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 47241e4
Browse filesBrowse the repository at this point in the historyBrowse files
[fix](fe) Retry split fetches while file listing is still running
### What problem does this PR solve?
Issue Number: N/A
Related PR: #69683
Problem Summary: With the new batch-mode defaults, a Hive scan can dispatch after its first partition produces a split. If another partition is still listing files, a BE may exhaust its assigned ranges and reach the one-second split-fetch wait limit. FE previously returned a fatal timeout even though split production was still running. FE now returns a retryable INCOMPLETE status only for this temporary gap; BE retries after checking query cancellation and releases its FE RPC client between attempts. Completed scans still return EOF, while producer errors and cancellation retain their failure paths. The wait setting remains a per-RPC bound and now rejects nonpositive values to avoid a tight retry loop.
### Release note
External batch scans no longer fail solely because another partition takes longer than one split-fetch wait interval to list files.
### Check List (For Author)
- Test: Unit Test: SplitSourceTest 37/37 passed with the repository test script. FE Maven build and Checkstyle passed through build.sh --fe; the outer script returned nonzero during output packaging. BE clang-format 16 check and git diff --check passed. BE compilation was not run locally.
- Behavior changed: Yes. A temporary split gap produces a retryable response instead of a fatal timeout.
- Does this need documentation: No. The session variable description was updated in code.
0 commit comments