mxm commented on code in PR #762: URL: https://github.com/apache/flink-kubernetes-operator/pull/762#discussion_r1464915527
########## flink-autoscaler/src/main/java/org/apache/flink/autoscaler/utils/MemoryTuningUtils.java: ########## @@ -0,0 +1,132 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.flink.autoscaler.utils; + +import org.apache.flink.autoscaler.JobAutoScalerContext; +import org.apache.flink.autoscaler.ScalingSummary; +import org.apache.flink.autoscaler.config.AutoScalerOptions; +import org.apache.flink.autoscaler.metrics.EvaluatedMetrics; +import org.apache.flink.autoscaler.metrics.EvaluatedScalingMetric; +import org.apache.flink.autoscaler.metrics.ScalingMetric; +import org.apache.flink.configuration.MemorySize; +import org.apache.flink.configuration.TaskManagerOptions; +import org.apache.flink.runtime.jobgraph.JobVertexID; + +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import java.util.Map; +import java.util.Optional; + +/** Tunes the TaskManager memory. */ +public class MemoryTuningUtils { + + private static final Logger LOG = LoggerFactory.getLogger(MemoryTuningUtils.class); + + public static Optional<MemorySize> tuneTaskManagerHeapMemory( + JobAutoScalerContext<?> context, + EvaluatedMetrics evaluatedMetrics, + Map<JobVertexID, ScalingSummary> scalingSummaries) { + + var config = context.getConfiguration(); + if (!config.get(AutoScalerOptions.MEMORY_TUNING_ENABLED)) { + return Optional.empty(); + } + + var globalMetrics = evaluatedMetrics.getGlobalMetrics(); + double avgHeapSize = globalMetrics.get(ScalingMetric.HEAP_AVERAGE_SIZE).getAverage(); + + double numTaskSlotsUsed = globalMetrics.get(ScalingMetric.NUM_TASK_SLOTS_USED).getCurrent(); + int taskSlotsPerTm = config.get(TaskManagerOptions.NUM_TASK_SLOTS); + int currentNumTMs = (int) Math.ceil(numTaskSlotsUsed / taskSlotsPerTm); + + double usedTotalHeapSize = currentNumTMs * avgHeapSize; + LOG.info("Total used heap size: {}", new MemorySize((long) usedTotalHeapSize)); + usedTotalHeapSize *= computeDataChangeRate(evaluatedMetrics); + LOG.info("Resized total heap size: {}", new MemorySize((long) usedTotalHeapSize)); + + int numTaskSlotsAfterRescale = + ResourceCheckUtils.estimateNumTaskSlotsAfterRescale( + evaluatedMetrics, scalingSummaries, numTaskSlotsUsed); + int newNumTms = (int) Math.ceil(numTaskSlotsAfterRescale / (double) taskSlotsPerTm); + LOG.info( + "Estimating {} task slots in use after rescale, spread across {} TaskManagers", + numTaskSlotsAfterRescale, + newNumTms); + + MemorySize newHeapSize = new MemorySize((long) (usedTotalHeapSize / newNumTms)); + // TM container memory can never grow beyond the user-specified max + Optional<MemorySize> maxMemory = context.getTaskManagerMemoryFromSpec(); Review Comment: > From the code, I understand that "Adjust memory" will only adjust the memory down to below the user-specified memory, but not above the user-specified memory, right? That is correct. > I'm not sure could we make this option name as generic as possible? Such as: `job.autoscaler.taskmanager-memory.scaling.enabled` and `job.autoscaler.taskmanager-memory.scaling.heap.min`. Sure, we can change those. Maybe `job.autoscaler.memory.taskmanager.heap.min` would be even more generic? It would also allow to add a jobmanager section under memory. > In the short term, we only turn down. In the long term, we might support turn up the memory. If the option name is more generic, it's easy to compatible. WDYT? I think the current option prefix would also allow to support increasing in the future, e.g. by adding `*.heap.max` option. ########## flink-kubernetes-operator/src/main/java/org/apache/flink/kubernetes/operator/autoscaler/KubernetesJobAutoScalerContext.java: ########## @@ -56,19 +57,47 @@ public KubernetesJobAutoScalerContext( jobStatus, configuration, metricGroup, - Optional.ofNullable(configuration.get(KubernetesConfigOptions.TASK_MANAGER_CPU)) - .orElse(0.), - Optional.ofNullable(configuration.get(TaskManagerOptions.TOTAL_PROCESS_MEMORY)) - .orElse(MemorySize.ZERO), restClientSupplier); this.resourceContext = resourceContext; } + @Override + public Optional<Double> getTaskManagerCpu() { + return Optional.ofNullable( + getConfiguration().get(KubernetesConfigOptions.TASK_MANAGER_CPU)); + } + + @Override + public Optional<MemorySize> getTaskManagerMemory() { + return Optional.ofNullable(getConfiguration().get(TaskManagerOptions.TOTAL_PROCESS_MEMORY)); + } + + @Override + public Optional<MemorySize> getTaskManagerMemoryFromSpec() { + return getJobDeployment() + .map( + flinkDeployment -> + MemorySize.parse( + flinkDeployment + .getSpec() + .getTaskManager() + .getResource() + .getMemory())); Review Comment: So the idea here is that the memory tuning works similarly to the config-based parallelism overrides, i.e. we do not change the original user spec but apply changes on top of the spec before the reconciliation loop runs. That way, we can always recover the original value. ########## flink-kubernetes-operator/src/main/java/org/apache/flink/kubernetes/operator/autoscaler/KubernetesJobAutoScalerContext.java: ########## @@ -56,19 +57,47 @@ public KubernetesJobAutoScalerContext( jobStatus, configuration, metricGroup, - Optional.ofNullable(configuration.get(KubernetesConfigOptions.TASK_MANAGER_CPU)) - .orElse(0.), - Optional.ofNullable(configuration.get(TaskManagerOptions.TOTAL_PROCESS_MEMORY)) - .orElse(MemorySize.ZERO), restClientSupplier); this.resourceContext = resourceContext; } + @Override + public Optional<Double> getTaskManagerCpu() { + return Optional.ofNullable( + getConfiguration().get(KubernetesConfigOptions.TASK_MANAGER_CPU)); + } + + @Override + public Optional<MemorySize> getTaskManagerMemory() { + return Optional.ofNullable(getConfiguration().get(TaskManagerOptions.TOTAL_PROCESS_MEMORY)); Review Comment: Sure thing. ########## flink-autoscaler/src/main/java/org/apache/flink/autoscaler/JobAutoScalerImpl.java: ########## @@ -151,7 +154,18 @@ protected void applyParallelismOverrides(Context ctx) throws Exception { userOverrides.put(k, v); } }); - scalingRealizer.realize(ctx, userOverrides); + scalingRealizer.realizeParallelismOverrides(ctx, userOverrides); + } + + @VisibleForTesting + void applyConfigOverrides(Context ctx) throws Exception { + Configuration configOverrides = stateStore.getConfigOverrides(ctx); + var tmHeapOverride = configOverrides.get(TaskManagerOptions.TASK_HEAP_MEMORY); + + if (tmHeapOverride != null) { + LOG.info("Applying heap memory override: {}", tmHeapOverride); + scalingRealizer.realizeMemoryOverrides(ctx, tmHeapOverride); Review Comment: Good catch, will rename. ########## flink-autoscaler/src/main/java/org/apache/flink/autoscaler/state/AutoScalerStateStore.java: ########## @@ -69,6 +70,13 @@ void storeParallelismOverrides(Context jobContext, Map<String, String> paralleli void removeParallelismOverrides(Context jobContext) throws Exception; + void storeConfigOverrides(Context jobContext, Configuration configOverrides) throws Exception; Review Comment: The idea is that those are applied on top of the regular config, that's why I chose configOverrides over config. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: issues-unsubscr...@flink.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
