underflow_error thrown in Flusher task results in flushing halting

With the N1QL transactions the bucket RAM grows quickly and OPs drops low double digits.

OS: centos7
Last working version 7.0.0-4003 (starts 29MB and stays 29MB)
First Start seeing issue 7.0.0-4007 (starts 29MB and grows in couple of minutes to 240+MB)
Build failures in between
Latest version test 7.0.0-4061 Attached cbcollectinfo

Repro:
cluster :

/opt/couchbase/bin/couchbase-cli cluster-init -c localhost --cluster-username Administrator --cluster-password password --services data,index,query --cluster-ramsize 2648 --cluster-index-ramsize 1024

/opt/couchbase/bin/couchbase-cli bucket-create -c localhost -u Administrator -p password --bucket b1 --bucket-type couchbase --bucket-ramsize 256 --bucket-replica 0

CREATE SCOPE b1.s1;

CREATE COLLECTION b1.s1.c1;

CREATE INDEX ix1 ON b1.s1.c1(a);

Run the Attached s4c.py on query node

python s4c.py

Observe the bucket RAM, ops ( RAM keep increases, ops drops)
The script does the following
insert 64 documents of each size around 50 bytes
16 threads run transactions in loop each transaction updates 4 documents.

Attached cbcollectinfo (I have issue first time collection, so did Flush b1 and reran repro)
Flush released all the memory.