as you mention
quorum, I guess you refer to the
quorum is the default, so if less than a quorum of the replicas succeeds, the write is not successful (see a few more details in the Definitive Guide). However, note that this is actually just a pre-check before the actual replication takes place.
We will keep writes on replicas that succeed and get a new replica node assignment from the master if one of the replicas fails. After a new replica node is assigned, the shard in question is synced to the new replica.
However, consistency in distributed systems is a hard problem and we document known edge cases and Elasticsearch's behavior / the status of our fixes on the resiliency page.