kazantsev-maksim commented on PR #5042:
URL:
https://github.com/apache/datafusion-comet/pull/5042#issuecomment-5655394266
Thanks @andygrove — agreed on both counts, both are addressed in the
current head.
1. Residual state in the scratch buffers: the scratch path now fully
re-initializes on every call (`clear()` + `resize(len, default_val)`), so
no
cell can carry a value from a previous row. There is a regression test
that
runs a whole column in one `invoke` and asserts forward-order ==
reversed-order
results, plus the small-threshold / long-string shape (`"a"`, `"bb"`,
`1`).
2. Non-ASCII benchmark coverage: added `non-ascii` and `mixed` datasets to
the
bench alongside the existing `ascii` one. On the current head, non-ASCII
is
*not* slower — it improves by 2.8-7.3% (p < 0.05), i.e. the extra
`is_ascii()`
scan is paid back by scratch reuse. Full numbers are in the PR
description.
One caveat I want to be explicit about: the `all_null` row moves ~-31% on all
three datasets, but that case does no DP and no scratch reuse at all, so I
don't
attribute that delta to this change.
Also worth noting for context: the thread-local buffers are now capped
(`MAX_RETAINED_CAPACITY`), and rows larger than the cap allocate call-local
vectors instead, which is what bounds the retained TLS memory from the
earlier
P2 finding.
--
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]
