On a DR test with 9k+ document, the secondary site receives only 1,5k documents.
As the elasticsearch replication received 9k+ documents, it seems that the problem is on the source and that we miss documents.
Here is the explanation of the problem:
1. We are in a processTimer call, consuming MongoDB oplog.
2. A Kafka rebalance happens
3. A call to MongoDBComputation#init() happens
4. Start timestamp of the query is updated by what has been already committed to Kafka
5. processTimer ends by consuming more logs
6. a new processTimer starts with the update query (TS > lastCommitedTimestamp) *but* the n-th page of that query
7. we miss the n-th first pages
Here is the interesting part of server.log
Fix is to init the page number at Computation init.