eldenmoon commented on PR #69761:
URL: https://github.com/apache/doris/pull/69761#issuecomment-6602324617
Update on F-02 (replaces the F-02 part of my previous reply): the
native-merge veto from 3a63ad32a85 is reverted in e263ab5668f, and F-02 stays a
documented limitation.
The veto had to rebuild every root with sparse or DOC input, because those
cells carry no per-path type metadata, so most wide documents would lose native
merging. The drift itself is narrow. It needs all three of:
1. the same path stored with different types in different input rowsets: a
boolean and a number, an integer beyond 2^53 and a double, or FLOAT and DOUBLE;
2. a native merge of those rowsets, that is, the generic SNII eligibility
check passes;
3. a MATCH term equal to such a leaf's text before or after the promotion,
for example `true` or `1` in your case.
Other promotions keep each leaf's text: integer width changes and the INT →
DOUBLE → STRING promotion that `test_variant_values_index_compaction` checks
for index-on/index-off parity in all four native/vertical combinations.
The underlying issue is that compaction changes the stored value:
`{"n":true}` reads back as `{"n":1}` with or without an index. That is VARIANT
type promotion, and we will fix it separately; once the stored value no longer
changes, the native postings match it again. The PR description now lists these
trigger conditions under the known limitations.
The F-01 fix in 3a63ad32a85 stays: BUILD INDEX is rejected on a table with a
values index, and the generic builder returns an error Status instead of the
fatal check.
--
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]
