You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 43a27b0
Browse filesBrowse the repository at this point in the historyBrowse files
### What problem does this PR solve?
Issue Number: None
Related PR: #68196
Problem Summary: On a mode-2 Paimon follower, a names refresh can rebind a logged Foo database to replacement FOO while a cold REFRESH replay still needs to retire SDK-only Foo handles and old Foo row counts. Resolve this ambiguous replay with an opening and closing catalog row-count fence around Paimon SDK retirement. A held table can also read after permanent DROP and recreate a row-count catalog generation; bound that registry so repeated overlapping drops cannot grow FE memory indefinitely. Finally, no-op DROP matching used equalsIgnoreCase, which equates I with dotless i even though catalog lookup's Locale.ROOT fold distinguishes them; use the same fold for resident identity matching.
### Release note
Cold Paimon REFRESH replay and no-op database DROP now preserve the correct external metadata identity; late row-count reads after catalog DROP remain memory bounded.
### Check List (For Author)
- Test: Unit Test: four related FE test classes, 153 tests passed; full ./build.sh --fe -j4 and Checkstyle passed
- Behavior changed: Yes, ambiguous Paimon replay retires old SDK and row-count state, no-op DROP uses catalog name folding, and catalog generation state is bounded
- Does this need documentation: No
0 commit comments