lkokhreidze commented on a change in pull request #10851: URL: https://github.com/apache/kafka/pull/10851#discussion_r810019916
########## File path: streams/src/main/java/org/apache/kafka/streams/processor/internals/assignment/ClientTagAwareStandbyTaskAssignor.java ########## @@ -0,0 +1,295 @@ +/* + * 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.kafka.streams.processor.internals.assignment; + +import org.apache.kafka.streams.processor.TaskId; +import org.apache.kafka.streams.processor.internals.assignment.AssignorConfiguration.AssignmentConfigs; +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import java.util.HashMap; +import java.util.HashSet; +import java.util.Map; +import java.util.Map.Entry; +import java.util.Objects; +import java.util.Set; +import java.util.UUID; + +import static org.apache.kafka.streams.processor.internals.assignment.StandbyTaskAssignmentUtils.computeTasksToRemainingStandbys; +import static org.apache.kafka.streams.processor.internals.assignment.StandbyTaskAssignmentUtils.pollClientAndMaybeAssignRemainingStandbyTasks; + +/** + * Distributes standby tasks over different tag dimensions. Standby task distribution is on a best-effort basis. + * If rack aware standby task assignment is not possible, implementation fall backs to distributing standby tasks on least-loaded clients. + * + * @see DefaultStandbyTaskAssignor + */ +class ClientTagAwareStandbyTaskAssignor implements StandbyTaskAssignor { + private static final Logger log = LoggerFactory.getLogger(ClientTagAwareStandbyTaskAssignor.class); + + /** + * The algorithm distributes standby tasks for the {@param statefulTaskIds} over different tag dimensions. + * For each stateful task, the number of standby tasks will be assigned based on configured {@link AssignmentConfigs#numStandbyReplicas}. + * Rack aware standby tasks distribution only takes into account tags specified via {@link AssignmentConfigs#rackAwareAssignmentTags}. + * Ideally, all standby tasks for any given stateful task will be located on different tag dimensions to have the best possible distribution. + * However, if the ideal distribution is impossible, the algorithm will fall back to the least-loaded clients without considering rack awareness constraints into consideration. + * The least-loaded clients are determined based on the total number of tasks (active and standby tasks) assigned to the client. + */ + @Override + public boolean assign(final Map<UUID, ClientState> clients, + final Set<TaskId> allTaskIds, + final Set<TaskId> statefulTaskIds, + final AssignorConfiguration.AssignmentConfigs configs) { + final int numStandbyReplicas = configs.numStandbyReplicas; + final Set<String> rackAwareAssignmentTags = new HashSet<>(configs.rackAwareAssignmentTags); + + final Map<TaskId, Integer> tasksToRemainingStandbys = computeTasksToRemainingStandbys( + numStandbyReplicas, + allTaskIds + ); + + final Map<String, Set<String>> tagKeyToValues = new HashMap<>(); + final Map<TagEntry, Set<UUID>> tagEntryToClients = new HashMap<>(); + + fillClientsTagStatistics(clients, tagEntryToClients, tagKeyToValues); + + final ConstrainedPrioritySet standbyTaskClientsByTaskLoad = new ConstrainedPrioritySet( + (client, t) -> !clients.get(client).hasAssignedTask(t), + client -> clients.get(client).assignedTaskLoad() + ); + + final Map<TaskId, UUID> pendingStandbyTasksToClientId = new HashMap<>(); + + for (final TaskId statefulTaskId : statefulTaskIds) { + for (final Map.Entry<UUID, ClientState> entry : clients.entrySet()) { + final UUID clientId = entry.getKey(); + final ClientState clientState = entry.getValue(); + + if (clientState.activeTasks().contains(statefulTaskId)) { + final int numberOfRemainingStandbys = assignStandbyTasksToClientsWithDifferentTags( + standbyTaskClientsByTaskLoad, + statefulTaskId, + clientId, + rackAwareAssignmentTags, + clients, + tasksToRemainingStandbys, + tagKeyToValues, + tagEntryToClients + ); + + if (numberOfRemainingStandbys > 0) { + pendingStandbyTasksToClientId.put(statefulTaskId, clientId); + } else { + tasksToRemainingStandbys.remove(statefulTaskId); + } + } + } + } + + if (!tasksToRemainingStandbys.isEmpty()) { + log.debug("Rack aware standby task assignment was not able to assign all standby tasks. " + + "tasksToRemainingStandbys=[{}], pendingStandbyTasksToClientId=[{}]. " + + "Will distribute the remaining standby tasks to least loaded clients.", + tasksToRemainingStandbys, pendingStandbyTasksToClientId); + + assignPendingStandbyTasksToLeastLoadedClients(clients, + numStandbyReplicas, + rackAwareAssignmentTags, + standbyTaskClientsByTaskLoad, + tasksToRemainingStandbys, + pendingStandbyTasksToClientId); + } + + // returning false, because standby task assignment will never require a follow-up probing rebalance. + return false; + } + + private static void assignPendingStandbyTasksToLeastLoadedClients(final Map<UUID, ClientState> clients, + final int numStandbyReplicas, + final Set<String> rackAwareAssignmentTags, + final ConstrainedPrioritySet standbyTaskClientsByTaskLoad, + final Map<TaskId, Integer> pendingStandbyTaskToNumberRemainingStandbys, + final Map<TaskId, UUID> pendingStandbyTaskToClientId) { + // We need to re offer all the clients to find the least loaded ones + standbyTaskClientsByTaskLoad.offerAll(clients.keySet()); + + for (final Entry<TaskId, Integer> pendingStandbyTaskAssignmentEntry : pendingStandbyTaskToNumberRemainingStandbys.entrySet()) { + final TaskId activeTaskId = pendingStandbyTaskAssignmentEntry.getKey(); + final UUID clientId = pendingStandbyTaskToClientId.get(activeTaskId); + + final int numberOfRemainingStandbys = pollClientAndMaybeAssignRemainingStandbyTasks( + clients, + pendingStandbyTaskToNumberRemainingStandbys, + standbyTaskClientsByTaskLoad, + activeTaskId + ); + + if (numberOfRemainingStandbys > 0) { + log.warn("Unable to assign {} of {} standby tasks for task [{}] with client tags [{}]. " + + "There is not enough available capacity. You should " + + "increase the number of application instances " + + "on different client tag dimensions " + + "to maintain the requested number of standby replicas. " + + "Rack awareness is configured with [{}] tags.", + numberOfRemainingStandbys, numStandbyReplicas, activeTaskId, + clients.get(clientId).clientTags(), rackAwareAssignmentTags); + } + } + } + + @Override + public boolean isAllowedTaskMovement(final ClientState source, final ClientState destination) { + final Map<String, String> sourceClientTags = source.clientTags(); + final Map<String, String> destinationClientTags = destination.clientTags(); + + for (final Entry<String, String> sourceClientTagEntry : sourceClientTags.entrySet()) { + if (!sourceClientTagEntry.getValue().equals(destinationClientTags.get(sourceClientTagEntry.getKey()))) { + return false; + } + } + + return true; + } + + private static void fillClientsTagStatistics(final Map<UUID, ClientState> clientStates, + final Map<TagEntry, Set<UUID>> tagEntryToClients, + final Map<String, Set<String>> tagKeyToValues) { + for (final Entry<UUID, ClientState> clientStateEntry : clientStates.entrySet()) { + final UUID clientId = clientStateEntry.getKey(); + final ClientState clientState = clientStateEntry.getValue(); + + clientState.clientTags().forEach((tagKey, tagValue) -> { + tagKeyToValues.computeIfAbsent(tagKey, ignored -> new HashSet<>()).add(tagValue); + tagEntryToClients.computeIfAbsent(new TagEntry(tagKey, tagValue), ignored -> new HashSet<>()).add(clientId); + }); + } + } + + private static int assignStandbyTasksToClientsWithDifferentTags(final ConstrainedPrioritySet standbyTaskClientsByTaskLoad, + final TaskId activeTaskId, + final UUID activeTaskClient, + final Set<String> rackAwareAssignmentTags, + final Map<UUID, ClientState> clientStates, + final Map<TaskId, Integer> tasksToRemainingStandbys, + final Map<String, Set<String>> tagKeyToValues, + final Map<TagEntry, Set<UUID>> tagEntryToClients) { + standbyTaskClientsByTaskLoad.offerAll(clientStates.keySet()); + + // We set countOfUsedClients as 1 because client where active task is located has to be considered as used. + int countOfUsedClients = 1; + int numRemainingStandbys = tasksToRemainingStandbys.get(activeTaskId); + + final Set<UUID> clientsOnAlreadyUsedTagDimensions = new HashSet<>( + findClientsOnUsedClientTagDimensions( + activeTaskClient, + countOfUsedClients, + rackAwareAssignmentTags, + clientStates, + tagEntryToClients, + tagKeyToValues + ) + ); + + while (numRemainingStandbys > 0) { + final UUID clientOnUnusedTagDimensions = standbyTaskClientsByTaskLoad.poll( + activeTaskId, uuid -> !clientsOnAlreadyUsedTagDimensions.contains(uuid) + ); + + if (clientOnUnusedTagDimensions == null) { + break; + } + + clientStates.get(clientOnUnusedTagDimensions).assignStandby(activeTaskId); + + countOfUsedClients++; + + clientsOnAlreadyUsedTagDimensions.addAll( + findClientsOnUsedClientTagDimensions( + clientOnUnusedTagDimensions, + countOfUsedClients, + rackAwareAssignmentTags, + clientStates, + tagEntryToClients, + tagKeyToValues + ) + ); + + numRemainingStandbys--; + } + + return numRemainingStandbys; + } + + private static Set<UUID> findClientsOnUsedClientTagDimensions(final UUID usedClient, + final int countOfUsedClients, + final Set<String> rackAwareAssignmentTags, + final Map<UUID, ClientState> clientStates, + final Map<TagEntry, Set<UUID>> tagEntryToClients, + final Map<String, Set<String>> tagKeyToValues) { + final Set<UUID> filteredClients = new HashSet<>(); + + final Map<String, String> usedClientTags = clientStates.get(usedClient).clientTags(); + + for (final Entry<String, String> usedClientTagEntry : usedClientTags.entrySet()) { + final String tagKey = usedClientTagEntry.getKey(); + + if (!rackAwareAssignmentTags.contains(tagKey)) { + log.warn("Client tag with key [{}] will be ignored when computing rack aware standby " + + "task assignment because it is not part of the configured rack awareness [{}].", + tagKey, rackAwareAssignmentTags); + continue; + } + + final Set<String> allTagValues = tagKeyToValues.get(tagKey); + final String tagValue = usedClientTagEntry.getValue(); + + // If we have used more clients than all the tag's unique values, + // we can't filter out clients located on that tag. + if (allTagValues.size() <= countOfUsedClients) { + continue; Review comment: Consider the following example ``` # Kafka Streams Client 1 client.tag.zone: eu-central-1a client.tag.cluster: k8s-cluster1 rack.aware.assignment.tags: zone,cluster # Kafka Streams Client 2 client.tag.zone: eu-central-1b client.tag.cluster: k8s-cluster1 rack.aware.assignment.tags: zone,cluster # Kafka Streams Client 3 client.tag.zone: eu-central-1c client.tag.cluster: k8s-cluster1 rack.aware.assignment.tags: zone,cluster # Kafka Streams Client 4 client.tag.zone: eu-central-1a client.tag.cluster: k8s-cluster2 rack.aware.assignment.tags: zone,cluster # Kafka Streams Client 5 client.tag.zone: eu-central-1b client.tag.cluster: k8s-cluster2 rack.aware.assignment.tags: zone,cluster # Kafka Streams Client 6 client.tag.zone: eu-central-1c client.tag.cluster: k8s-cluster2 rack.aware.assignment.tags: zone,cluster ``` With the above we have following number of unique tag values: ``` zone: 3 (eu-central-1a, eu-central-1b, eu-central-1c) cluster: 2 (k8s-cluster1, k8s-cluster2) ``` Now lets say we have standby replica count as `2` and we want to active task is located on `Client 1` `usedClients=1` (because of the active task) ### 1s standby assignment During 1st standby takes assignment, we will exclude clients on following dimensions: ``` zone: eu-central-1a cluster: k8s-cluster1 ``` Used clients will get incremented since we can allocate the client on different `zone` and `cluster`. `usedClients=2` ### 2nd standby assignment We will have to exclude `zone: eu-central-1a and (eu-central-1b || eu-central-1c)` tag values. We can do that, because after we exclude clients on the new tag, we still have clients on the one free tag value we can assign the next standby to. We can't exclude `cluster` because we have already used two clients, and we just have two unique values for the `cluster` tag, so it's impossible to get the ideal distribution with this configuration and number of instances. So we can only achieve partial distribution. So idea of this check is to ignore tags where we have less unique values than the clients we have already used. If we don't have this check, for the 2nd standby task assignment we would have excluded all the clients located on `k8s-cluster1` and `k8s-cluster2`, and there wouldn't be any client left to assign the standby task to. We would fall back to the least loaded client, BUT there will be no guarantee that least loaded client assignment would honor partial rack awareness. Hope this makes sense. -- 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: jira-unsubscr...@kafka.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
