Consider the 2 types of kv_vb_map in use to calculate the stats for a paused replication in UpdateStats(...):

A. cur_kv_vb_map, calculated as: 

B. sourceVBMap, calculated as: 

Say there are N KV nodes in the source cluster and for the sake for simplicity, let's say that all the N nodes have T total_docs each and each have processed (docs_processed) P docs.

The difference between the maps is that:

(A) contains all the N nodes and stats calculated using this will be stats aggregated across the cluster level

AND

(B) contains only 1 node (the current node) in its map i.e. the stats calculated using this will be the stats for itself only.

And when we hit the following codepath, we use (A)

constructStatsForReplication calculates the following:

1. total_docs: highSeqNo (gotten from KV) for all the nodes in (B) = 1*T. Example, For a 3 KV node setup, because of this bug we get: 

      2. docs_processed: seqno from checkpoints of all the VBs of nodes in (A) = N*P

      3. changes_left = total_docs - docs_processed = 1*T - N*P which potentially will go negative.

This has to be fixed for all of this path to use (B), so that we are also consistent and calculate each overview stat for that node only.

Additionally, in this same code path we always read the overview_stats from the main pipeline, but endup storing it in backfill pipline's stats store sometimes:

This may need revisiting as well.