In the sharded cluster python test fixture, we call refreshLogicalSessionCacheNow on the config server to ensure the config.system.sessions collection exists and is sharded. After this, we call refreshLogicalSessionCacheNow on each shard. The shard version of this command will fail if config.system.sessions is found to be unsharded. Because this command is run directly against the shard primary, we have no guarantee of causal consistency between the shardCollection on the config server and the refresh on the shard server, though, which can cause failures in the test setup if the refresh reads from a config server secondary which is behind the primary and so does not know the sessions collection has been sharded yet.

We should change the test setup either to ensure causal consistency here or to add retries upon failure of the refreshLogicalSessionCacheNow function when run on the shards.