RadosGW slow when reaching a certain amount of objects

[Dmitry Sutyagin]

RadosGW has reached 16M+ objects in .rgw.buckets in ceph and has become very slow.
A similar issue has been discussed here - http://lists.ceph.com/pipermail/ceph-users-ceph.com/2016-September/012987.html

It may be necessary to implement multiple services per host or upgrade ceph (http://ceph.com/community/new-luminous-bluestore/) to avoid this issue.

Environment is MOS 9.2.

S3 API is used, but the same issue is with Swift API (tried).

[Denis Meltsaykin (dmeltsaykin)]

Dmitry, I understand your point, but unfortunately we cannot implement anything (i.e. develop a new feature) in the stable released product. And of course we cannot do the major upgrade of Ceph in MOS 9.2. I propose to create a note in the release notes about this limitation. If you have any good solution to fix this bug without a major upgrade or implementing a new feature, please suggest.

[Dmitry Sutyagin (dsutyagin)]

Denis, I understand. I may be able to test multiworker approach for rgw and see if it improves anything -
 depends on customer's decision.