>> 3. This also means the overhead of lifecycle is massively increased. Lifecycle scans every object in every bucket with policy every <interval>. This is not usually a problem, because it happens once per day, and has 24 hours to complete (but in large systems, it can even take more than 24 hours to complete). Make sure your cluster is small enough that all lifecycle can be processed in <interval>, or it will generally back up and not process things in time. >> >> >> So, basically, it can sort of be done, but it's a very bad idea. One strategy that can partly mitigate the overhead is to have a cohort of RGWs that are dedicated to LC and/or GC processing that are not behind whatever ingress / LB endpoint is used to federate client operations. With such a strategy the client RGWs have LC/GC processing disabled via a central config mask, and the dedicated cohort have LC/GC enabled via a similar config/mask. This way clients and LC/GC don’t contend directly for each other within the same RGW instances, though of course the CPU and underlying storage demands remain, especially if one has bought into the false economy of HDDs for bucket data. _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
 |