[FLINK-31215] [autoscaler] Backpropagate processing rate limits from non-scalable bottlenecks to upstream operators

docs/layouts/shortcodes/generated/auto_scaler_configuration.html

@@ -116,6 +116,18 @@
             <td>Double</td>
             <td>Percentage threshold for switching to observed from busy time based true processing rate if the measurement is off by at least the configured fraction. For example 0.15 means we switch to observed if the busy time based computation is at least 15% higher during catchup.</td>
         </tr>
+        <tr>
+            <td><h5>job.autoscaler.processing.rate.backpropagation.enabled</h5></td>
+            <td style="word-wrap: break-word;">false</td>
+            <td>Boolean</td>
+            <td>Enable backpropagation of processing rate during autoscaling to reduce resources usage.</td>
+        </tr>
+        <tr>
+            <td><h5>job.autoscaler.processing.rate.backpropagation.impact</h5></td>
+            <td style="word-wrap: break-word;">0.0</td>
+            <td>Double</td>
+            <td>How strong should backpropagated values affect scaling. 0 - means no effect, 1 - use backpropagated values. It is not recommended to set this factor greater than 0.8</td>
+        </tr>
         <tr>
             <td><h5>job.autoscaler.quota.cpu</h5></td>
             <td style="word-wrap: break-word;">(none)</td>
@@ -210,7 +222,7 @@
             <td><h5>job.autoscaler.vertex.exclude.ids</h5></td>
             <td style="word-wrap: break-word;"></td>
             <td>List&lt;String&gt;</td>
-            <td>A (semicolon-separated) list of vertex ids in hexstring for which to disable scaling. Caution: For non-sink vertices this will still scale their downstream operators until https://issues.apache.org/jira/browse/FLINK-31215 is implemented.</td>
+            <td>A (semicolon-separated) list of vertex ids in hexstring for which to disable scaling.</td>
         </tr>
         <tr>
             <td><h5>job.autoscaler.vertex.max-parallelism</h5></td>

flink-autoscaler/src/main/java/org/apache/flink/autoscaler/JobVertexScaler.java

@@ -20,8 +20,10 @@
 import org.apache.flink.annotation.VisibleForTesting;
 import org.apache.flink.autoscaler.config.AutoScalerOptions;
 import org.apache.flink.autoscaler.event.AutoScalerEventHandler;
+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.autoscaler.topology.JobTopology;
 import org.apache.flink.autoscaler.topology.ShipStrategy;
 import org.apache.flink.autoscaler.utils.AutoScalerUtils;
 import org.apache.flink.configuration.Configuration;
@@ -36,6 +38,7 @@
 import java.time.Instant;
 import java.time.ZoneId;
 import java.util.Collection;
+import java.util.List;
 import java.util.Map;
 import java.util.SortedMap;
 
@@ -49,6 +52,7 @@
 import static org.apache.flink.autoscaler.metrics.ScalingMetric.EXPECTED_PROCESSING_RATE;
 import static org.apache.flink.autoscaler.metrics.ScalingMetric.MAX_PARALLELISM;
 import static org.apache.flink.autoscaler.metrics.ScalingMetric.PARALLELISM;
+import static org.apache.flink.autoscaler.metrics.ScalingMetric.TARGET_DATA_RATE;
 import static org.apache.flink.autoscaler.metrics.ScalingMetric.TRUE_PROCESSING_RATE;
 import static org.apache.flink.autoscaler.topology.ShipStrategy.HASH;
 
@@ -150,6 +154,116 @@ public int computeScaleTargetParallelism(
         return newParallelism;
     }
 
+    public void propagateBackpropScaleFactor(
+            Configuration conf,
+            JobVertexID vertex,
+            JobTopology topology,
+            EvaluatedMetrics evaluatedMetrics,
+            Map<JobVertexID, Double> realDataRates,
+            List<String> excludedVertices) {
+
+        double targetDataRate =
+                evaluatedMetrics.getVertexMetrics().get(vertex).get(TARGET_DATA_RATE).getCurrent();
+
+        // we cannot adjust data rate of the vertex if data rate is undefined or the vertex is
+        // excluded from scaling
+        if (Double.isNaN(targetDataRate) || excludedVertices.contains(vertex.toString())) {
+            LOG.debug(
+                    "Vertex {} is excluded from scaling or it's target data rate is undefined. Excluding it from backpropagation",
+                    vertex.toHexString());
+            realDataRates.put(vertex, targetDataRate);
+            return;
+        }
+
+        // how target data rate should be lowered
+        double dataRateDecrease = 0.0;
+
+        // check how the data rate of the vertex should be lowered depending on downstream
+        for (var downstream : topology.getVertexInfos().get(vertex).getOutputs().keySet()) {
+            double downstreamTargetDataRate =
+                    evaluatedMetrics
+                            .getVertexMetrics()
+                            .get(downstream)
+                            .get(TARGET_DATA_RATE)
+                            .getCurrent();
+            double downstreamRealDataRate = realDataRates.getOrDefault(downstream, Double.NaN);
+            int upstreamVertices = topology.getVertexInfos().get(downstream).getInputs().size();
+
+            double outputRatio =
+                    topology.getVertexInfos()
+                            .get(downstream)
+                            .getInputRatios()
+                            .getOrDefault(vertex, Double.NaN);
+
+            // if real data rate cannot be updated by the downstream vertex
+            if (Double.isNaN(downstreamRealDataRate) || Double.isInfinite(downstreamRealDataRate)) {
+                continue;
+            }
+            if (Double.isNaN(downstreamTargetDataRate)
+                    || Double.isInfinite(downstreamTargetDataRate)) {
+                continue;
+            }
+            if (Double.isNaN(outputRatio) || Double.isInfinite(outputRatio)) {
+                continue;
+            }
+
+            // distribute downstream's data rate delta over all it's upstream vertices
+            double downstreamDataRateDelta =
+                    (downstreamTargetDataRate - downstreamRealDataRate) / upstreamVertices;
+
+            dataRateDecrease = Math.max(dataRateDecrease, downstreamDataRateDelta / outputRatio);
+        }
+
+        LOG.debug(
+                "Data rate of {} is decreased by {} from downstream",
+                vertex.toHexString(),
+                dataRateDecrease);
+
+        if (dataRateDecrease > targetDataRate) {
+            LOG.warn(
+                    "Required data rate decrease {} for vertex {} exceeds target data of the vertex {}.",
+                    dataRateDecrease,
+                    vertex.toHexString(),
+                    targetDataRate);
+            dataRateDecrease = targetDataRate;
+        }
+
+        targetDataRate -= dataRateDecrease;
+
+        // check, if target data rate should be lowered even more due to scaling limitations
+        double averageTrueProcessingRate =
+                evaluatedMetrics
+                        .getVertexMetrics()
+                        .get(vertex)
+                        .get(TRUE_PROCESSING_RATE)
+                        .getAverage();
+
+        if (Double.isNaN(averageTrueProcessingRate)) {
+            LOG.info(
+                    "True Processing Rate of {} is undefined, use target data rate adjusted by downstream",
+                    vertex);
+            realDataRates.put(vertex, targetDataRate);
+            return;
+        }
+
+        // determine upper limit for scaling of the vertex
+        double parallelism =
+                evaluatedMetrics.getVertexMetrics().get(vertex).get(PARALLELISM).getCurrent();
+        double maxParallelism =
+                evaluatedMetrics.getVertexMetrics().get(vertex).get(MAX_PARALLELISM).getCurrent();
+        double maxScaleFactor =
+                Math.min(1 + conf.get(MAX_SCALE_UP_FACTOR), maxParallelism / parallelism);
+
+        double maxDataProcessingRate = maxScaleFactor * averageTrueProcessingRate;
+        targetDataRate = Math.min(targetDataRate, maxDataProcessingRate);
+
+        LOG.debug(
+                "Real data rate of {} after backpropagation is {}",
+                vertex.toHexString(),
+                targetDataRate);
+        realDataRates.put(vertex, targetDataRate);
+    }
+
     private boolean blockScalingBasedOnPastActions(
             Context context,
             JobVertexID vertex,

flink-autoscaler/src/main/java/org/apache/flink/autoscaler/ScalingExecutor.java

@@ -35,6 +35,7 @@
 import org.apache.flink.configuration.TaskManagerOptions;
 import org.apache.flink.runtime.instance.SlotSharingGroupId;
 import org.apache.flink.runtime.jobgraph.JobVertexID;
+import org.apache.flink.util.Preconditions;
 
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
@@ -49,6 +50,7 @@
 import java.util.SortedMap;
 
 import static org.apache.flink.autoscaler.config.AutoScalerOptions.EXCLUDED_PERIODS;
+import static org.apache.flink.autoscaler.config.AutoScalerOptions.PROCESSING_RATE_BACKPROPAGATION_ENABLED;
 import static org.apache.flink.autoscaler.config.AutoScalerOptions.SCALING_ENABLED;
 import static org.apache.flink.autoscaler.config.AutoScalerOptions.SCALING_EVENT_INTERVAL;
 import static org.apache.flink.autoscaler.event.AutoScalerEventHandler.SCALING_EXECUTION_DISABLED_REASON;
@@ -57,6 +59,7 @@
 import static org.apache.flink.autoscaler.metrics.ScalingHistoryUtils.addToScalingHistoryAndStore;
 import static org.apache.flink.autoscaler.metrics.ScalingMetric.SCALE_DOWN_RATE_THRESHOLD;
 import static org.apache.flink.autoscaler.metrics.ScalingMetric.SCALE_UP_RATE_THRESHOLD;
+import static org.apache.flink.autoscaler.metrics.ScalingMetric.TARGET_DATA_RATE;
 import static org.apache.flink.autoscaler.metrics.ScalingMetric.TRUE_PROCESSING_RATE;
 
 /** Class responsible for executing scaling decisions. */
@@ -221,8 +224,14 @@ Map<JobVertexID, ScalingSummary> computeScalingSummary(
         }
 
         var out = new HashMap<JobVertexID, ScalingSummary>();
+
+        if (context.getConfiguration().get(PROCESSING_RATE_BACKPROPAGATION_ENABLED)) {
+            backpropagateProcessingRate(context, evaluatedMetrics, restartTime, jobTopology);
+        }
+
         var excludeVertexIdList =
                 context.getConfiguration().get(AutoScalerOptions.VERTEX_EXCLUDE_IDS);
+
         evaluatedMetrics
                 .getVertexMetrics()
                 .forEach(
@@ -255,6 +264,73 @@ Map<JobVertexID, ScalingSummary> computeScalingSummary(
         return out;
     }
 
+    private void backpropagateProcessingRate(
+            Context context,
+            EvaluatedMetrics evaluatedMetrics,
+            Duration restartTime,
+            JobTopology jobTopology) {
+        var conf = context.getConfiguration();
+        double backpropagationImpact =
+                conf.get(AutoScalerOptions.PROCESSING_RATE_BACKPROPAGATION_IMPACT);
+        Preconditions.checkState(
+                0 <= backpropagationImpact && backpropagationImpact <= 1.0,
+                "Backpropagation impact should be in range [0, 1]");
+        var realDataRates = new HashMap<JobVertexID, Double>();
+        var excludeVertexIdList =
+                context.getConfiguration().get(AutoScalerOptions.VERTEX_EXCLUDE_IDS);
+        var vertexIterator =
+                jobTopology
+                        .getVerticesInTopologicalOrder()
+                        .listIterator(jobTopology.getVerticesInTopologicalOrder().size());
+
+        // backpropagate scale factors
+        while (vertexIterator.hasPrevious()) {
+            var vertex = vertexIterator.previous();
+            jobVertexScaler.propagateBackpropScaleFactor(
+                    conf,
+                    vertex,
+                    jobTopology,
+                    evaluatedMetrics,
+                    realDataRates,
+                    excludeVertexIdList);
+        }
+
+        // use an extra map to not lose precision
+        Map<JobVertexID, Double> adjustedDataRate = new HashMap<>();
+
+        // re-evaluating vertices capacity
+        // Target data rate metric is rewritten for parallelism evaluation
+        for (var vertex : jobTopology.getVerticesInTopologicalOrder()) {
+            double adjustedCapacity = 0.0;
+
+            if (jobTopology.isSource(vertex)) {
+                adjustedCapacity +=
+                        evaluatedMetrics
+                                                .getVertexMetrics()
+                                                .get(vertex)
+                                                .get(TARGET_DATA_RATE)
+                                                .getAverage()
+                                        * (1.0 - backpropagationImpact)
+                                + backpropagationImpact * realDataRates.get(vertex);
+            } else {
+                for (var input : jobTopology.getVertexInfos().get(vertex).getInputs().keySet()) {
+                    adjustedCapacity +=
+                            adjustedDataRate.get(input)
+                                    * jobTopology
+                                            .getVertexInfos()
+                                            .get(vertex)
+                                            .getInputRatios()
+                                            .get(input);
+                }
+            }
+            adjustedDataRate.put(vertex, adjustedCapacity);
+            evaluatedMetrics
+                    .getVertexMetrics()
+                    .get(vertex)
+                    .put(TARGET_DATA_RATE, EvaluatedScalingMetric.avg(adjustedCapacity));
+        }
+    }
+
     private boolean isJobUnderMemoryPressure(
             Context ctx, Map<ScalingMetric, EvaluatedScalingMetric> evaluatedMetrics) {
 

flink-autoscaler/src/main/java/org/apache/flink/autoscaler/ScalingMetricEvaluator.java

@@ -346,6 +346,7 @@ private void computeTargetDataRate(
                 var inputTargetRate = inputEvaluatedMetrics.get(TARGET_DATA_RATE);
                 var outputRatio =
                         computeEdgeOutputRatio(inputVertex, vertex, topology, metricsHistory);
+                topology.get(vertex).getInputRatios().put(inputVertex, outputRatio);
                 LOG.debug(
                         "Computed output ratio for edge ({} -> {}) : {}",
                         inputVertex,

flink-autoscaler/src/main/java/org/apache/flink/autoscaler/config/AutoScalerOptions.java

@@ -58,6 +58,24 @@ private static ConfigOptions.OptionBuilder autoScalerConfig(String key) {
                     .withDescription(
                             "Enable vertex scaling execution by the autoscaler. If disabled, the autoscaler will only collect metrics and evaluate the suggested parallelism for each vertex but will not upgrade the jobs.");
 
+    public static final ConfigOption<Boolean> PROCESSING_RATE_BACKPROPAGATION_ENABLED =
+            autoScalerConfig("processing.rate.backpropagation.enabled")
+                    .booleanType()
+                    .defaultValue(false)
+                    .withFallbackKeys(
+                            oldOperatorConfigKey("processing.rate.backpropagation.enabled"))
+                    .withDescription(
+                            "Enable backpropagation of processing rate during autoscaling to reduce resources usage.");
+
+    public static final ConfigOption<Double> PROCESSING_RATE_BACKPROPAGATION_IMPACT =
+            autoScalerConfig("processing.rate.backpropagation.impact")
+                    .doubleType()
+                    .defaultValue(0.0)
+                    .withFallbackKeys(
+                            oldOperatorConfigKey("processing.rate.backpropagation.impact"))
+                    .withDescription(
+                            "How strong should backpropagated values affect scaling. 0 - means no effect, 1 - use backpropagated values. It is not recommended to set this factor greater than 0.8");
+
     public static final ConfigOption<Duration> METRICS_WINDOW =
             autoScalerConfig("metrics.window")
                     .durationType()
@@ -313,7 +331,7 @@ private static ConfigOptions.OptionBuilder autoScalerConfig(String key) {
                     .defaultValues()
                     .withFallbackKeys(oldOperatorConfigKey("vertex.exclude.ids"))
                     .withDescription(
-                            "A (semicolon-separated) list of vertex ids in hexstring for which to disable scaling. Caution: For non-sink vertices this will still scale their downstream operators until https://issues.apache.org/jira/browse/FLINK-31215 is implemented.");
+                            "A (semicolon-separated) list of vertex ids in hexstring for which to disable scaling.");
 
     public static final ConfigOption<Duration> SCALING_EVENT_INTERVAL =
             autoScalerConfig("scaling.event.interval")

flink-autoscaler/src/main/java/org/apache/flink/autoscaler/topology/VertexInfo.java

@@ -25,6 +25,7 @@
 import lombok.Data;
 import lombok.Setter;
 
+import java.util.HashMap;
 import java.util.Map;
 
 /** Job vertex information. */
@@ -36,6 +37,9 @@ public class VertexInfo {
     // All input vertices and the ship_strategy
     private final Map<JobVertexID, ShipStrategy> inputs;
 
+    // Output ratios from input vertices. Used for backpropagation
+    private final Map<JobVertexID, Double> inputRatios;
+
     private final SlotSharingGroupId slotSharingGroupId;
 
     // All output vertices and the ship_strategy
@@ -68,6 +72,7 @@ public VertexInfo(
         this.originalMaxParallelism = maxParallelism;
         this.finished = finished;
         this.ioMetrics = ioMetrics;
+        this.inputRatios = new HashMap<>();
     }
 
     @VisibleForTesting