Details
-
Improvement
-
Resolution: Unresolved
-
Major
-
7.6.0
-
0
Description
KV Range Scans currently use by-key KVStore scans to retrieve the requested items. As highlighted by recent DCP backfill performance work (MB-56547), retrieving a given range of documents via the by-key index is not always faster than a "brute force" scan of the entire bucket in seqno order and returning only documents matching the requested key range.
As such, it could be beneficial for KV Range Scan use-cases which request a more than a small fraction of documents (e.g. >4% of bucket) to instead use a seqno scan. This should result in higher throughput, however at the cost of the results no longer being returned in key order. Therefore such a feature would need to be opt-in from the client - if (potential) higher throughput is desired over ordered result set.