n1ql query throughput up 20x after restoring backup

Although a 20x increase in throughput is desired, I'm a bit suspicious of this one: 

Avg Throughput (queries/sec), Workload SJ, 4 node

http://showfast.sc.couchbase.com/#/timeline/Linux/n1ql/soe/all

3023 - http://perf.jenkins.couchbase.com/job/hebe/7837/ - 27967
3022 - http://perf.jenkins.couchbase.com/job/hebe/7836/ - 1539

Graphs: http://cbmonitor.sc.couchbase.com/reports/html/?snapshot=hebe_700-3022_access_c132&snapshot=hebe_700-3023_access_1f4f

Looks like a lot more inserts are happening. write/sec start off high and decay exponentially to a constant rate.

http://172.23.123.43:10000/?product=couchbase-server&fromVersion=7.0.0&fromBuild=3022&toVersion=7.0.0&toBuild=3023&f_backup=on&f_build=off&f_cbgt=off&f_goxdcr=off&f_kv_engine=off&f_ns_server=off&f_query=on&f_testrunner=off

There is ns_server stats commit that could be responsible: 

https://github.com/couchbase/ns_server/commit/6c94d6828d906a06efbc4df196a03388b6efeab4

But I think the tools commit should be considered first since it mentions N1QL explicitly:

https://github.com/couchbase/backup/commit/24ee2a5cb1e291134d751f05c4cc016e368a94e8

This is the command used in the test:

/opt/couchbase/bin/cbbackupmgr restore --archive /backups --repo soe10M --threads 24 --cluster http://localhost:8091 --username Administrator --password password --disable-ft-indexes --disable-gsi-indexes

The test will restore a backup called soe10M, then load additional data into CB, before running 90% join queries and 5% kv inserts and 5% kv deletes simultaneously.

The performance impact can be seen in workload SJ, SL1, SL2, SMIX, and SMIX2 in YCSB soe branch.