[Fix][Connector-V2] Fix Arrow memory leak by correcting close order in ArrowToSeatunnelRowReader

[davidzollo]

Purpose

Fix IllegalStateException: Memory was leaked by query when using the Doris connector (and any connector backed by ArrowToSeatunnelRowReader).

Closes #9863

Problem

In ArrowToSeatunnelRowReader.close(), the previous close order was:

1. root.close() — close VectorSchemaRoot
2. rootAllocator.close() — close RootAllocator ← wrong: arrowStreamReader still holds unreleased buffers
3. arrowStreamReader.close() — too late, allocator already closed

Apache Arrow's memory management requires that all child allocations are released before the RootAllocator is closed. The ArrowStreamReader internally allocates Arrow buffers via the RootAllocator. Closing the allocator while the reader still holds references causes the allocator to detect leaked memory (~64 bytes of internal metadata) and throw:

java.lang.IllegalStateException: Memory was leaked by query. Memory leaked: (64)
Allocator(ROOT) 0/64/64/2147483647 (res/actual/peak/limit)

This exception propagated up through DorisValueReader.hasNext() → DorisSourceReader.pollNext() → SourceFlowLifeCycle.collect() causing the task to fail with state FAILED.

Fix

Corrected the close order to:

1. arrowStreamReader.close() — releases all Arrow buffers and internally closes VectorSchemaRoot
2. rootAllocator.close() — now safe, all allocations already released

The separate root.close() call is removed since ArrowStreamReader.close() already handles it.

Impact

• Affected connectors: Doris, StarRocks, and any connector using ArrowToSeatunnelRowReader
• Trigger: Any read operation that creates an ArrowToSeatunnelRowReader instance
• Backward compatible: Yes, pure bug fix with no API or behavior changes

[DanielLeens]

Thanks for the contribution. I reviewed the full diff on the current head and retraced the close path instead of only reading the PR description.

What this PR fixes

• User pain: the Arrow reader can throw a leaked-memory exception during shutdown even after the read path itself succeeds.
• Fix approach: close ArrowStreamReader before closing the root allocator.
• One-line summary: this fixes the real Arrow resource-release order on the normal close path.

Runtime chain I checked

Arrow source read path
  -> ArrowStreamReader owns VectorSchemaRoot buffers
  -> ArrowToSeatunnelRowReader.close()
      -> ArrowStreamReader.close()
      -> VectorSchemaRoot buffers released back to allocator
      -> RootAllocator.close()

Findings

• The normal close path definitely hits ArrowToSeatunnelRowReader.close().
• The new order at ArrowToSeatunnelRowReader.java:320-334 is the correct one for Arrow ownership semantics.
• I do not see a reopened source-level blocker on this revision.

Merge conclusion

Conclusion: can merge after fixes

1. Blocking items

• No code blocker from my side.
• Please let the latest GitHub Build finish green before merging.

2. Suggested non-blocking follow-up

• None from my side on the source path.

Overall, this is a focused and correct fix for the allocator-leak shutdown path.

[nzw921rx]

+1

[DanielLeens]

Thanks for taking a look, @nzw921rx. On the unchanged head 25744b5, Daniel's source-level conclusion is still the same as before: I do not see a reopened code blocker on this revision, and from my side the remaining item is still to let the latest GitHub Build finish green before merge.

There is no new code on the PR head, so I am not starting another full review in this round. Happy to re-check if the branch changes again.