Fly-Style commented on code in PR #18819: URL: https://github.com/apache/druid/pull/18819#discussion_r2620577303
########## indexing-service/src/main/java/org/apache/druid/indexing/seekablestream/supervisor/autoscaler/CostBasedAutoScaler.java: ########## @@ -0,0 +1,259 @@ +/* + * 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.druid.indexing.seekablestream.supervisor.autoscaler; + +import org.apache.druid.indexing.overlord.supervisor.SupervisorSpec; +import org.apache.druid.indexing.overlord.supervisor.autoscaler.LagStats; +import org.apache.druid.indexing.overlord.supervisor.autoscaler.SupervisorTaskAutoScaler; +import org.apache.druid.indexing.seekablestream.supervisor.SeekableStreamSupervisor; +import org.apache.druid.java.util.common.StringUtils; +import org.apache.druid.java.util.common.concurrent.Execs; +import org.apache.druid.java.util.emitter.EmittingLogger; +import org.apache.druid.java.util.emitter.service.ServiceEmitter; +import org.apache.druid.java.util.emitter.service.ServiceMetricEvent; +import org.apache.druid.query.DruidMetrics; + +import java.util.ArrayList; +import java.util.List; +import java.util.concurrent.Callable; +import java.util.concurrent.ScheduledExecutorService; +import java.util.concurrent.TimeUnit; + +/** + * Cost-based auto-scaler for seekable stream supervisors. + * Uses a cost function combining lag and idle time metrics to determine optimal task counts. + * Task counts are selected from pre-calculated values (not arbitrary factors). + * Scale-up happens incrementally, scale-down only during task rollover. + */ +public class CostBasedAutoScaler implements SupervisorTaskAutoScaler +{ + private static final EmittingLogger log = new EmittingLogger(CostBasedAutoScaler.class); + + private static final int SCALE_UP_FACTOR_DISCRETE_DISTANCE = 2; + private static final int SCALE_DOWN_FACTOR_DISCRETE_DISTANCE = SCALE_UP_FACTOR_DISCRETE_DISTANCE * 2; + public static final String OPTIMAL_TASK_COUNT_METRIC = "task/autoScaler/costBased/optimalTaskCount"; + + private final String supervisorId; + private final SeekableStreamSupervisor supervisor; + private final ServiceEmitter emitter; + private final SupervisorSpec spec; + private final CostBasedAutoScalerConfig config; + private final ServiceMetricEvent.Builder metricBuilder; + private final ScheduledExecutorService autoscalerExecutor; + private final WeightedCostFunction costFunction; + + public CostBasedAutoScaler( + SeekableStreamSupervisor supervisor, + CostBasedAutoScalerConfig config, + SupervisorSpec spec, + ServiceEmitter emitter + ) + { + this.config = config; + this.spec = spec; + this.supervisor = supervisor; + this.supervisorId = spec.getId(); + this.emitter = emitter; + + this.costFunction = new WeightedCostFunction(); + + this.autoscalerExecutor = Execs.scheduledSingleThreaded(StringUtils.encodeForFormat(spec.getId())); + this.metricBuilder = ServiceMetricEvent.builder() + .setDimension(DruidMetrics.SUPERVISOR_ID, supervisorId) + .setDimension( + DruidMetrics.STREAM, + this.supervisor.getIoConfig().getStream() + ); + } + + @Override + public void start() + { + Callable<Integer> scaleAction = () -> computeOptimalTaskCount(this.collectMetrics()); + Runnable onSuccessfulScale = () -> { + }; + + autoscalerExecutor.scheduleAtFixedRate( + supervisor.buildDynamicAllocationTask(scaleAction, onSuccessfulScale, emitter), + config.getScaleActionPeriodMillis(), + config.getScaleActionPeriodMillis(), + TimeUnit.MILLISECONDS + ); + + log.info( + "CostBasedAutoScaler started for supervisorId [%s]: evaluating scaling every [%d]ms", + supervisorId, + config.getScaleActionPeriodMillis() + ); + } + + @Override + public void stop() + { + autoscalerExecutor.shutdownNow(); + log.info("CostBasedAutoScaler stopped for supervisorId [%s]", supervisorId); + } + + @Override + public void reset() + { + // No-op. + } + + private CostMetrics collectMetrics() + { + if (spec.isSuspended()) { + log.debug("Supervisor [%s] is suspended, skipping a metrics collection", supervisorId); + return null; + } + + final LagStats lagStats = supervisor.computeLagStats(); + if (lagStats == null) { + log.debug("Lag stats unavailable for supervisorId [%s], skipping collection", supervisorId); + return null; + } + + final int currentTaskCount = supervisor.getIoConfig().getTaskCount(); + final int partitionCount = supervisor.getPartitionCount(); + final double pollIdleRatio = supervisor.getPollIdleRatioMetric(); + + return new CostMetrics(lagStats.getAvgLag(), currentTaskCount, partitionCount, pollIdleRatio); + } + + /** + * Computes the optimal task count based on current metrics. + * <p> + * Returns -1 (no scaling needed) in the following cases: + * <ul> + * <li>Metrics are not available</li> + * <li>The current idle ratio is in the ideal range [0.2, 0.6] - optimal utilization achieved</li> + * <li>Optimal task count equals current task count</li> + * </ul> + * + * @return optimal task count for scale-up, or -1 if no scaling action needed + */ + public int computeOptimalTaskCount(CostMetrics metrics) + { + if (metrics == null) { + log.debug("No metrics available yet for supervisorId [%s]", supervisorId); + return -1; + } + + final int partitionCount = metrics.getPartitionCount(); + final int currentTaskCount = metrics.getCurrentTaskCount(); + if (partitionCount <= 0 || currentTaskCount <= 0) { + return -1; + } + + final int[] validTaskCounts = CostBasedAutoScaler.computeValidTaskCounts(partitionCount, currentTaskCount); + + if (validTaskCounts.length == 0) { + log.warn("No valid task counts after applying constraints for supervisorId [%s]", supervisorId); + return -1; + } + + // If idle is already in the ideal range [0.2, 0.6], optimal utilization has been achieved. + // No scaling is needed - maintain stability by staying at current task count. + final double currentIdleRatio = metrics.getPollIdleRatio(); + if (currentIdleRatio >= 0 && WeightedCostFunction.isIdleInIdealRange(currentIdleRatio)) { + log.info( + "Idle ratio [%.3f] is in ideal range for supervisorId [%s], no scaling needed", + currentIdleRatio, + supervisorId + ); + return -1; + } + + // Update bounds with observed lag BEFORE optimization loop + // This ensures normalization uses historical observed values, not predicted values + costFunction.updateLagBounds(metrics.getAvgPartitionLag()); + + int optimalTaskCount = -1; + double optimalCost = Double.POSITIVE_INFINITY; + + for (int taskCount : validTaskCounts) { + double cost = costFunction.computeCost(metrics, taskCount, config); + log.debug("Proposed task count: %d, Cost: %.4f", taskCount, cost); + if (cost < optimalCost) { + optimalTaskCount = taskCount; + optimalCost = cost; + } + } + + emitter.emit(metricBuilder.setMetric(OPTIMAL_TASK_COUNT_METRIC, (long) optimalTaskCount)); + + log.debug( + "Cost-based scaling evaluation for supervisorId [%s]: current=%d, optimal=%d, cost=%.4f, " + + "avgPartitionLag=%.2f, pollIdleRatio=%.3f", + supervisorId, + metrics.getCurrentTaskCount(), + optimalTaskCount, + optimalCost, + metrics.getAvgPartitionLag(), + metrics.getPollIdleRatio() + ); + + if (optimalTaskCount == currentTaskCount) { + return -1; + } + // Temporarily, we equalize scaleup and scaledown effects, due to uncompleted state of task rollover. + // The behaviour will be changed by complementing task rollover state machine. + return optimalTaskCount; + } + + /** + * Generates valid task counts based on partitions-per-task ratios. + * This enables gradual scaling and avoids large jumps. + * Limits the range of task counts considered to avoid excessive computation. + * + * @return sorted list of valid task counts within bounds + */ + static int[] computeValidTaskCounts(int partitionCount, int currentTaskCount) + { + if (partitionCount <= 0) { + return new int[]{}; + } + + List<Integer> result = new ArrayList<>(); + final int currentPartitionsPerTask = partitionCount / currentTaskCount; + // To avoid confusion: minimum partitions per task means maximum amount of tasks (scale up) and vice versa. + final int minPartitionsPerTask = Math.max(1, currentPartitionsPerTask - SCALE_UP_FACTOR_DISCRETE_DISTANCE); + final int maxPartitionsPerTask = Math.min( + partitionCount, + Math.max( + minPartitionsPerTask, + currentPartitionsPerTask + SCALE_DOWN_FACTOR_DISCRETE_DISTANCE + ) + ); + + for (int partitionsPerTask = maxPartitionsPerTask; partitionsPerTask >= minPartitionsPerTask; partitionsPerTask--) { + final int taskCount = (partitionCount + partitionsPerTask - 1) / partitionsPerTask; + if (result.isEmpty() || result.get(result.size() - 1) != taskCount) { + result.add(taskCount); + } + } + return result.stream().mapToInt(Integer::intValue).toArray(); Review Comment: It's my personal belief: not a fan of boxed primitives A LOT :) -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
