[HBASE-20800] Master-orchestrated compactions

• Type: Umbrella
• Status: Resolved
• Priority: Major
• Resolution: Won't Fix
• Component/s: Compaction
• Assignee: Unassigned
• Reporter: Michael Stack
• Created: 27/Jun/18 21:26
• Updated: 11/Jun/22 19:03
• Resolved: 11/Jun/22 19:03

Description

An umbrella issue for having compactions go via the Master so we can have a centralized arbitrator of cluster i/o. If we put Master in the way, we can do stuff like:

• Ask the Master for current cluster compaction state; what is running, what is blocked
• Master can manage cluster-wide compaction policy and/or throttling/or blocking of compaction i/os.
• Master can schedule when and where compactions run so we can guard against the pathological where all RegionServers decide now is the time to major compact bringing on a compaction storm.

Other side-benefits might include being able to farm out the compaction work to another process – e.g. the splice machine model of having spark run the compactions – or just to a separate compactor that we might i/o nice.

• We'll need to figure how to externalize the CompactionRequest so it can be passed over RPC.
• We'll need to have something like a CompactionManager in the Master process that keeps up current cluster state.

MOB needs a compaction fabric it can use. Its compactions are currently Master-based only and so don't scale. It could make use of this mechanism to ask the Master to farm out its compaction requests.

This is an umbrella issue. I thought I'd filed one already on this topic but can't find it. Will shut it down if I trip over it.

Issue Links

• is related to: HBASE-25714 Offload the compaction job to independent Compaction Server

Sub-Tasks

• Sending compaction descriptions from all regionservers to master.: Resolved