[FLINK-22725] SlotManagers should unregister metrics at the start of suspend()

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.13.0
• Fix Version/s: 1.14.0, 1.13.2
• Component/s: Runtime / Coordination, Runtime / Metrics
• Labels: pull-request-available
• Assignee: Chesnay Schepler
• Reporter: Chesnay Schepler
• Created: 20/May/21 11:53
• Updated: 28/Aug/21 12:11
• Resolved: 25/May/21 20:34

Description

Slotmanagers register metrics in start(), but only unregister them in close().

This has 2 issues:

a) If the SM is restarted it cannot re-register the metrics because the old ones are still present; this isn't a critical problem (because the old ones still work) but it produces unnecesasry logging noise

b) The metric may produce an NPE when accessed by a reporter during suspend().

Issue Links

• is related to: FLINK-22646 Accessing DeclarativeSlotManager metrics during unregister throws NPE
• links to: GitHub Pull Request #15975
• links to: GitHub Pull Request #15977