Details
-
Bug
-
Resolution: Fixed
-
Major
-
5.0.0
-
4 data nodes, 1 query node, 1 index node
-
Untriaged
-
Centos 64-bit
-
-
No
Description
Query:
SELECT name, email FROM `bucket-1` WHERE capped_small > $1 ORDER BY capped_small LIMIT 10;
|
Index:
CREATE INDEX by_capped_small ON `bucket-1`(capped_small);
|
There are 20M documents in the database.
The rate of not_bounded queries is rather high. We can achieve to up 18K queries/sec in Spock.
But if I change scan_consistency from not_bounded to request_plus, the throughput drops to ~100 in 4.6 and ~700 in Spock. Spock results are almost as bad as 4.5.0 numbers.
From graphs I can see that cbq-engine receives way more items from 2i when scan_consistency is set to request_plus. That results in high CPU and memory usage, and also larger number of KV fetches.
I wonder if we have any limitation in this area.
See also http://showfast.sc.couchbase.com/#/timeline/Linux/n1ql/pagination