englefly commented on PR #69111:
URL: https://github.com/apache/doris/pull/69111#issuecomment-6210468926
Implemented the framework variant (`45a495f18b6`), as agreed: the consumer
maps are refreshed instead of keeping the producer outputs stable.
`RewriteCteChildren.visitLogicalCTEAnchor` now compares the producer outputs
before and after the rewrite of the producer body and refreshes both output
maps of every consumer of that CTE:
```text
producer outputs [k#0, v#1, cnt#3] -> [k#10, v#1, cnt#3]
consumer maps c2p={k#4=k#0, ...}, p2c={k#0=[k#4], ...} =>
c2p={k#4=k#10, ...}, p2c={k#10=[k#4], ...}
```
The correspondence is derived, not published by the rule, so it also covers
renames done by any other rule of the producer body:
* a column which keeps its exprId needs no mapping;
* a renamed column is the one whose exprId disappeared from the outputs, it
is matched by name (a rule which renames an output keeps its name) and the
producer slot of the maps is replaced by the new one;
* a pruned column (e.g. `ColumnPruning`, see
`CteLimitPushdownPlanTest#testPushLimitBeforeProducerOutputPruning`) is not
needed by any consumer and gets no mapping.
The refreshed consumers are written back to `RewrittenCteConsumer` and the
`CteIdToConsumers` collection is refreshed as well, so a later reuse of the
same CTE does not resurrect the stale maps. This runs inside the CTE subtree
rewrite, i.e. before `AdjustNullable` and before the CBO builds the CTE
consumer groups from these maps.
With that, the guard is gone: `ConstantPropagation` names the constant of a
producer output like anywhere else and the consumers follow it.
`ConstantPropagationFreshExprIdTest#testMaterializedCteConsumerMapsStayValid`
asserts both directions now - the producer really outputs different exprIds
than in the analyzed plan, and every producer slot a consumer maps to is an
output of the producer - plus a `checkPlannerResult` on the query.
Verification: 33 FE unit test classes (constant propagation, constant cte,
cte inline, cte limit pushdown, rewrite cte children limit pushdown, search cte
rewrite, data trait, sql plan suite, column pruning, ...) rerun: 403 tests, 0
failures; checkstyle passes.
Note for the earlier discussion: this replaces the "keep the producer output
exprId" approach of `8af10b7bee3` and `fe4ae780bbf`; the CTE part of those
commits is superseded, the sink part (P2) stays.
--
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]
