Status
Current state: [Under Discussion]
Discussion threads:
- FLIP-75 discussion about the initial design
- FLIP-102 discussion after splitting up FLIP-75 into sub-flips
JIRA:
Released: <Flink Version>
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
FLIP-49 has been accepted and merged in Flink 1.10, the metric in current task manager detail page could not correspond well to the design of FLIP-49.
The memory model which is exposed through the configuration parameters should be visualized in the same way in the TaskManager's details.
Proposed Changes
According to FLIP-49, we can correlate the configuration parameters and the metrics partially.
JVM Metrics
These JVM metrics are exposed and can be used through the TaskManager's metrics REST API.
JVM | Metric | Used key | Total key |
---|---|---|---|
Heap | Status.JVM.Memory.Heap | Used | Max |
Direct | Status.JVM.Memory.Direct | Used | Max |
Metaspace | Status.JVM.Memory.Metaspace
| Used | Max |
Mapped | Status.JVM.Memory.Mapped | MemoryUsed | TotalCapacity |
NonHeap | Status.JVM.Memory.NonHeap | MemoryUsed | TotalCapacity |
Memory Configuration
Flink's memory model (as described in org.apache.flink.runtime.clusterframework.TaskExecutorProcessSpec
) can be mapped to the following Flink configuration parameters. There are a few that have a correlating Flink metric.
Flink Memory Model | Flink configuration1 | Effective Configuration REST API2 | Metric3 | Used key | Total key |
---|---|---|---|---|---|
Framework Heap | taskmanager.memory.framework.heap.size | memoryConfiguration.frameworkHeap | Status.JVM.Memory.Heap | Used | Max |
Task Heap | taskmanager.memory.task.heap.size | memoryConfiguration.taskHeap | |||
Framework OffHeap | taskmanager.memory.framework.off-heap.size | memoryConfiguration.frameworkOffHeap | - | - | - |
Task OffHeap | taskmanager.memory.task.off-heap.size | memoryConfiguration.taskOffHeap | |||
Network Memory | memoryConfiguration.networkMemory | Status.Shuffle.Netty | UsedMemory | TotalMemory | |
Managed Memory | taskmanager.memory.managed.size | memoryConfiguration.managedMemory | Status.ManagedMemory | (is up for discussion, still) | (is up for discussion, still) |
JVM Metaspace | taskmanager.memory.jvm-metaspace.size | memoryConfiguration.jvmMetaspace | Status.JVM.Memory.Metaspace | Used | Max |
JVM Overhead | memoryConfiguration.jvmOverhead | - | - | - |
1 These are the configuration parameters used in the Flink configuration.
2 These are the Json paths to address the properties in the HTTP REST API response. Additionally, memoryConfiguration.totalFlinkMemory
and totalProcessMemory
are exposed through the REST API.
3 The metrics which are exposed through the TaskManager's metrics REST API.
Frontend Design (out-dated)
Redesign the task manager metric page, this would allow users to more clearly understand the relationship between these metrics.
REST API Design
- task manager's resource contains this information, show it in
url: /taskmanagers/:taskmanagerid
Implementation Proposal
Step 1: Expose effective configuration parameters of TaskExecutorn
TaskManagerResourceInfo
is introduced as a POJO containing the relevant values proposed in the REST response.- The
TaskManagerResourceInfo
is initialized when initializing theTaskExecutor
in the same way as we do it with theHardwareDescription
. It will be handed over in the same way throughTaskExecutorRegistry
→WorkerRegistration
. - The
TaskManagerResourceInfo
will be added along withHardwareDescription
inResourceManager::requestTaskManagerInfo(ResourceId, Time)
.
Step 2: Introduce new metric for memory usage of NetworkBufferPool
add shuffle memory's size metric
public long getTotalMemorySize() { return 1L * getTotalNumberOfMemorySegments() * memorySegmentSize; } public long getUsedMemorySize() { return getTotalMemorySize() - (getNumberOfAvailableMemorySegments() * memorySegmentSize); }
update
NettyShuffleMetricFactory#registerShuffleMetrics
private static final String METRIC_TOTAL_MEMORY_IN_BYTES = "TotalMemory"; private static final String METRIC_USED_MEMORY_IN_BYTES = "UsedMemory"; // ... private static void registerShuffleMetrics( String groupName, MetricGroup metricGroup, NetworkBufferPool networkBufferPool) { MetricGroup networkGroup = metricGroup.addGroup(groupName); networkGroup.<Integer, Gauge<Integer>>gauge(METRIC_TOTAL_MEMORY_SEGMENT, networkBufferPool::getTotalNumberOfMemorySegments); networkGroup.<Integer, Gauge<Integer>>gauge(METRIC_AVAILABLE_MEMORY_SEGMENT, networkBufferPool::getNumberOfAvailableMemorySegments); // === new === networkGroup.<Long, Gauge<Long>>gauge(METRIC_TOTAL_MEMORY_IN_BYTES, networkBufferPool::getTotalMemorySize); networkGroup.<Long, Gauge<Long>>gauge(METRIC_USED_MEMORY_IN_BYTES, networkBufferPool::getUsedMemorySize); }
Step 3: Introduce new metrics for Task's managed memory usage
We still have to discuss how to implement that in the right way. A brief proposal is the following one:
We would have to introduce a new metric that represents the aggregated memory usage of each TaskSlot
. The aggregation can be maintained in the TaskExecutor
.
- add managed memory metric and expose it in rest API, as now managed memory is allocated in the slot, we need to sum it in TaskExecutor.
add default memory type in MemoryManager
public static final MemoryType DEFAULT_MEMORY_TYPE = MemoryType.OFF_HEAP;
add register ManagedMemoryUsage in TaskExecutor#requestSlot:
public long getManagedMemoryUsed() { return this.taskSlotTable.getAllocatedSlots().stream().mapToLong( slot -> slot.getMemoryManager().getMemorySizeByType(MemoryManager.DEFAULT_MEMORY_TYPE) - slot.getMemoryManager().availableMemory(MemoryManager.DEFAULT_MEMORY_TYPE) ).sum(); }
add instantiateMemoryManagerMetrics in MetricUtils
public static void instantiateMemoryManagerMetrics( MetricGroup statusMetricGroup, TaskExecutor taskExecutor) { checkNotNull(statusMetricGroup); MetricGroup memoryManagerGroup = statusMetricGroup.addGroup("Managed").addGroup("Memory"); memoryManagerGroup.<Long, Gauge<Long>>gauge("TotalCapacity", taskExecutor::getManagedMemoryTotal); memoryManagerGroup.<Long, Gauge<Long>>gauge("MemoryUsed", taskExecutor::getManagedMemoryUsed); }
- register it in TaskManagerRunner#startTaskManager
- taskmanager’metrics add Status.Managed.Memory.TotalCapacity and Status.Managed.Memory.MemoryUsed
- url: /taskmanagers/:taskmanagerid
- response: metric add managedMemoryTotal and managedMemoryUsed
Step 4: Add Metaspace metrics
There are no metrics present, yet, monitoring the JVM's Metaspace pool. The newly introduced metrics are going to be exposed through the {{/taskmanagers/metrics}} REST API.
Step 5: Update TaskManager details pages
The web UI has to be updated as proposed above.
Follow-Ups
- Create a separate independent endpoint for the effective memory configuration.
Test Plan
Existing tests are updated to verify feature.