dwsmith1983 commented on PR #6041: URL: https://github.com/apache/datafusion-comet/pull/6041#issuecomment-5743601961
> Please include this deterministic eligibility check in the fallback and extend the new regression with `maxFields=1`. Done in eb57d7dcf. The guard now applies `WholeStageCodegenExec.isTooManyFields` to the aggregate's own schema and to each child's schema, the same two checks `CollapseCodegenStages` makes, and the regression runs `SUM(v)` with `spark.sql.codegen.maxFields=1` in both ANSI modes. Without the field-count check that variant fails the same way the sibling case did. The compatibility note now names only the runtime codegen compile fallback. > Please add a focused before/after scalar accumulator microbenchmark for common in-range decimal inputs and report time per row or throughput, including batch size and null rate. Added `sum_decimal_accumulator` to `native/spark-expr/benches/aggregate.rs`: the scalar accumulator over ten batches of 8,192 `DECIMAL(38,10)` in-range values, with no nulls and with a 10 percent null rate, in legacy and ANSI mode. Criterion medians per iteration of 81,920 rows, this branch against the previous `sum_decimal.rs` with the same bench: | case | before | after | |---|---|---| | no nulls, legacy | 67.0 us (0.82 ns/row) | 64.9 us (0.79 ns/row) | | no nulls, ANSI | 102.4 us (1.25 ns/row) | 77.2 us (0.94 ns/row) | | 10% nulls, legacy | 84.3 us (1.03 ns/row) | 78.2 us (0.95 ns/row) | | 10% nulls, ANSI | 107.8 us (1.32 ns/row) | 72.1 us (0.88 ns/row) | The wider checked add costs less than the per-row precision check and ANSI branch it replaces. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
