As part of SERVER-120403 I added some skipped test cases for retryable writes to sharded time series collections. This uncovered some inconsistent behavior:

• Metrics from the shards are not propagated to the router (e.g, nModified reported as zero)
• Retries (with same lsid and txnid) are not deduped. They will look like multiple executions, even if performed just once.

We should decide the right behavior for this case. If we don't plan on supporting time series updates for query stats on shard servers, we should probably explicitly opt out registering query stats for such updates on the router as well.