In the present scenario, there are issues in left_token_ring transition state
execution in the decommissioning path. In case of concurrent mutation race
conditions, we enter left_token_ring more than once, and apparently if
we enter left token ring second time, we try to barrier the decommisioned
node, which at this point is no longer possible. That's what causes the errors.
This pr resolves the issue by adding a check right in the start of
left_token_ring to check if the first topology state update, which marks
the request as done is completed. In this case, its confirmed that this
is the second time flow is entering left_token_ring and the steps preceding
the request status update should be skipped. In such cases, all the rest
steps are skipped and topology node status update( which threw error in
previous trial) is executed directly. Node removal status from group0 is
also checked and remove operation is retried if failed last time.
Although these changes are done with regard to the decommission operation
behavior in `left_token_ring` transition state, but since the pr doesn't
interfere with the core logic, it should not derail any rollback specific
logic. The changes just prevent some non-idempotent operations from
re-occuring in case of failures. Rest of the core logic remain intact.
Test is also added to confirm the proper working of the same.
Fixes: scylladb/scylladb#20865
Backport is not needed, since this is not a super critical bug fix.
Closesscylladb/scylladb#26717
ab0e0eab90