englefly commented on PR #68340:
URL: https://github.com/apache/doris/pull/68340#issuecomment-5774138684
Both new P1 cases are addressed in `ee1bf37378b`.
**Representative selection is now constrained by the source value domain.**
The pushed filter is the one of the deepest builder of the identity group,
which is a conservative substitute only when it prunes at most as many rows as
every other filter of the group. `keepsSourceValue()` therefore requires that
the source expression keeps its values between the deepest builder and the
shallowest one:
* every node in between must be in `SPJ_PLAN`. This is the all-node check
the original code intended (its SPJ line was commented out) and it rejects the
`Repeat` and the aggregate of `GROUP BY GROUPING SETS` — the case of comment 1
— as well as set operations and nested loop joins;
* a join which null extends the child the source comes from is rejected: the
left child of a right outer join (and of an asof right outer join) — the case
of comment 2. Tracking the child the source travels through is what
distinguishes it from the preserved right child, which stays allowed.
Verified on a local cluster for the grouping-sets shape (comment 1): `t c1
JOIN (SELECT x FROM (SELECT c2.k AS k2, b.x AS x FROM t c2 JOIN b ON c2.k <=>
b.x) z GROUP BY GROUPING SETS ((x), ())) g ON c1.k <=> g.x` with a NULL in `t`
and `b.x = {3}` returns the row which matches the synthesized NULL with the
NULL of the CTE, and the runtime filters stay on the branch sinks — the
producer scan gets none (the plan contains `VAGGREGATE` and `VREPEAT_NODE`
between the two builders). The same shape is now the `shared_cte_grouping_sets`
case of the regression suite.
Tests: `RuntimeFilterTest` 42/42 (the two new ones are
`testDoNotPushRuntimeFilterWhichCrossesAValueSynthesizingNode` for the
repeat/grouping-sets shape, and
`testSourceValueIsNotKeptWhenAnOuterJoinNullExtendsIt`, which drives
`keepsSourceValue` with a right outer join whose source comes from the null
extended left child versus the preserved right child). The regression suite
passes with six ordered cases. TPC-DS q95 with `runtime_filter_type=12` still
installs the same four runtime filters on the inner scan of its CTE body with
an unchanged result, and the reported case still returns 4 rows instead of 0.
--
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]
