When a multi-document transaction needs the filtering metadata for a namespace and a refresh is in progress, it will serialize with that refresh using a deadline of 500 ms. If that deadline is exceeded, the shard will throw an ExceededTimeLimit error.

We would like to investigate whether this ExceededTimeLimit is retried at the level of mongos (because it is a retriable error) or at the level of the driver (because we attach the TransientTransactionError label). Ideally, mongos should be the component retrying this error, and we would prefer not to make the driver responsible for retrying the entire transaction from the beginning.