[FLINK-15300] Shuffle memory fraction sanity check does not account for its min/max limit

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: 1.10.0
• Component/s: Runtime / Configuration
• Labels: pull-request-available
• Assignee: Andrey Zagrebin
• Reporter: Andrey Zagrebin
• Created: 17/Dec/19 14:28
• Updated: 08/Jan/20 09:03
• Resolved: 08/Jan/20 09:03

Description

If we have a configuration which results in setting shuffle memory size to its min or max, not fraction during TM startup then starting TM parses generated dynamic properties and while doing the sanity check (TaskExecutorResourceUtils#sanityCheckShuffleMemory) it fails because it checks the exact fraction for min/max value.

Example, start TM with the following Flink config:

taskmanager.memory.total-flink.size: 350m
taskmanager.memory.framework.heap.size: 16m
taskmanager.memory.shuffle.fraction: 0.1

The calculation will happen for total Flink memory and will result in the following extra program args:

taskmanager.memory.shuffle.max: 67108864b
taskmanager.memory.framework.off-heap.size: 134217728b
taskmanager.memory.managed.size: 146800642b
taskmanager.cpu.cores: 1.0
taskmanager.memory.task.heap.size: 2097150b
taskmanager.memory.task.off-heap.size: 0b
taskmanager.memory.shuffle.min: 67108864b

where the derived fraction is less than shuffle memory min size (64mb), so it was set to the min value: 64mb.

While TM starts, the calculation happens now for the explicit task heap and managed memory but also with the explicit total Flink memory and TaskExecutorResourceUtils#sanityCheckShuffleMemory throws the following exception:

org.apache.flink.configuration.IllegalConfigurationException:
Derived Shuffle Memory size(64 Mb (67108864 bytes)) does not match configured Shuffle Memory fraction (0.10000000149011612).
at org.apache.flink.runtime.clusterframework.TaskExecutorResourceUtils.sanityCheckShuffleMemory(TaskExecutorResourceUtils.java:552)
at org.apache.flink.runtime.clusterframework.TaskExecutorResourceUtils.deriveResourceSpecWithExplicitTaskAndManagedMemory(TaskExecutorResourceUtils.java:183)
at org.apache.flink.runtime.clusterframework.TaskExecutorResourceUtils.resourceSpecFromConfig(TaskExecutorResourceUtils.java:135)

This can be fixed by checking whether the fraction to assert is within the min/max range.

Issue Links

• relates to: FLINK-13980 FLIP-49 Unified Memory Configuration for TaskExecutors
• links to: GitHub Pull Request #10608