Motivation
After generating splits for a non-empty table, TieringSourceEnumerator.generateAndAssignSplits() assigns them and returns. If all pending splits were assigned but some readers are still waiting, it does not immediately claim another table.
Example: eight readers have already requested work and are waiting. A periodic poll claims table A with two splits, while table B is also available. After assigning A, six readers remain idle until another event or periodic poll triggers a claim. The default polling interval is 30 seconds; an earlier split completion can shorten this wait.
This is a scheduling gap identified by code inspection, separate from continuation after an empty table.
Solution
After assignment, asynchronously continue claiming tables while readers remain available and the pending split queue is empty. Stop when no table is available or the concurrency limit is reached. Coalesce claim requests and retain backoff for empty responses and failures.
Add a test for the example above: B should be assigned without completing A's splits or advancing the periodic timer. Do not buffer extra splits on busy readers.
Anything else?
Related: #4372 addresses empty tables. This case concerns non-empty tables whose splits do not fill the waiting readers.
Willingness to contribute
Motivation
After generating splits for a non-empty table,
TieringSourceEnumerator.generateAndAssignSplits()assigns them and returns. If all pending splits were assigned but some readers are still waiting, it does not immediately claim another table.Example: eight readers have already requested work and are waiting. A periodic poll claims table A with two splits, while table B is also available. After assigning A, six readers remain idle until another event or periodic poll triggers a claim. The default polling interval is 30 seconds; an earlier split completion can shorten this wait.
This is a scheduling gap identified by code inspection, separate from continuation after an empty table.
Solution
After assignment, asynchronously continue claiming tables while readers remain available and the pending split queue is empty. Stop when no table is available or the concurrency limit is reached. Coalesce claim requests and retain backoff for empty responses and failures.
Add a test for the example above: B should be assigned without completing A's splits or advancing the periodic timer. Do not buffer extra splits on busy readers.
Anything else?
Related: #4372 addresses empty tables. This case concerns non-empty tables whose splits do not fill the waiting readers.
Willingness to contribute