Details
Major Description
In one of our larger production clusters we're seeing issues with snapshot performance that cause the scheduler to failover before completing a snapshot.
For background, the scheduler writes a compressed (when -deflate_snapshots is enabled), binary-encoded Snapshot (from api.thrift) to the mesos replicated log every hour (or -dlog_snapshot_interval). This snapshot represents most of the scheduler's heap usage, including the configuration for all tasks running in the cluster.
Add appropriate instrumentation to the snapshot routine and patch any obvious performance bottlenecks.
Attachments
Issue Links
- relates to
-
AURORA-74 Write snapshots and backups in a way that requires less memory
-
- Open
-