[Segment Replication] Reduce/remove document parsing on replicas

[andrross]

When using segment replication, TransportShardBulkAction.performOpOnReplica() is invoked on replicas in order to persist the document into the translog on the replica and give the same durability guarantees as document replication. Here are two snapshots of CPU profiling for the same workload for both docrep and segrep:

Document Replication

Segment Replication

It clearly shows that NRTReplicationEngine does not invoke IndexWriter, whereas InternalEngine does. This is expected since the replicas do not need to index the document. However, a lot of work is spent doing DocumentMapper#parse in both cases. The question is: can this be refactored so that with segment replication the replicas avoid doing some or all of that document parsing?

[anasalkouz]

@mch2 Can you share new profiler after fixing the issue?