[AURORA-930] GC performance issues with snapshot deduplication on large clusters

• Type: Story
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.7.0
• Component/s: Scheduler
• Sprint: Twitter Aurora Q4 Sprint 4, Twitter Aurora Q4 Sprint 5
• Assignee: Kevin Sweeney
• Reporter: Chris Lambert
• Created: 17/Nov/14 21:08
• Updated: 17/Apr/15 22:27
• Resolved: 01/Dec/14 20:17

Description

We observed very long GC pauses in our production cluster we believe are caused by the extra heap pressure caused by SnapshotDeduplicatorImpl's extra deep copy of a Snapshot object.

Also, EntrySerializer fully buffers the full binary-encoded thrift Snapshot before writing it to the replicated log - it can avoid memory pressure by streaming instead.

Issue Links

• relates to: AURORA-953 Change the log format so that FrameHeader doesn't need to know the size and checksum of the serialized data ahead-of-time.
• relates to: AURORA-74 Write snapshots and backups in a way that requires less memory