start_workflow_execution fails with duplicate key error despite TERMINATE_EXISTING policy (possible DB inconsistency)

Expected Behavior

Starting a workflow with the following parameters should either succeed or terminate the existing one as specified by the conflict policy:

• id_reuse_policy: ALLOW_DUPLICATE
• execution_timeout: 12h
• id_conflict_policy: TERMINATE_EXISTING

Actual Behavior

The gRPC call start_workflow_execution fails after multiple retries with the following error:

WARN temporal_client::retry: gRPC call start_workflow_execution retried 7 times error=Status { code: Unavailable, message: "createOrUpdateCurrentExecution failed. Failed to insert into current_executions table. Error: pq: duplicate key value violates unique constraint \"current_executions_pkey\"", metadata: MetadataMap { headers: {"content-type": "application/grpc"} }, source: None }

This suggests that a previous workflow with the same ID was not properly cleaned up, possibly due to a failed rollback transaction in Temporal, resulting in data inconsistency in the current_executions table.

Manual deletion of the row in the database resolves the issue:

DELETE FROM current_executions WHERE workflow_id = 'problem-id';

Steps to Reproduce the Problem

I haven't found a way to reproduce the problem, but the process looks like this:

1. Attempt to start a workflow with the given parameters (ALLOW_DUPLICATE, TERMINATE_EXISTING, long timeout).
2. Observe retries and eventual failure with the duplicate key value violates unique constraint error.
3. Verify that the conflicting workflow ID still exists in current_executions table.
4. Manually delete the row to allow the new workflow to start successfully.

Specifications

• Version: temporalio (python sdk) 1.10.0
• Platform: Temporal Version 1.27.1 (PostgreSQL).

Start at the start_workflow_execution gRPC path and the createOrUpdateCurrentExecution write to PostgreSQL's current_executions table. Trace how TERMINATE_EXISTING and rollback behave across retries, and determine whether the duplicate-key failure can be reproduced or explained as inconsistent state.