verified

• Red, full build 43617993890 on c62ef9f, the tests alone. In all ten jobs the three Pipeline Evidence tests failed in the core tests ("3 failed, 481 passed") and the two usage sharing record tests in fiftyone_pipeline_engines_fiftyone ("2 failed, 95 passed"). In on-premise Device Detection, test_a_key_in_any_case_reaches_the_native_engine in test_evidence_key_case.py failed for Header.User-Agent, HEADER.USER-AGENT and Query.User-Agent at line 61 ("3 failed, 77 passed"), so this run is that test's red too.
• Green, full build 43658210022 on the head 0c514e7, from rc 28e5bf2, is green. Its ten Build and Test jobs, Python 3.10 to 3.14 on Ubuntu and Windows, each passed, with test_pipeline_evidence_key_filter.py (3 tests), test_share_usage_record_fields.py (9) and test_evidence_key_case.py passing in each. The 11 Setting 'PullRequestSha' lines all name 0c514e7, and Configure printed "Left out 5 macOS entries". macOS was left out of this draft's full build under the 9 November decision. The rc head is built with macOS before it merges into main.
• An earlier green build, 43654951393 on 34dcb7a, failed in all ten jobs only at the teardown of test_evidence_key_case.py, whose flow data stayed open ("6 native object(s) still alive"). 0c514e76 reads each value inside a with block.

What it releases

fiftyone_pipeline_core, fiftyone_pipeline_engines_fiftyone and fiftyone_devicedetection_onpremise and their dependants, which Monday's release plan releases already.

What remains

Nothing.