diff options
author | Jon Bratseth <bratseth@gmail.com> | 2023-02-07 13:52:31 +0100 |
---|---|---|
committer | GitHub <noreply@github.com> | 2023-02-07 13:52:31 +0100 |
commit | f38de98559a5a3dd2ed7586bf689964db47bccc9 (patch) | |
tree | 00166c2e8f1a51e0af8a030eec7f0644ff32453e /controller-server/src/main/java/com/yahoo | |
parent | 628ab7604f016e4062a1e937436d35a8354a8dcc (diff) | |
parent | 9be10074a8b0e31c9dfcdcb2d8e9048c86b1fadc (diff) |
Merge pull request #25912 from vespa-engine/bratseth/bcp-rebased
Bratseth/bcp rebased
Diffstat (limited to 'controller-server/src/main/java/com/yahoo')
-rw-r--r-- | controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/TrafficShareUpdater.java | 111 |
1 files changed, 96 insertions, 15 deletions
diff --git a/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/TrafficShareUpdater.java b/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/TrafficShareUpdater.java index 832dbb6b921..aea01ae36d3 100644 --- a/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/TrafficShareUpdater.java +++ b/controller-server/src/main/java/com/yahoo/vespa/hosted/controller/maintenance/TrafficShareUpdater.java @@ -1,6 +1,10 @@ // Copyright Yahoo. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root. package com.yahoo.vespa.hosted.controller.maintenance; +import com.yahoo.config.application.api.Bcp; +import com.yahoo.config.application.api.DeploymentSpec; +import com.yahoo.config.provision.InstanceName; +import com.yahoo.config.provision.RegionName; import com.yahoo.vespa.hosted.controller.ApplicationController; import com.yahoo.vespa.hosted.controller.Controller; import com.yahoo.vespa.hosted.controller.Instance; @@ -8,7 +12,11 @@ import com.yahoo.vespa.hosted.controller.api.integration.configserver.NodeReposi import com.yahoo.vespa.hosted.controller.application.Deployment; import java.time.Duration; +import java.util.ArrayList; +import java.util.List; +import java.util.Map; import java.util.logging.Level; +import java.util.stream.Collectors; /** * This computes, for every application deployment @@ -41,12 +49,11 @@ public class TrafficShareUpdater extends ControllerMaintainer { int failures = 0; for (var application : applications.asList()) { for (var instance : application.instances().values()) { - for (var deployment : instance.deployments().values()) { - if ( ! deployment.zone().environment().isProduction()) continue; + for (var deployment : instance.productionDeployments().values()) { if (shuttingDown()) return 1.0; try { attempts++; - updateTrafficFraction(instance, deployment); + updateTrafficFraction(instance, deployment, application.deploymentSpec()); } catch (Exception e) { // Some failures due to locked applications are expected and benign @@ -62,20 +69,94 @@ public class TrafficShareUpdater extends ControllerMaintainer { return successFactor; } - private void updateTrafficFraction(Instance instance, Deployment deployment) { - double qpsInZone = deployment.metrics().queriesPerSecond(); - double totalQps = instance.deployments().values().stream() - .filter(i -> i.zone().environment().isProduction()) - .mapToDouble(i -> i.metrics().queriesPerSecond()).sum(); - long prodRegions = instance.deployments().values().stream() - .filter(i -> i.zone().environment().isProduction()) - .count(); - double currentReadShare = totalQps == 0 ? 0 : qpsInZone / totalQps; - double maxReadShare = prodRegions < 2 ? 1.0 : 1.0 / ( prodRegions - 1.0); - if (currentReadShare > maxReadShare) // This can happen because the assumption of equal traffic - maxReadShare = currentReadShare; // distribution can be incorrect + private void updateTrafficFraction(Instance instance, Deployment deployment, DeploymentSpec deploymentSpec) { + // maxReadShare / currentReadShare = how much additional traffic must the zone be able to handle + double currentReadShare = 0; // How much of the total traffic of the group(s) this is a member of does this deployment receive + double maxReadShare = 0; // How much of the total traffic of the group(s) this is a member of might this deployment receive if a member of the group fails + for (BcpGroup group : BcpGroup.groupsFrom(instance, deploymentSpec)) { + if ( ! group.contains(deployment.zone().region())) continue; + double deploymentQps = deployment.metrics().queriesPerSecond(); + double groupQps = group.totalQps(); + double fraction = group.fraction(deployment.zone().region()); + currentReadShare += groupQps == 0 ? 0 : fraction * deploymentQps / groupQps; + maxReadShare += group.size() == 1 + ? currentReadShare + : fraction * ( deploymentQps + group.maxQpsExcluding(deployment.zone().region()) / (group.size() - 1) ) / groupQps; + } nodeRepository.patchApplication(deployment.zone(), instance.id(), currentReadShare, maxReadShare); } + /** + * A set of regions which will take over traffic from each other if one of them fails. + * Each region will take an equal share (modulated by fraction) of the failing region's traffic. + * + * A regions membership in a group may be partial, represented by a fraction [0, 1], + * in which case the other regions will collectively only take that fraction of the failing regions traffic, + * and symmetrically, the region will only take its fraction of its share of traffic of any other failing region. + */ + private static class BcpGroup { + + /** The instance which has this group. */ + private final Instance instance; + + /** Regions in this group, with their fractions. */ + private final Map<RegionName, Double> regions; + + /** Creates a group of a subset of the deployments in this instance. */ + private BcpGroup(Instance instance, Map<RegionName, Double> regions) { + this.instance = instance; + this.regions = regions; + } + + /** Returns the sum of the fractional memberships of this. */ + double size() { + return regions.values().stream().mapToDouble(f -> f).sum(); + } + + double fraction(RegionName region) { + return regions.getOrDefault(region, 0.0); + } + + boolean contains(RegionName region) { + return regions.containsKey(region); + } + + double totalQps() { + return instance.productionDeployments().values().stream() + .mapToDouble(i -> i.metrics().queriesPerSecond()).sum(); + } + + double maxQpsExcluding(RegionName region) { + return instance.productionDeployments().values().stream() + .filter(d -> ! d.zone().region().equals(region)) + .mapToDouble(d -> d.metrics().queriesPerSecond() * fraction(d.zone().region())) + .max() + .orElse(0); + } + private static Bcp bcpOf(InstanceName instanceName, DeploymentSpec deploymentSpec) { + var instanceSpec = deploymentSpec.instance(instanceName); + if (instanceSpec.isEmpty()) return deploymentSpec.bcp(); + return instanceSpec.get().bcp().orElse(deploymentSpec.bcp()); + } + + private static Map<RegionName, Double> regionsFrom(Instance instance) { + return instance.productionDeployments().values().stream() + .collect(Collectors.toMap(deployment -> deployment.zone().region(), __ -> 1.0)); + } + + private static Map<RegionName, Double> regionsFrom(Bcp.Group groupSpec) { + return groupSpec.members().stream() + .collect(Collectors.toMap(member -> member.region(), member -> member.fraction())); + } + + static List<BcpGroup> groupsFrom(Instance instance, DeploymentSpec deploymentSpec) { + Bcp bcp = bcpOf(instance.name(), deploymentSpec); + if (bcp.isEmpty()) + return List.of(new BcpGroup(instance, regionsFrom(instance))); + return bcp.groups().stream().map(groupSpec -> new BcpGroup(instance, regionsFrom(groupSpec))).toList(); + } + + } + } |