[KUDU-2069] Add a maintenance mode

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 1.11.0
• Assignee: Andrew Wong
• Reporter: Jean-Daniel Cryans
• Created: 13/Jul/17 22:19
• Updated: 09/Oct/19 07:32
• Resolved: 09/Oct/19 07:32

Description

Adding a maintenance mode in Kudu would mainly prevent replicas from getting evicted while the cluster remains in that mode. We might also consider preventing table create/alter/delete.

Maintenance mode is a good thing for two use cases:

• When starting a cluster, some nodes might take longer than others to come back. We shouldn't kick their replicas out.
• When nodes need to be taken out for a known short period of time, say to change a disk or perform an OS upgrade, we shouldn't necessarily evict all its replicas. Right now, taking a node out for more than 5 minutes is equivalent to wiping it out since it will delete all its replicas when coming back. In fact, wiping it out might even be preferable since it cleans up the data dirs.

Issue Links

• relates to: KUDU-1827 Add capability and tool to gracefully decommission a tablet server
• relates to: KUDU-2054 Rolling Restart and Upgrade