Timeseries will try writing measurements by calling 
performOrderedTimeseriesWritesAtomically. This can fail with ShardCannotRefreshDueToLocksHeld if a resharding operation is ongoing at the same time. We should handle these errors similar to how scoped_operation_completion_sharding_actions does - by doing a refresh first and then retrying. But timeseries will do a retry loop immediately instead. Resharding will eventually do a refresh, but this causes an extra bucket to get created (because the first failure will close that bucket and allocate a new one for retries).
 
 
More context in this comment.