[sakthivelmanii]

Problem

• When a read-write transaction executes statements that acquire locks and the user callback returns normally, TransactionRunnerImpl, TransactionManagerImpl, and AsyncTransactionManagerImpl do not roll back if Commit fails with an indeterminate error (DEADLINE_EXCEEDED, CANCELLED, UNAVAILABLE, or RESOURCE_EXHAUSTED with RetryInfo).
• If the Commit RPC does not complete on Cloud Spanner, the server-side transaction remains ACTIVE (or PRECOMMITTED) holding row locks for ~15 seconds until server-side idle cleanup aborts it.

Solution

• When a read-write transaction callback returns normally and Commit fails with an indeterminate error (DEADLINE_EXCEEDED, CANCELLED, UNAVAILABLE, or RESOURCE_EXHAUSTED with RetryInfo), TransactionRunnerImpl, TransactionManagerImpl, and AsyncTransactionManagerImpl now trigger a best-effort fire-and-forget rollbackAsync().
• If the Commit RPC did not complete on Cloud Spanner, this background rollback promptly releases server-side row locks rather than leaving the transaction in ACTIVE or PRECOMMITTED state until server-side idle timeout (~15s).
• Deterministic commit failures (such as ALREADY_EXISTS or INVALID_ARGUMENT) and RESOURCE_EXHAUSTED without RetryInfo do not trigger a rollback, as the transaction has already failed definitively on Cloud Spanner.
• The original commit exception is always rethrown with any synchronous rollback failure attached via addSuppressed(...).

[gemini-code-assist]

Code Review

This pull request introduces automatic rollback behavior on commit failures for specific error codes (such as DEADLINE_EXCEEDED, CANCELLED, UNAVAILABLE, and RESOURCE_EXHAUSTED with a retry delay) across AsyncTransactionManagerImpl, TransactionManagerImpl, and TransactionRunnerImpl. It also adds comprehensive unit tests to verify these rollback scenarios. The review feedback suggests two improvements: adding null checks for txn and commitError in rollbackBestEffort to prevent potential NullPointerExceptions, and increasing the assertion timeout in OrphanedTransactionTest from 200ms to 2000ms to prevent flaky test failures in busy CI/CD environments.

[gemini-code-assist]

To prevent potential NullPointerExceptions, it is safer to add a null check for txn before calling txn.rollbackAsync(). Additionally, checking if commitError is non-null before calling addSuppressed ensures robust exception handling.

  static void rollbackBestEffort(TransactionContextImpl txn, Throwable commitError) {
    if (txn == null) {
      return;
    }
    try {
      ApiFuture<Empty> unused = txn.rollbackAsync();
    } catch (RuntimeException rollbackError) {
      if (commitError != null) {
        commitError.addSuppressed(rollbackError);
      }
    }
  }

[gemini-code-assist]

Using a short timeout of 200ms for checking if a rollback request was received can lead to flaky test failures in busy CI/CD environments where thread scheduling and gRPC mock server processing might be delayed. Increasing this timeout to 2000ms provides a safer margin while still keeping the test fast due to the short sleep interval.

Suggested change

	long deadline = System.currentTimeMillis() + 200L;
	long deadline = System.currentTimeMillis() + 2000L;