dajac commented on code in PR #13443: URL: https://github.com/apache/kafka/pull/13443#discussion_r1188857767
########## group-coordinator/src/main/java/org/apache/kafka/coordinator/group/assignor/RangeAssignor.java: ########## @@ -0,0 +1,238 @@ +/* + * 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.coordinator.group.assignor; + +import org.apache.kafka.common.Uuid; +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import java.util.ArrayList; +import java.util.Collection; +import java.util.Collections; +import java.util.HashMap; +import java.util.HashSet; +import java.util.List; +import java.util.Map; +import java.util.Set; + +import static java.lang.Math.min; + +/** + * This Range Assignor inherits properties of both the range assignor and the sticky assignor. + * The properties are as follows: + * <ol> + * <li> Each member must get at least one partition from every topic that it is subscribed to. The only exception is when + * the number of subscribed members is greater than the number of partitions for that topic. (Range) </li> + * <li> Partitions should be assigned to members in a way that facilitates the join operation when required. (Range) + * This can only be done if every member is subscribed to the same topics and the topics are co-partitioned. + * Two streams are co-partitioned if the following conditions are met: + * <ul> + * <li> The keys must have the same schemas. </li> + * <li> The topics involved must have the same number of partitions. </li> + * </ul> + * </li> + * <li> Members should retain as much of their previous assignment as possible to reduce the number of partition movements during reassignment. (Sticky) </li> + * </ol> + */ +public class RangeAssignor implements PartitionAssignor { + private static final Logger log = LoggerFactory.getLogger(RangeAssignor.class); + + public static final String RANGE_ASSIGNOR_NAME = "range"; + + @Override + public String name() { + return RANGE_ASSIGNOR_NAME; + } + + /** + * Pair of memberId and remaining partitions to meet the quota. + */ + private static class MemberWithRemainingAssignments { + /** + * Member Id. + */ + private final String memberId; + /** + * Number of partitions required to meet the assignment quota. + */ + private final Integer remaining; + + public MemberWithRemainingAssignments(String memberId, Integer remaining) { + this.memberId = memberId; + this.remaining = remaining; + } + } + + /** + * @return Map of topicIds to a list of members subscribed to them. + */ + private Map<Uuid, List<String>> membersPerTopic(final AssignmentSpec assignmentSpec) { + Map<Uuid, List<String>> membersPerTopic = new HashMap<>(); + Map<String, AssignmentMemberSpec> membersData = assignmentSpec.members(); + + membersData.forEach((memberId, memberMetadata) -> { + Collection<Uuid> topics = memberMetadata.subscribedTopicIds(); + for (Uuid topicId: topics) { + // Only topics that are present in both the subscribed topics list and the topic metadata should be considered for assignment. + if (assignmentSpec.topics().containsKey(topicId)) { + membersPerTopic + .computeIfAbsent(topicId, k -> new ArrayList<>()) + .add(memberId); + } else { + log.warn("Member " + memberId + " subscribed to topic " + topicId + " which doesn't exist in the topic metadata"); + } + } + }); + + return membersPerTopic; + } + + /** + * <p> The algorithm includes the following steps: + * <ol> + * <li> Generate a map of <code>membersPerTopic</code> using the given member subscriptions.</li> + * <li> Generate a list of members (<code>potentiallyUnfilledMembers</code>) that have not met the minimum required quota of partitions for the assignment AND + * get a list (<code>assignedStickyPartitionsPerTopic</code>) of partitions that will be retained in the new assignment.</li> + * <li> Add members from the <code>potentiallyUnfilledMembers</code> list to the <code>unfilledMembersPerTopic</code> map if they haven't met the total required quota + * i.e. minRequiredQuota + 1, if the member is designated to receive one of the excess partitions OR minRequiredQuota otherwise. </li> + * <li> Generate a list of unassigned partitions by calculating the difference between the total partitions for the topic and the assigned (sticky) partitions. </li> + * <li> Check if unfilled members exist for the current topicId and assign partitions to them in ranges from the <code>unassignedPartitionsPerTopic</code> map + * based on the remaining partitions value stored. </li> + * </ol> + * </p> + */ + @Override + public GroupAssignment assign(final AssignmentSpec assignmentSpec) throws PartitionAssignorException { + Map<String, MemberAssignment> newAssignment = new HashMap<>(); + + // Step 1 + Map<Uuid, List<String>> membersPerTopic = membersPerTopic(assignmentSpec); + + membersPerTopic.forEach((topicId, membersForTopic) -> { + + AssignmentTopicMetadata topicData = assignmentSpec.topics().get(topicId); + int numPartitionsForTopic = topicData.numPartitions(); + int minRequiredQuota = numPartitionsForTopic / membersForTopic.size(); + // Each member can get only ONE extra partition per topic after receiving the minimum quota. + int numMembersWithExtraPartition = numPartitionsForTopic % membersForTopic.size(); + + // Idle members case : When the number of members subscribed to a topic is greater than the total number of partitions, + // all members get assigned via the "extra partitions" logic since minRequiredQuota = 0. + + // Step 2 + Set<Integer> assignedStickyPartitionsForTopic = new HashSet<>(); + List<MemberWithRemainingAssignments> potentiallyUnfilledMembers = new ArrayList<>(); + + for (String memberId : membersForTopic) { + Set<Integer> assignedPartitionsForTopic = assignmentSpec.members().get(memberId) + .assignedPartitions().getOrDefault(topicId, Collections.emptySet()); + + int currentAssignmentSize = assignedPartitionsForTopic.size(); + List<Integer> currentAssignmentListForTopic = new ArrayList<>(assignedPartitionsForTopic); + + // If there were partitions from this topic that were previously assigned to this member present, retain as much as possible. + // We need to retain: + // 1) currentSize number of partitions (when currentSize < required) OR + // 2) required number of partitions (when currentSize > required). + if (currentAssignmentSize > 0) { + int retainedPartitionsCount = min(currentAssignmentSize, minRequiredQuota); + // The partitions need to be in ascending order since we want the same partition numbers from each topic + // to go to the same member to facilitate joins in case of co-partitioned topics. + Collections.sort(currentAssignmentListForTopic); + for (int i = 0; i < retainedPartitionsCount; i++) { + assignedStickyPartitionsForTopic + .add(currentAssignmentListForTopic.get(i)); + newAssignment.computeIfAbsent(memberId, k -> new MemberAssignment(new HashMap<>())) + .targetPartitions() + .computeIfAbsent(topicId, k -> new HashSet<>()) + .add(currentAssignmentListForTopic.get(i)); + } + } + + // Number of partitions required to meet the minRequiredQuota. + // There are 3 cases w.r.t the value of remaining: + // 1) remaining < 0: this means that the member has more than the min required amount. + // 2) If remaining = 0: member has the minimum required partitions, but it may get an extra partition, so it is a potentially unfilled member. + // 3) If remaining > 0: member doesn't have the minimum required partitions, so it should be added to potentiallyUnfilledMembers. + int remaining = minRequiredQuota - currentAssignmentSize; + + // In order to remain as sticky as possible, since the order of members can be different, we want the members that have more than the required number of partitions, + // to get the extra partition instead of assigning the extras to the first few members directly. + if (remaining < 0 && numMembersWithExtraPartition > 0) { + numMembersWithExtraPartition--; + // Since we already added the minimumRequiredQuota of partitions in the previous step (until minReq - 1), we just need to + // add the extra partition that will be present at the index right after min quota was satisfied. + assignedStickyPartitionsForTopic + .add(currentAssignmentListForTopic.get(minRequiredQuota)); + newAssignment.computeIfAbsent(memberId, k -> new MemberAssignment(new HashMap<>())) + .targetPartitions() + .computeIfAbsent(topicId, k -> new HashSet<>()) + .add(currentAssignmentListForTopic.get(minRequiredQuota)); + } else { + MemberWithRemainingAssignments newPair = new MemberWithRemainingAssignments(memberId, remaining); + potentiallyUnfilledMembers.add(newPair); + } + } + + List<MemberWithRemainingAssignments> unfilledMembersForTopic = new ArrayList<>(); + + // Step 3 + // If remaining > 0 after increasing the required quota due to the extra partition, add potentially unfilled member to the unfilled members list. + for (MemberWithRemainingAssignments pair : potentiallyUnfilledMembers) { + String memberId = pair.memberId; + Integer remaining = pair.remaining; + if (numMembersWithExtraPartition > 0) { + remaining++; + numMembersWithExtraPartition--; + } + if (remaining > 0) { + MemberWithRemainingAssignments newPair = new MemberWithRemainingAssignments(memberId, remaining); + unfilledMembersForTopic + .add(newPair); + } + } + + // Step 4 + // Find the difference between the total partitions per topic and the already assigned sticky partitions for the topic to get unassigned partitions. + // List of unassigned partitions per topic contains the partitions in ascending order. Review Comment: Interesting... I feel like this part is more important than all the rest, no? ########## group-coordinator/src/main/java/org/apache/kafka/coordinator/group/assignor/RangeAssignor.java: ########## @@ -0,0 +1,238 @@ +/* + * 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.coordinator.group.assignor; + +import org.apache.kafka.common.Uuid; +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import java.util.ArrayList; +import java.util.Collection; +import java.util.Collections; +import java.util.HashMap; +import java.util.HashSet; +import java.util.List; +import java.util.Map; +import java.util.Set; + +import static java.lang.Math.min; + +/** + * This Range Assignor inherits properties of both the range assignor and the sticky assignor. + * The properties are as follows: + * <ol> + * <li> Each member must get at least one partition from every topic that it is subscribed to. The only exception is when + * the number of subscribed members is greater than the number of partitions for that topic. (Range) </li> + * <li> Partitions should be assigned to members in a way that facilitates the join operation when required. (Range) + * This can only be done if every member is subscribed to the same topics and the topics are co-partitioned. + * Two streams are co-partitioned if the following conditions are met: + * <ul> + * <li> The keys must have the same schemas. </li> + * <li> The topics involved must have the same number of partitions. </li> + * </ul> + * </li> + * <li> Members should retain as much of their previous assignment as possible to reduce the number of partition movements during reassignment. (Sticky) </li> + * </ol> + */ +public class RangeAssignor implements PartitionAssignor { + private static final Logger log = LoggerFactory.getLogger(RangeAssignor.class); + + public static final String RANGE_ASSIGNOR_NAME = "range"; + + @Override + public String name() { + return RANGE_ASSIGNOR_NAME; + } + + /** + * Pair of memberId and remaining partitions to meet the quota. + */ + private static class MemberWithRemainingAssignments { + /** + * Member Id. + */ + private final String memberId; + /** + * Number of partitions required to meet the assignment quota. + */ + private final Integer remaining; + + public MemberWithRemainingAssignments(String memberId, Integer remaining) { + this.memberId = memberId; + this.remaining = remaining; + } + } + + /** + * @return Map of topicIds to a list of members subscribed to them. + */ + private Map<Uuid, List<String>> membersPerTopic(final AssignmentSpec assignmentSpec) { + Map<Uuid, List<String>> membersPerTopic = new HashMap<>(); + Map<String, AssignmentMemberSpec> membersData = assignmentSpec.members(); + + membersData.forEach((memberId, memberMetadata) -> { + Collection<Uuid> topics = memberMetadata.subscribedTopicIds(); + for (Uuid topicId: topics) { + // Only topics that are present in both the subscribed topics list and the topic metadata should be considered for assignment. + if (assignmentSpec.topics().containsKey(topicId)) { + membersPerTopic + .computeIfAbsent(topicId, k -> new ArrayList<>()) + .add(memberId); + } else { + log.warn("Member " + memberId + " subscribed to topic " + topicId + " which doesn't exist in the topic metadata"); + } + } + }); + + return membersPerTopic; + } + + /** + * <p> The algorithm includes the following steps: + * <ol> + * <li> Generate a map of <code>membersPerTopic</code> using the given member subscriptions.</li> + * <li> Generate a list of members (<code>potentiallyUnfilledMembers</code>) that have not met the minimum required quota of partitions for the assignment AND + * get a list (<code>assignedStickyPartitionsPerTopic</code>) of partitions that will be retained in the new assignment.</li> + * <li> Add members from the <code>potentiallyUnfilledMembers</code> list to the <code>unfilledMembersPerTopic</code> map if they haven't met the total required quota + * i.e. minRequiredQuota + 1, if the member is designated to receive one of the excess partitions OR minRequiredQuota otherwise. </li> + * <li> Generate a list of unassigned partitions by calculating the difference between the total partitions for the topic and the assigned (sticky) partitions. </li> + * <li> Check if unfilled members exist for the current topicId and assign partitions to them in ranges from the <code>unassignedPartitionsPerTopic</code> map + * based on the remaining partitions value stored. </li> + * </ol> + * </p> + */ + @Override + public GroupAssignment assign(final AssignmentSpec assignmentSpec) throws PartitionAssignorException { + Map<String, MemberAssignment> newAssignment = new HashMap<>(); + + // Step 1 + Map<Uuid, List<String>> membersPerTopic = membersPerTopic(assignmentSpec); + + membersPerTopic.forEach((topicId, membersForTopic) -> { + + AssignmentTopicMetadata topicData = assignmentSpec.topics().get(topicId); + int numPartitionsForTopic = topicData.numPartitions(); + int minRequiredQuota = numPartitionsForTopic / membersForTopic.size(); + // Each member can get only ONE extra partition per topic after receiving the minimum quota. + int numMembersWithExtraPartition = numPartitionsForTopic % membersForTopic.size(); + + // Idle members case : When the number of members subscribed to a topic is greater than the total number of partitions, + // all members get assigned via the "extra partitions" logic since minRequiredQuota = 0. + + // Step 2 + Set<Integer> assignedStickyPartitionsForTopic = new HashSet<>(); + List<MemberWithRemainingAssignments> potentiallyUnfilledMembers = new ArrayList<>(); + + for (String memberId : membersForTopic) { + Set<Integer> assignedPartitionsForTopic = assignmentSpec.members().get(memberId) + .assignedPartitions().getOrDefault(topicId, Collections.emptySet()); + + int currentAssignmentSize = assignedPartitionsForTopic.size(); + List<Integer> currentAssignmentListForTopic = new ArrayList<>(assignedPartitionsForTopic); + + // If there were partitions from this topic that were previously assigned to this member present, retain as much as possible. + // We need to retain: + // 1) currentSize number of partitions (when currentSize < required) OR + // 2) required number of partitions (when currentSize > required). + if (currentAssignmentSize > 0) { + int retainedPartitionsCount = min(currentAssignmentSize, minRequiredQuota); + // The partitions need to be in ascending order since we want the same partition numbers from each topic + // to go to the same member to facilitate joins in case of co-partitioned topics. + Collections.sort(currentAssignmentListForTopic); + for (int i = 0; i < retainedPartitionsCount; i++) { + assignedStickyPartitionsForTopic + .add(currentAssignmentListForTopic.get(i)); + newAssignment.computeIfAbsent(memberId, k -> new MemberAssignment(new HashMap<>())) + .targetPartitions() + .computeIfAbsent(topicId, k -> new HashSet<>()) + .add(currentAssignmentListForTopic.get(i)); + } + } + + // Number of partitions required to meet the minRequiredQuota. + // There are 3 cases w.r.t the value of remaining: + // 1) remaining < 0: this means that the member has more than the min required amount. + // 2) If remaining = 0: member has the minimum required partitions, but it may get an extra partition, so it is a potentially unfilled member. + // 3) If remaining > 0: member doesn't have the minimum required partitions, so it should be added to potentiallyUnfilledMembers. + int remaining = minRequiredQuota - currentAssignmentSize; + + // In order to remain as sticky as possible, since the order of members can be different, we want the members that have more than the required number of partitions, + // to get the extra partition instead of assigning the extras to the first few members directly. + if (remaining < 0 && numMembersWithExtraPartition > 0) { + numMembersWithExtraPartition--; + // Since we already added the minimumRequiredQuota of partitions in the previous step (until minReq - 1), we just need to + // add the extra partition that will be present at the index right after min quota was satisfied. + assignedStickyPartitionsForTopic + .add(currentAssignmentListForTopic.get(minRequiredQuota)); + newAssignment.computeIfAbsent(memberId, k -> new MemberAssignment(new HashMap<>())) + .targetPartitions() + .computeIfAbsent(topicId, k -> new HashSet<>()) + .add(currentAssignmentListForTopic.get(minRequiredQuota)); + } else { + MemberWithRemainingAssignments newPair = new MemberWithRemainingAssignments(memberId, remaining); + potentiallyUnfilledMembers.add(newPair); + } + } + + List<MemberWithRemainingAssignments> unfilledMembersForTopic = new ArrayList<>(); + + // Step 3 + // If remaining > 0 after increasing the required quota due to the extra partition, add potentially unfilled member to the unfilled members list. + for (MemberWithRemainingAssignments pair : potentiallyUnfilledMembers) { + String memberId = pair.memberId; + Integer remaining = pair.remaining; + if (numMembersWithExtraPartition > 0) { + remaining++; + numMembersWithExtraPartition--; + } + if (remaining > 0) { + MemberWithRemainingAssignments newPair = new MemberWithRemainingAssignments(memberId, remaining); + unfilledMembersForTopic + .add(newPair); + } + } + + // Step 4 + // Find the difference between the total partitions per topic and the already assigned sticky partitions for the topic to get unassigned partitions. + // List of unassigned partitions per topic contains the partitions in ascending order. + List<Integer> unassignedPartitionsForTopic = new ArrayList<>(); + for (int i = 0; i < numPartitionsForTopic; i++) { + if (!assignedStickyPartitionsForTopic.contains(i)) { + unassignedPartitionsForTopic.add(i); + } + } + + // Step 5 + // Assign the remaining number of partitions from the unassigned partitions list. + int unassignedPartitionsListStartPointer = 0; + for (MemberWithRemainingAssignments memberWithRemainingAssignments : unfilledMembersForTopic) { + String memberId = memberWithRemainingAssignments.memberId; + int remaining = memberWithRemainingAssignments.remaining; + List<Integer> partitionsToAssign = unassignedPartitionsForTopic + .subList(unassignedPartitionsListStartPointer, unassignedPartitionsListStartPointer + remaining); + unassignedPartitionsListStartPointer += remaining; + newAssignment.computeIfAbsent(memberId, k -> new MemberAssignment(new HashMap<>())) + .targetPartitions() + .computeIfAbsent(topicId, k -> new HashSet<>()) + .addAll(partitionsToAssign); Review Comment: Ack. -- 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
