kubernetes-sigs / inference-perf / 34504974826

80%

push

github

feat(datagen): support the filter config key in weka replay (#773)

WekaTraceReplayConfig had no `filter` field, so configs are strict and a
`filter:` key under weka_trace_replay failed validation outright.

Add it and apply it to all three trace sources, reusing the OTel
datagen's _compile_filter. Weka traces record only per-request lengths,
so derive the aggregates a filter needs: max_tokens, total_tokens and
num_turns. A failing expression raises rather than being skipped as bad
input data, so a typo'd key cannot quietly empty the corpus.
num_dataset_entries now caps accepted traces rather than lines scanned,
so a filter narrows which traces are picked instead of how many.

---------

Signed-off-by: Angelo Ruocco <ang@zurich.ibm.com>

61 of 65 new or added lines in 4 files covered. (93.85%)

9360 of 11864 relevant lines covered (78.89%)

0.79 hits per line