[FLINK-10246] Harden and separate MetricQueryService

• Type: Improvement
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.5.3, 1.6.0, 1.7.0
• Component/s: Runtime / Coordination, Runtime / Metrics
• Assignee: Unassigned
• Reporter: Till Rohrmann
• Created: 29/Aug/18 07:46
• Updated: 11/Dec/18 15:24
• Resolved: 11/Dec/18 15:24

Description

This is an umbrella issue to track the effort to harden Flink's MetricQueryService and to separate it from the rest of the system.

The idea is to setup the MetricQueryService and the metric system in general in such a way that it cannot interfere with or even bring the main Flink components down. Moreover, the metric system also should not degrade performance by simply using any free CPU cycles but not more. Ideally, the user does not see a difference between running Flink with metric query service turned on or off.

Issue Links

• is related to: FLINK-10226 Latency metrics can choke job-manager

Sub-Tasks

• Run MetricQueryService in separate thread pool: Resolved
• Handle oversized metric messages: Closed
• Run MetricQueryService with lower priority: Closed
• Run MetricFetcher in metrics ActorSystem: Resolved