[IMPALA-5119] Don't make RPCs from Coordinator::UpdateBackendExecStatus()

• Type: Sub-task
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: Impala 2.9.0
• Component/s: Distributed Exec
• Target Version: Product Backlog
• Epic Color: ghx-label-2
• Assignee: Unassigned
• Reporter: Henry Robinson
• Created: 27/Mar/17 21:10
• Updated: 07/Jun/19 00:00

Description

If it reports a bad status, UpdateFragmentExecStatus() will call UpdateStatus(), which takes Coordinator::lock_ and then calls Cancel(). That method issues one RPC per fragment instance.

In KRPC, doing so much work from UpdateFragmentExecStatus() - which is an RPC handler - is a bad idea, even if the RPCs are issued asynchronously. There's still some serialization cost.

It's also a bad idea to do all this work while holding lock_. We should address both of these to ensure scalability of the cancellation path.

Issue Links

• depends upon: IMPALA-2990 Coordinator should timeout and cancel queries with unresponsive / stuck executors
• depends upon: IMPALA-5384 Simplify coordinator locking protocol