diff options
Diffstat (limited to 'node-repository/src/main/java')
15 files changed, 108 insertions, 73 deletions
diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/applications/Cluster.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/applications/Cluster.java index e0ccbe10b10..ad20f68ca33 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/applications/Cluster.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/applications/Cluster.java @@ -26,7 +26,7 @@ public class Cluster { private final ClusterSpec.Id id; private final boolean exclusive; private final ClusterResources min, max; - private boolean required; + private final boolean required; private final Optional<Suggestion> suggested; private final Optional<ClusterResources> target; diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/AllocatableClusterResources.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/AllocatableClusterResources.java index 078b0621a99..849ea03665b 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/AllocatableClusterResources.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/AllocatableClusterResources.java @@ -13,6 +13,7 @@ import com.yahoo.vespa.hosted.provision.provisioning.NodeResourceLimits; import java.util.List; import java.util.Optional; +import java.util.stream.Collectors; /** * @author bratseth @@ -139,23 +140,21 @@ public class AllocatableClusterResources { public static Optional<AllocatableClusterResources> from(ClusterResources wantedResources, ClusterSpec clusterSpec, Limits applicationLimits, - boolean required, NodeList hosts, NodeRepository nodeRepository) { - var capacityPolicies = new CapacityPolicies(nodeRepository); var systemLimits = new NodeResourceLimits(nodeRepository); boolean exclusive = clusterSpec.isExclusive(); - int actualNodes = capacityPolicies.decideSize(wantedResources.nodes(), required, true, false, clusterSpec); if ( !clusterSpec.isExclusive() && !nodeRepository.zone().getCloud().dynamicProvisioning()) { // We decide resources: Add overhead to what we'll request (advertised) to make sure real becomes (at least) cappedNodeResources var advertisedResources = nodeRepository.resourcesCalculator().realToRequest(wantedResources.nodeResources(), exclusive); advertisedResources = systemLimits.enlargeToLegal(advertisedResources, clusterSpec.type(), exclusive); // Ask for something legal advertisedResources = applicationLimits.cap(advertisedResources); // Overrides other conditions, even if it will then fail - advertisedResources = capacityPolicies.decideNodeResources(advertisedResources, required, clusterSpec); // Adjust to what we can request var realResources = nodeRepository.resourcesCalculator().requestToReal(advertisedResources, exclusive); // What we'll really get - if ( ! systemLimits.isWithinRealLimits(realResources, clusterSpec.type())) return Optional.empty(); + if ( ! systemLimits.isWithinRealLimits(realResources, clusterSpec.type())) + return Optional.empty(); + if (matchesAny(hosts, advertisedResources)) - return Optional.of(new AllocatableClusterResources(wantedResources.withNodes(actualNodes).with(realResources), + return Optional.of(new AllocatableClusterResources(wantedResources.with(realResources), advertisedResources, wantedResources, clusterSpec)); @@ -168,7 +167,6 @@ public class AllocatableClusterResources { for (Flavor flavor : nodeRepository.flavors().getFlavors()) { // Flavor decide resources: Real resources are the worst case real resources we'll get if we ask for these advertised resources NodeResources advertisedResources = nodeRepository.resourcesCalculator().advertisedResourcesOf(flavor); - advertisedResources = capacityPolicies.decideNodeResources(advertisedResources, required, clusterSpec); // Adjust to what we can get NodeResources realResources = nodeRepository.resourcesCalculator().requestToReal(advertisedResources, exclusive); // Adjust where we don't need exact match to the flavor @@ -184,7 +182,7 @@ public class AllocatableClusterResources { if ( ! between(applicationLimits.min().nodeResources(), applicationLimits.max().nodeResources(), advertisedResources)) continue; if ( ! systemLimits.isWithinRealLimits(realResources, clusterSpec.type())) continue; - var candidate = new AllocatableClusterResources(wantedResources.withNodes(actualNodes).with(realResources), + var candidate = new AllocatableClusterResources(wantedResources.with(realResources), advertisedResources, wantedResources, clusterSpec); diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/AllocationOptimizer.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/AllocationOptimizer.java index b8a80a9bd2b..30432c1c078 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/AllocationOptimizer.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/AllocationOptimizer.java @@ -66,15 +66,12 @@ public class AllocationOptimizer { groupsAdjustedForRedundancy, limits, target, current, clusterModel)); var allocatableResources = AllocatableClusterResources.from(next, current.clusterSpec(), limits, - clusterModel.cluster().required(), hosts, nodeRepository); - if (allocatableResources.isEmpty()) continue; if (bestAllocation.isEmpty() || allocatableResources.get().preferableTo(bestAllocation.get())) bestAllocation = allocatableResources; } } - return bestAllocation; } diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/ClusterModel.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/ClusterModel.java index e1e670c5b01..3c26eef41d9 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/ClusterModel.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/autoscale/ClusterModel.java @@ -32,9 +32,11 @@ public class ClusterModel { static final double idealQueryCpuLoad = 0.8; static final double idealWriteCpuLoad = 0.95; static final double idealMemoryLoad = 0.65; - static final double idealDiskLoad = 0.6; + static final double idealContainerDiskLoad = 0.95; + static final double idealContentDiskLoad = 0.6; private final Application application; + private final ClusterSpec clusterSpec; private final Cluster cluster; /** The current nodes of this cluster, or empty if this models a new cluster not yet deployed */ private final NodeList nodes; @@ -54,6 +56,7 @@ public class ClusterModel { MetricsDb metricsDb, Clock clock) { this.application = application; + this.clusterSpec = clusterSpec; this.cluster = cluster; this.nodes = clusterNodes; this.clock = clock; @@ -64,12 +67,14 @@ public class ClusterModel { /** For testing */ ClusterModel(Application application, + ClusterSpec clusterSpec, Cluster cluster, Clock clock, Duration scalingDuration, ClusterTimeseries clusterTimeseries, ClusterNodesTimeseries nodeTimeseries) { this.application = application; + this.clusterSpec = clusterSpec; this.cluster = cluster; this.nodes = null; this.clock = clock; @@ -79,6 +84,8 @@ public class ClusterModel { this.nodeTimeseries = nodeTimeseries; } + public Application application() { return application; } + public ClusterSpec clusterSpec() { return clusterSpec; } public Cluster cluster() { return cluster; } /** Returns the predicted duration of a rescaling of this cluster */ @@ -110,7 +117,7 @@ public class ClusterModel { public Load averageLoad() { return nodeTimeseries().averageLoad(clock.instant().minus(scalingDuration())); } public Load idealLoad() { - return new Load(idealCpuLoad(), idealMemoryLoad, idealDiskLoad); + return new Load(idealCpuLoad(), idealMemoryLoad, idealDiskLoad()); } /** Ideal cpu load must take the application traffic fraction into account */ @@ -193,6 +200,12 @@ public class ClusterModel { return duration; } + private double idealDiskLoad() { + // Stateless clusters are not expected to consume more disk over time - + // if they do it is due to logs which will be rotated away right before the disk is full + return clusterSpec.isStateful() ? idealContentDiskLoad : idealContainerDiskLoad; + } + /** * Create a cluster model if possible and logs a warning and returns empty otherwise. * This is useful in cases where it's possible to continue without the cluser model, diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/LoadBalancerExpirer.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/LoadBalancerExpirer.java index 3b74533772b..fbc3d236421 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/LoadBalancerExpirer.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/LoadBalancerExpirer.java @@ -79,7 +79,7 @@ public class LoadBalancerExpirer extends NodeRepositoryMaintainer { allocatedNodes(lb.id()).isEmpty(), lb -> { try { attempts.add(1); - log.log(Level.INFO, () -> "Removing expired inactive load balancer " + lb.id()); + log.log(Level.INFO, () -> "Removing expired inactive " + lb.id()); service.remove(lb.id().application(), lb.id().cluster()); db.removeLoadBalancer(lb.id()); } catch (Exception e){ diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/NodeMover.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/NodeMover.java index 6c103627ad4..57db874fb84 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/NodeMover.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/NodeMover.java @@ -59,7 +59,7 @@ public abstract class NodeMover<MOVE> extends NodeRepositoryMaintainer { protected final MOVE findBestMove(NodesAndHosts<? extends NodeList> allNodes) { HostCapacity capacity = new HostCapacity(allNodes, nodeRepository().resourcesCalculator()); MOVE bestMove = emptyMove; - // Shuffle nodes so we did not get stuck if the chosen move is consistently discarded. Node moves happen through + // Shuffle nodes to not get stuck if the chosen move is consistently discarded. Node moves happen through // a soft request to retire (preferToRetire), which node allocation can disregard NodeList activeNodes = allNodes.nodes().nodeType(NodeType.tenant) .state(Node.State.active) diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/SwitchRebalancer.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/SwitchRebalancer.java index 7bea671fbac..f01e8ecd301 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/SwitchRebalancer.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/maintenance/SwitchRebalancer.java @@ -16,6 +16,7 @@ import java.time.Duration; import java.util.HashSet; import java.util.List; import java.util.Set; +import java.util.logging.Logger; /** * Ensure that nodes within a cluster a spread across hosts on exclusive network switches. @@ -24,6 +25,8 @@ import java.util.Set; */ public class SwitchRebalancer extends NodeMover<Move> { + private static final Logger LOG = Logger.getLogger(SwitchRebalancer.class.getName()); + private final Metric metric; private final Deployer deployer; @@ -40,7 +43,12 @@ public class SwitchRebalancer extends NodeMover<Move> { NodesAndHosts<NodeList> allNodes = NodesAndHosts.create(nodeRepository().nodes().list()); // Lockless as strong consistency is not needed if (!zoneIsStable(allNodes.nodes())) return 1.0; - findBestMove(allNodes).execute(false, Agent.SwitchRebalancer, deployer, metric, nodeRepository()); + Move bestMove = findBestMove(allNodes); + if (!bestMove.isEmpty()) { + LOG.info("Trying " + bestMove + " (" + bestMove.fromHost().switchHostname().orElse("<none>") + + " -> " + bestMove.toHost().switchHostname().orElse("<none>") + ")"); + } + bestMove.execute(false, Agent.SwitchRebalancer, deployer, metric, nodeRepository()); return 1.0; } diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/persistence/NodeResourcesSerializer.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/persistence/NodeResourcesSerializer.java index 8c421443a65..1c3d3f5c489 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/persistence/NodeResourcesSerializer.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/persistence/NodeResourcesSerializer.java @@ -20,6 +20,7 @@ public class NodeResourcesSerializer { private static final String storageTypeKey = "storageType"; static void toSlime(NodeResources resources, Cursor resourcesObject) { + if (resources.isUnspecified()) return; resourcesObject.setDouble(vcpuKey, resources.vcpu()); resourcesObject.setDouble(memoryKey, resources.memoryGb()); resourcesObject.setDouble(diskKey, resources.diskGb()); @@ -29,6 +30,7 @@ public class NodeResourcesSerializer { } static NodeResources resourcesFromSlime(Inspector resources) { + if ( ! resources.field(vcpuKey).valid()) return NodeResources.unspecified(); return new NodeResources(resources.field(vcpuKey).asDouble(), resources.field(memoryKey).asDouble(), resources.field(diskKey).asDouble(), diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/Activator.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/Activator.java index 0d32b21016c..8c358301b85 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/Activator.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/Activator.java @@ -113,7 +113,8 @@ class Activator { var cluster = modified.cluster(clusterEntry.getKey()).get(); var previousResources = oldNodes.cluster(clusterEntry.getKey()).toResources(); var currentResources = clusterEntry.getValue().toResources(); - if ( ! previousResources.justNumbers().equals(currentResources.justNumbers())) { + if ( previousResources.nodeResources().isUnspecified() + || ! previousResources.justNumbers().equals(currentResources.justNumbers())) { cluster = cluster.with(ScalingEvent.create(previousResources, currentResources, generation, at)); } if (cluster.targetResources().isPresent() diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/CapacityPolicies.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/CapacityPolicies.java index 0c2c3c48df1..4088d717a67 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/CapacityPolicies.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/CapacityPolicies.java @@ -3,6 +3,7 @@ package com.yahoo.vespa.hosted.provision.provisioning; import com.yahoo.config.provision.ApplicationId; import com.yahoo.config.provision.Capacity; +import com.yahoo.config.provision.ClusterResources; import com.yahoo.config.provision.ClusterSpec; import com.yahoo.config.provision.Environment; import com.yahoo.config.provision.NodeResources; @@ -29,10 +30,21 @@ public class CapacityPolicies { this.sharedHosts = type -> PermanentFlags.SHARED_HOST.bindTo(nodeRepository.flagSource()).value().isEnabled(type.name()); } - public int decideSize(int requested, boolean required, boolean canFail, boolean isTester, ClusterSpec cluster) { + public Capacity applyOn(Capacity capacity, ApplicationId application) { + return capacity.withLimits(applyOn(capacity.minResources(), capacity, application), + applyOn(capacity.maxResources(), capacity, application)); + } + + private ClusterResources applyOn(ClusterResources resources, Capacity capacity, ApplicationId application) { + int nodes = decideSize(resources.nodes(), capacity.isRequired(), application.instance().isTester()); + int groups = Math.min(resources.groups(), nodes); // cannot have more groups than nodes + var nodeResources = decideNodeResources(resources.nodeResources(), capacity.isRequired()); + return new ClusterResources(nodes, groups, nodeResources); + } + + private int decideSize(int requested, boolean required, boolean isTester) { if (isTester) return 1; - ensureRedundancy(requested, cluster, canFail); if (required) return requested; switch(zone.environment()) { case dev : case test : return 1; @@ -43,11 +55,9 @@ public class CapacityPolicies { } } - public NodeResources decideNodeResources(NodeResources target, boolean required, ClusterSpec cluster) { - if (target.isUnspecified()) - target = defaultNodeResources(cluster.type()); - + private NodeResources decideNodeResources(NodeResources target, boolean required) { if (required) return target; + if (target.isUnspecified()) return target; // Cannot be modified // Dev does not cap the cpu or network of containers since usage is spotty: Allocate just a small amount exclusively if (zone.environment() == Environment.dev && !zone.getCloud().dynamicProvisioning()) @@ -77,28 +87,11 @@ public class CapacityPolicies { } /** - * Whether or not the nodes requested can share physical host with other applications. + * Returns whether the nodes requested can share physical host with other applications. * A security feature which only makes sense for prod. */ public boolean decideExclusivity(Capacity capacity, boolean requestedExclusivity) { return requestedExclusivity && (capacity.isRequired() || zone.environment() == Environment.prod); } - /** - * Throw if the node count is 1 for container and content clusters and we're in a production zone - * - * @throws IllegalArgumentException if only one node is requested and we can fail - */ - private void ensureRedundancy(int nodeCount, ClusterSpec cluster, boolean canFail) { - if (canFail && - nodeCount == 1 && - requiresRedundancy(cluster.type()) && - zone.environment().isProduction()) - throw new IllegalArgumentException("Deployments to prod require at least 2 nodes per cluster for redundancy. Not fulfilled for " + cluster); - } - - private static boolean requiresRedundancy(ClusterSpec.Type clusterType) { - return clusterType.isContent() || clusterType.isContainer(); - } - } diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/GroupPreparer.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/GroupPreparer.java index ba46f0a9535..2d93763c631 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/GroupPreparer.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/GroupPreparer.java @@ -71,7 +71,7 @@ public class GroupPreparer { // Try preparing in memory without global unallocated lock. Most of the time there should be no changes and we // can return nodes previously allocated. NodeAllocation probeAllocation = prepareAllocation(application, cluster, requestedNodes, surplusActiveNodes, - indices::probeNext, wantedGroups, allNodesAndHosts); + indices::probeNext, wantedGroups, allNodesAndHosts); if (probeAllocation.fulfilledAndNoChanges()) { List<Node> acceptedNodes = probeAllocation.finalNodes(); surplusActiveNodes.removeAll(acceptedNodes); diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeAllocation.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeAllocation.java index 7cc4acc20b0..6c22a26d88a 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeAllocation.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeAllocation.java @@ -99,12 +99,12 @@ class NodeAllocation { * Note that if unallocated nodes are offered before allocated nodes, this will unnecessarily * reject allocated nodes due to index duplicates. * - * @param nodesPrioritized the nodes which are potentially on offer. These may belong to a different application etc. + * @param candidates the nodes which are potentially on offer. These may belong to a different application etc. * @return the subset of offeredNodes which was accepted, with the correct allocation assigned */ - List<Node> offer(List<NodeCandidate> nodesPrioritized) { + List<Node> offer(List<NodeCandidate> candidates) { List<Node> accepted = new ArrayList<>(); - for (NodeCandidate candidate : nodesPrioritized) { + for (NodeCandidate candidate : candidates) { if (candidate.allocation().isPresent()) { Allocation allocation = candidate.allocation().get(); ClusterMembership membership = allocation.membership(); @@ -121,7 +121,7 @@ class NodeAllocation { if ((! saturated() && hasCompatibleFlavor(candidate) && requestedNodes.acceptable(candidate)) || acceptToRetire) { candidate = candidate.withNode(); if (candidate.isValid()) - accepted.add(acceptNode(candidate, shouldRetire(candidate, nodesPrioritized), resizeable)); + accepted.add(acceptNode(candidate, shouldRetire(candidate, candidates), resizeable)); } } else if (! saturated() && hasCompatibleFlavor(candidate)) { diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeCandidate.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeCandidate.java index 4f0ae688b1c..62ac1f0d0e6 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeCandidate.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeCandidate.java @@ -238,7 +238,6 @@ public abstract class NodeCandidate implements Nodelike, Comparable<NodeCandidat private double skewWith(NodeResources resources) { if (parent.isEmpty()) return 0; - NodeResources free = freeParentCapacity.justNumbers().subtract(resources.justNumbers()); return Node.skew(parent.get().flavor().resources(), free); } diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeRepositoryProvisioner.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeRepositoryProvisioner.java index b35b0a5e301..7d15a2b30b1 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeRepositoryProvisioner.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/NodeRepositoryProvisioner.java @@ -84,8 +84,8 @@ public class NodeRepositoryProvisioner implements Provisioner { @Override public List<HostSpec> prepare(ApplicationId application, ClusterSpec cluster, Capacity requested, ProvisionLogger logger) { - log.log(Level.FINE, () -> "Received deploy prepare request for " + requested + - " for application " + application + ", cluster " + cluster); + log.log(Level.FINE, "Received deploy prepare request for " + requested + + " for application " + application + ", cluster " + cluster); if (cluster.group().isPresent()) throw new IllegalArgumentException("Node requests cannot specify a group"); @@ -96,21 +96,21 @@ public class NodeRepositoryProvisioner implements Provisioner { NodeResources resources; NodeSpec nodeSpec; if (requested.type() == NodeType.tenant) { - ClusterResources target = decideTargetResources(application, cluster, requested); - int nodeCount = capacityPolicies.decideSize(target.nodes(), - requested.isRequired(), - requested.canFail(), - application.instance().isTester(), - cluster); - groups = Math.min(target.groups(), nodeCount); // cannot have more groups than nodes - resources = capacityPolicies.decideNodeResources(target.nodeResources(), requested.isRequired(), cluster); - boolean exclusive = capacityPolicies.decideExclusivity(requested, cluster.isExclusive()); - nodeSpec = NodeSpec.from(nodeCount, resources, exclusive, requested.canFail()); - logIfDownscaled(target.nodes(), nodeCount, cluster, logger); + var actual = capacityPolicies.applyOn(requested, application); + ClusterResources target = decideTargetResources(application, cluster, actual); + boolean exclusive = capacityPolicies.decideExclusivity(actual, cluster.isExclusive()); + ensureRedundancy(target.nodes(), cluster, actual.canFail(), application); + logIfDownscaled(requested.minResources().nodes(), actual.minResources().nodes(), cluster, logger); + + groups = target.groups(); + resources = target.nodeResources().isUnspecified() ? capacityPolicies.defaultNodeResources(cluster.type()) + : target.nodeResources(); + nodeSpec = NodeSpec.from(target.nodes(), resources, exclusive, actual.canFail()); } else { groups = 1; // type request with multiple groups is not supported - resources = requested.minResources().nodeResources(); + resources = requested.minResources().nodeResources().isUnspecified() ? capacityPolicies.defaultNodeResources(cluster.type()) + : requested.minResources().nodeResources(); nodeSpec = NodeSpec.from(requested.type()); } return asSortedHosts(preparer.prepare(application, cluster, nodeSpec, groups), resources); @@ -164,12 +164,20 @@ public class NodeRepositoryProvisioner implements Provisioner { boolean firstDeployment = nodes.isEmpty(); AllocatableClusterResources currentResources = firstDeployment // start at min, preserve current resources otherwise - ? new AllocatableClusterResources(requested.minResources(), clusterSpec, nodeRepository) + ? new AllocatableClusterResources(initialResourcesFrom(requested, clusterSpec), clusterSpec, nodeRepository) : new AllocatableClusterResources(nodes.asList(), nodeRepository); var clusterModel = new ClusterModel(application, cluster, clusterSpec, nodes, nodeRepository.metricsDb(), nodeRepository.clock()); return within(Limits.of(requested), currentResources, firstDeployment, clusterModel); } + private ClusterResources initialResourcesFrom(Capacity requested, ClusterSpec clusterSpec) { + var initial = requested.minResources(); + if (initial.nodeResources().isUnspecified()) + initial = initial.with(capacityPolicies.defaultNodeResources(clusterSpec.type())); + return initial; + } + + /** Make the minimal adjustments needed to the current resources to stay within the limits */ private ClusterResources within(Limits limits, AllocatableClusterResources current, @@ -190,10 +198,28 @@ public class NodeRepositoryProvisioner implements Provisioner { .advertisedResources(); } - private void logIfDownscaled(int targetNodes, int actualNodes, ClusterSpec cluster, ProvisionLogger logger) { - if (zone.environment().isManuallyDeployed() && actualNodes < targetNodes) - logger.log(Level.INFO, "Requested " + targetNodes + " nodes for " + cluster + - ", downscaling to " + actualNodes + " nodes in " + zone.environment()); + /** + * Throw if the node count is 1 for container and content clusters and we're in a production zone + * + * @throws IllegalArgumentException if only one node is requested and we can fail + */ + private void ensureRedundancy(int nodeCount, ClusterSpec cluster, boolean canFail, ApplicationId application) { + if (! application.instance().isTester() && + canFail && + nodeCount == 1 && + requiresRedundancy(cluster.type()) && + zone.environment().isProduction()) + throw new IllegalArgumentException("Deployments to prod require at least 2 nodes per cluster for redundancy. Not fulfilled for " + cluster); + } + + private static boolean requiresRedundancy(ClusterSpec.Type clusterType) { + return clusterType.isContent() || clusterType.isContainer(); + } + + private void logIfDownscaled(int requestedMinNodes, int actualMinNodes, ClusterSpec cluster, ProvisionLogger logger) { + if (zone.environment().isManuallyDeployed() && actualMinNodes < requestedMinNodes) + logger.log(Level.INFO, "Requested " + requestedMinNodes + " nodes for " + cluster + + ", downscaling to " + actualMinNodes + " nodes in " + zone.environment()); } private List<HostSpec> asSortedHosts(List<Node> nodes, NodeResources requestedResources) { diff --git a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/Preparer.java b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/Preparer.java index 282b0d96cf4..b12368b2834 100644 --- a/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/Preparer.java +++ b/node-repository/src/main/java/com/yahoo/vespa/hosted/provision/provisioning/Preparer.java @@ -10,7 +10,6 @@ import com.yahoo.vespa.hosted.provision.Node; import com.yahoo.vespa.hosted.provision.NodeList; import com.yahoo.vespa.hosted.provision.NodeRepository; import com.yahoo.vespa.hosted.provision.NodesAndHosts; -import com.yahoo.vespa.hosted.provision.node.Nodes; import java.util.ArrayList; import java.util.List; @@ -25,13 +24,11 @@ import java.util.stream.Collectors; */ class Preparer { - private final NodeRepository nodeRepository; private final GroupPreparer groupPreparer; private final Optional<LoadBalancerProvisioner> loadBalancerProvisioner; public Preparer(NodeRepository nodeRepository, Optional<HostProvisioner> hostProvisioner, Optional<LoadBalancerProvisioner> loadBalancerProvisioner) { - this.nodeRepository = nodeRepository; this.loadBalancerProvisioner = loadBalancerProvisioner; this.groupPreparer = new GroupPreparer(nodeRepository, hostProvisioner); } @@ -69,9 +66,10 @@ class Preparer { for (int groupIndex = 0; groupIndex < wantedGroups; groupIndex++) { ClusterSpec clusterGroup = cluster.with(Optional.of(ClusterSpec.Group.from(groupIndex))); - GroupPreparer.PrepareResult result = groupPreparer.prepare( - application, clusterGroup, requestedNodes.fraction(wantedGroups), - surplusNodes, indices, wantedGroups, allNodesAndHosts); + GroupPreparer.PrepareResult result = groupPreparer.prepare(application, clusterGroup, + requestedNodes.fraction(wantedGroups), + surplusNodes, indices, wantedGroups, + allNodesAndHosts); allNodesAndHosts = result.allNodesAndHosts; // Might have changed List<Node> accepted = result.prepared; if (requestedNodes.rejectNonActiveParent()) { |