in streams/src/main/java/org/apache/kafka/streams/processor/internals/StreamPartitionAssignor.java [280:554]
public Map<String, Assignment> assign(Cluster metadata, Map<String, Subscription> subscriptions) {
// construct the client metadata from the decoded subscription info
Map<UUID, ClientMetadata> clientsMetadata = new HashMap<>();
for (Map.Entry<String, Subscription> entry : subscriptions.entrySet()) {
String consumerId = entry.getKey();
Subscription subscription = entry.getValue();
SubscriptionInfo info = SubscriptionInfo.decode(subscription.userData());
// create the new client metadata if necessary
ClientMetadata clientMetadata = clientsMetadata.get(info.processId);
if (clientMetadata == null) {
clientMetadata = new ClientMetadata(info.userEndPoint);
clientsMetadata.put(info.processId, clientMetadata);
}
// add the consumer to the client
clientMetadata.addConsumer(consumerId, info);
}
log.debug("Constructed client metadata {} from the member subscriptions.", clientsMetadata);
// ---------------- Step Zero ---------------- //
// parse the topology to determine the repartition source topics,
// making sure they are created with the number of partitions as
// the maximum of the depending sub-topologies source topics' number of partitions
final Map<Integer, InternalTopologyBuilder.TopicsInfo> topicGroups = taskManager.builder().topicGroups();
final Map<String, InternalTopicMetadata> repartitionTopicMetadata = new HashMap<>();
for (InternalTopologyBuilder.TopicsInfo topicsInfo : topicGroups.values()) {
for (InternalTopicConfig topic: topicsInfo.repartitionSourceTopics.values()) {
repartitionTopicMetadata.put(topic.name(), new InternalTopicMetadata(topic));
}
}
boolean numPartitionsNeeded;
do {
numPartitionsNeeded = false;
for (InternalTopologyBuilder.TopicsInfo topicsInfo : topicGroups.values()) {
for (String topicName : topicsInfo.repartitionSourceTopics.keySet()) {
int numPartitions = repartitionTopicMetadata.get(topicName).numPartitions;
// try set the number of partitions for this repartition topic if it is not set yet
if (numPartitions == UNKNOWN) {
for (InternalTopologyBuilder.TopicsInfo otherTopicsInfo : topicGroups.values()) {
final Set<String> otherSinkTopics = otherTopicsInfo.sinkTopics;
if (otherSinkTopics.contains(topicName)) {
// if this topic is one of the sink topics of this topology,
// use the maximum of all its source topic partitions as the number of partitions
for (final String sourceTopicName : otherTopicsInfo.sourceTopics) {
final Integer numPartitionsCandidate;
// It is possible the sourceTopic is another internal topic, i.e,
// map().join().join(map())
if (repartitionTopicMetadata.containsKey(sourceTopicName)) {
numPartitionsCandidate = repartitionTopicMetadata.get(sourceTopicName).numPartitions;
} else {
numPartitionsCandidate = metadata.partitionCountForTopic(sourceTopicName);
if (numPartitionsCandidate == null) {
repartitionTopicMetadata.get(topicName).numPartitions = NOT_AVAILABLE;
}
}
if (numPartitionsCandidate != null && numPartitionsCandidate > numPartitions) {
numPartitions = numPartitionsCandidate;
}
}
}
}
// if we still have not find the right number of partitions,
// another iteration is needed
if (numPartitions == UNKNOWN) {
numPartitionsNeeded = true;
} else {
repartitionTopicMetadata.get(topicName).numPartitions = numPartitions;
}
}
}
}
} while (numPartitionsNeeded);
// ensure the co-partitioning topics within the group have the same number of partitions,
// and enforce the number of partitions for those repartition topics to be the same if they
// are co-partitioned as well.
ensureCopartitioning(taskManager.builder().copartitionGroups(), repartitionTopicMetadata, metadata);
// make sure the repartition source topics exist with the right number of partitions,
// create these topics if necessary
prepareTopic(repartitionTopicMetadata);
// augment the metadata with the newly computed number of partitions for all the
// repartition source topics
final Map<TopicPartition, PartitionInfo> allRepartitionTopicPartitions = new HashMap<>();
for (Map.Entry<String, InternalTopicMetadata> entry : repartitionTopicMetadata.entrySet()) {
final String topic = entry.getKey();
final Integer numPartitions = entry.getValue().numPartitions;
for (int partition = 0; partition < numPartitions; partition++) {
allRepartitionTopicPartitions.put(new TopicPartition(topic, partition),
new PartitionInfo(topic, partition, null, new Node[0], new Node[0]));
}
}
final Cluster fullMetadata = metadata.withPartitions(allRepartitionTopicPartitions);
taskManager.setClusterMetadata(fullMetadata);
log.debug("Created repartition topics {} from the parsed topology.", allRepartitionTopicPartitions.values());
// ---------------- Step One ---------------- //
// get the tasks as partition groups from the partition grouper
final Set<String> allSourceTopics = new HashSet<>();
final Map<Integer, Set<String>> sourceTopicsByGroup = new HashMap<>();
for (Map.Entry<Integer, InternalTopologyBuilder.TopicsInfo> entry : topicGroups.entrySet()) {
allSourceTopics.addAll(entry.getValue().sourceTopics);
sourceTopicsByGroup.put(entry.getKey(), entry.getValue().sourceTopics);
}
final Map<TaskId, Set<TopicPartition>> partitionsForTask = partitionGrouper.partitionGroups(sourceTopicsByGroup, fullMetadata);
// check if all partitions are assigned, and there are no duplicates of partitions in multiple tasks
final Set<TopicPartition> allAssignedPartitions = new HashSet<>();
final Map<Integer, Set<TaskId>> tasksByTopicGroup = new HashMap<>();
for (Map.Entry<TaskId, Set<TopicPartition>> entry : partitionsForTask.entrySet()) {
final Set<TopicPartition> partitions = entry.getValue();
for (TopicPartition partition : partitions) {
if (allAssignedPartitions.contains(partition)) {
log.warn("Partition {} is assigned to more than one tasks: {}", partition, partitionsForTask);
}
}
allAssignedPartitions.addAll(partitions);
final TaskId id = entry.getKey();
Set<TaskId> ids = tasksByTopicGroup.get(id.topicGroupId);
if (ids == null) {
ids = new HashSet<>();
tasksByTopicGroup.put(id.topicGroupId, ids);
}
ids.add(id);
}
for (String topic : allSourceTopics) {
final List<PartitionInfo> partitionInfoList = fullMetadata.partitionsForTopic(topic);
if (!partitionInfoList.isEmpty()) {
for (PartitionInfo partitionInfo : partitionInfoList) {
final TopicPartition partition = new TopicPartition(partitionInfo.topic(), partitionInfo.partition());
if (!allAssignedPartitions.contains(partition)) {
log.warn("Partition {} is not assigned to any tasks: {}", partition, partitionsForTask);
}
}
} else {
log.warn("No partitions found for topic {}", topic);
}
}
// add tasks to state change log topic subscribers
final Map<String, InternalTopicMetadata> changelogTopicMetadata = new HashMap<>();
for (Map.Entry<Integer, InternalTopologyBuilder.TopicsInfo> entry : topicGroups.entrySet()) {
final int topicGroupId = entry.getKey();
final Map<String, InternalTopicConfig> stateChangelogTopics = entry.getValue().stateChangelogTopics;
for (InternalTopicConfig topicConfig : stateChangelogTopics.values()) {
// the expected number of partitions is the max value of TaskId.partition + 1
int numPartitions = UNKNOWN;
if (tasksByTopicGroup.get(topicGroupId) != null) {
for (TaskId task : tasksByTopicGroup.get(topicGroupId)) {
if (numPartitions < task.partition + 1)
numPartitions = task.partition + 1;
}
final InternalTopicMetadata topicMetadata = new InternalTopicMetadata(topicConfig);
topicMetadata.numPartitions = numPartitions;
changelogTopicMetadata.put(topicConfig.name(), topicMetadata);
} else {
log.debug("No tasks found for topic group {}", topicGroupId);
}
}
}
prepareTopic(changelogTopicMetadata);
log.debug("Created state changelog topics {} from the parsed topology.", changelogTopicMetadata.values());
// ---------------- Step Two ---------------- //
// assign tasks to clients
final Map<UUID, ClientState> states = new HashMap<>();
for (Map.Entry<UUID, ClientMetadata> entry : clientsMetadata.entrySet()) {
states.put(entry.getKey(), entry.getValue().state);
}
log.debug("Assigning tasks {} to clients {} with number of replicas {}",
partitionsForTask.keySet(), states, numStandbyReplicas);
final StickyTaskAssignor<UUID> taskAssignor = new StickyTaskAssignor<>(states, partitionsForTask.keySet());
taskAssignor.assign(numStandbyReplicas);
log.info("Assigned tasks to clients as {}.", states);
// ---------------- Step Three ---------------- //
// construct the global partition assignment per host map
final Map<HostInfo, Set<TopicPartition>> partitionsByHostState = new HashMap<>();
for (Map.Entry<UUID, ClientMetadata> entry : clientsMetadata.entrySet()) {
final HostInfo hostInfo = entry.getValue().hostInfo;
if (hostInfo != null) {
final Set<TopicPartition> topicPartitions = new HashSet<>();
final ClientState state = entry.getValue().state;
for (final TaskId id : state.activeTasks()) {
topicPartitions.addAll(partitionsForTask.get(id));
}
partitionsByHostState.put(hostInfo, topicPartitions);
}
}
taskManager.setPartitionsByHostState(partitionsByHostState);
// within the client, distribute tasks to its owned consumers
final Map<String, Assignment> assignment = new HashMap<>();
for (Map.Entry<UUID, ClientMetadata> entry : clientsMetadata.entrySet()) {
final Set<String> consumers = entry.getValue().consumers;
final ClientState state = entry.getValue().state;
final List<List<TaskId>> interleavedActive = interleaveTasksByGroupId(state.activeTasks(), consumers.size());
final List<List<TaskId>> interleavedStandby = interleaveTasksByGroupId(state.standbyTasks(), consumers.size());
int consumerTaskIndex = 0;
for (String consumer : consumers) {
final Map<TaskId, Set<TopicPartition>> standby = new HashMap<>();
final ArrayList<AssignedPartition> assignedPartitions = new ArrayList<>();
final List<TaskId> assignedActiveList = interleavedActive.get(consumerTaskIndex);
for (final TaskId taskId : assignedActiveList) {
for (final TopicPartition partition : partitionsForTask.get(taskId)) {
assignedPartitions.add(new AssignedPartition(taskId, partition));
}
}
if (!state.standbyTasks().isEmpty()) {
final List<TaskId> assignedStandbyList = interleavedStandby.get(consumerTaskIndex);
for (final TaskId taskId : assignedStandbyList) {
Set<TopicPartition> standbyPartitions = standby.get(taskId);
if (standbyPartitions == null) {
standbyPartitions = new HashSet<>();
standby.put(taskId, standbyPartitions);
}
standbyPartitions.addAll(partitionsForTask.get(taskId));
}
}
consumerTaskIndex++;
Collections.sort(assignedPartitions);
final List<TaskId> active = new ArrayList<>();
final List<TopicPartition> activePartitions = new ArrayList<>();
for (AssignedPartition partition : assignedPartitions) {
active.add(partition.taskId);
activePartitions.add(partition.partition);
}
// finally, encode the assignment before sending back to coordinator
assignment.put(consumer, new Assignment(activePartitions, new AssignmentInfo(active, standby, partitionsByHostState).encode()));
}
}
return assignment;
}