morrySnow opened a new pull request, #67737:
URL: https://github.com/apache/doris/pull/67737
## Problem
Storage-layer aggregation could be selected for `COUNT`, `MIN`, and `MAX`
whose argument was a numeric `CAST` or `TRY_CAST`. The storage scan
aggregates
the source column directly and cannot preserve the cast's value and null
semantics, so the optimization can change query results.
For example, with a non-null `DOUBLE` column containing `42`, `2147483647`,
`2147483648`, `-2147483649`, and `1e20`, and with strict casts disabled:
```sql
SET enable_strict_cast = false;
SET enable_push_down_no_group_agg = true;
SELECT COUNT(CAST(d AS INT)), COUNT(TRY_CAST(d AS INT)) FROM count_t;
```
Only two casts are non-null, so both counts must be `2`. Before this change,
the scan used `pushAggOp=COUNT` and returned `5`. A similar narrowing cast
from
`BIGINT` values `-200, 0, 200` to `TINYINT` made cast-based `MIN`/`MAX` use
`pushAggOp=MINMAX` and return `NULL` instead of `0`.
## Root cause
The eligibility check treated a numeric cast of a slot as equivalent to the
slot itself. However, the storage aggregation descriptor carries the source
slot and operation, not an expression evaluator for the cast. Numeric casts
are not necessarily lossless: they can introduce nulls on overflow and can
change values relevant to extrema.
## Fix
Restrict storage-layer aggregation arguments to raw slot references. Keep any
explicit `CAST` or `TRY_CAST` in the regular row-evaluation and aggregation
path, including casts exposed through a project alias. This conservative rule
preserves correctness for all scan types and all supported aggregate
operations while retaining pushdown for unmodified columns.
## Tests
- Added unit coverage for direct `COUNT`, `MIN`, and `MAX` arguments using
`CAST`/`TRY_CAST`, plus a projected cast alias.
- Ran `PhysicalStorageLayerAggregateTest`: 8 tests passed, 0 failed.
- Rebuilt and deployed the FE in a local sandbox. Both counts return `2`, all
four cast-based extrema return `0`, and the scan plans show
`pushAggOp=NONE`.
Issue Number: None
--
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]
