The issue was reported in the context of a cluster having many sessions opened with username and password (instead of API tokens) by a monitoring system. A segfault triggered with certain usage patterns was fixed.The ha-simulator can now better help to test races in scheduling (on the different nodes) by introducing a skip-round.Handle an edge case where a node would get stuck in fence state, if all services were removed from it before the node actually fenced itself.This increased the number of configurable services to be well above the previous implementation. The issue was fixed by additionally sorting the services by the amount of time in which they hadn't been scheduled. Since the services that have already started must also be checked to ensure that they are still in the target state, it could happen during large deployments that the services starved at the end of the queue. Ha-manager uses a statically configurable number of workers to handle the services, by scheduling jobs in repeating scheduling rounds with a time-slice of 10s max. Improve handling of huge numbers of services per node, by improving the LRM scheduler that starts workers.Better handling of snapshot removal (for example, after finishing a backup) when storage replication is configured.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |