hachikuji commented on a change in pull request #10275: URL: https://github.com/apache/kafka/pull/10275#discussion_r598954950
########## File path: clients/src/main/java/org/apache/kafka/clients/admin/internals/AdminApiDriver.java ########## @@ -0,0 +1,461 @@ +/* + * 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.clients.admin.internals; + +import org.apache.kafka.clients.admin.internals.AdminApiHandler.Keys; +import org.apache.kafka.common.errors.DisconnectException; +import org.apache.kafka.common.internals.KafkaFutureImpl; +import org.apache.kafka.common.requests.AbstractRequest; +import org.apache.kafka.common.requests.AbstractResponse; +import org.apache.kafka.common.utils.LogContext; +import org.slf4j.Logger; + +import java.util.ArrayList; +import java.util.HashMap; +import java.util.HashSet; +import java.util.List; +import java.util.Map; +import java.util.Objects; +import java.util.Optional; +import java.util.OptionalInt; +import java.util.Set; +import java.util.function.BiFunction; + +/** + * The `KafkaAdminClient`'s internal `Call` primitive is not a good fit for multi-stage + * request workflows such as we see with the group coordinator APIs or any request which + * needs to be sent to a partition leader. Typically these APIs have two concrete stages: + * + * 1. Lookup: Find the broker that can fulfill the request (e.g. partition leader or group + * coordinator) + * 2. Fulfillment: Send the request to the broker found in the first step + * + * This is complicated by the fact that `Admin` APIs are typically batched, which + * means the Lookup stage may result in a set of brokers. For example, take a `ListOffsets` + * request for a set of topic partitions. In the Lookup stage, we will find the partition + * leaders for this set of partitions; in the Fulfillment stage, we will group together + * partition according to the IDs of the discovered leaders. + * + * Additionally, the flow between these two stages is bi-directional. We may find after + * sending a `ListOffsets` request to an expected leader that there was a leader change. + * This would result in a topic partition being sent back to the Lookup stage. + * + * Managing this complexity by chaining together `Call` implementations is challenging + * and messy, so instead we use this class to do the bookkeeping. It handles both the + * batching aspect as well as the transitions between the Lookup and Fulfillment stages. + * + * Note that the interpretation of the `retries` configuration becomes ambiguous + * for this kind of pipeline. We could treat it as an overall limit on the number + * of requests that can be sent, but that is not very useful because each pipeline + * has a minimum number of requests that need to be sent in order to satisfy the request. + * Instead, we treat this number of retries independently at each stage so that each + * stage has at least one opportunity to complete. So if a user sets `retries=1`, then + * the full pipeline can still complete as long as there are no request failures. + * + * @param <K> The key type, which is also the granularity of the request routing (e.g. + * this could be `TopicPartition` in the case of requests intended for a partition + * leader or the `GroupId` in the case of consumer group requests intended for + * the group coordinator) + * @param <V> The fulfillment type for each key (e.g. this could be consumer group state + * when the key type is a consumer `GroupId`) + */ +public class AdminApiDriver<K, V> { + private final Logger log; + private final long retryBackoffMs; + private final long deadlineMs; + private final AdminApiHandler<K, V> handler; + private final Keys<K> keys; + private final Map<K, KafkaFutureImpl<V>> futures; + + private final BiMultimap<ApiRequestScope, K> lookupMap = new BiMultimap<>(); + private final BiMultimap<FulfillmentScope, K> fulfillmentMap = new BiMultimap<>(); + private final Map<ApiRequestScope, RequestState> requestStates = new HashMap<>(); + + public AdminApiDriver( + AdminApiHandler<K, V> handler, + long deadlineMs, + long retryBackoffMs, + LogContext logContext + ) { + this.handler = handler; + this.deadlineMs = deadlineMs; + this.retryBackoffMs = retryBackoffMs; + this.log = logContext.logger(AdminApiDriver.class); + this.futures = new HashMap<>(); + this.keys = initializeKeys(handler); + } + + private Keys<K> initializeKeys(AdminApiHandler<K, V> handler) { + Keys<K> keys = handler.initializeKeys(); + + keys.staticKeys.forEach((key, brokerId) -> { + futures.put(key, new KafkaFutureImpl<>()); + map(key, brokerId); + }); + + keys.dynamicKeys.forEach(key -> { + futures.put(key, new KafkaFutureImpl<>()); + lookupMap.put(keys.lookupStrategy.lookupScope(key), key); + }); + + return keys; + } + + /** + * Associate a key with a brokerId. This is called after a response in the Lookup + * stage reveals the mapping (e.g. when the `FindCoordinator` tells us the the + * group coordinator for a specific consumer group). + */ + private void map(K key, Integer brokerId) { + lookupMap.remove(key); + fulfillmentMap.put(new FulfillmentScope(brokerId), key); + } + + /** + * Disassociate a key from the currently mapped brokerId. This will send the key + * back to the Lookup stage, which will allow us to attempt lookup again. + */ + private void unmap(K key) { + if (!keys.dynamicKeys.contains(key)) { + throw new IllegalStateException("Attempt to unmap key " + key + " which is not dynamically mapped"); + } + + fulfillmentMap.remove(key); + lookupMap.put(keys.lookupStrategy.lookupScope(key), key); + } + + private void clear(K key) { + lookupMap.remove(key); + fulfillmentMap.remove(key); + } + + OptionalInt keyToBrokerId(K key) { + Optional<FulfillmentScope> scope = fulfillmentMap.getKey(key); + if (scope.isPresent()) { + return OptionalInt.of(scope.get().destinationBrokerId); + } else { + return OptionalInt.empty(); + } + } + + /** + * Complete the future associated with the given key exceptionally. After is called, + * the key will be taken out of both the Lookup and Fulfillment stages so that request + * are not retried. + */ + private void completeExceptionally(K key, Throwable t) { + KafkaFutureImpl<V> future = futures.get(key); + if (future == null) { + log.warn("Attempt to complete future for {}, which was not requested", key); + } else { + clear(key); + future.completeExceptionally(t); + } + } + + /** + * Complete the future associated with the given key. After is called, the key will + * be taken out of both the Lookup and Fulfillment stages so that request are not retried. + */ + private void complete(K key, V value) { + KafkaFutureImpl<V> future = futures.get(key); + if (future == null) { + log.warn("Attempt to complete future for {}, which was not requested", key); + } else { + clear(key); + future.complete(value); + } + } + + /** + * Check whether any requests need to be sent. This should be called immediately + * after the driver is constructed and then again after each request returns + * (i.e. after {@link #onFailure(long, RequestSpec, Throwable)} or + * {@link #onResponse(long, RequestSpec, AbstractResponse)}). + * + * @return A list of requests that need to be sent + */ + public List<RequestSpec<K>> poll() { + List<RequestSpec<K>> requests = new ArrayList<>(); + collectLookupRequests(requests); + collectFulfillmentRequests(requests); + return requests; + } + + /** + * Get a map of the futures that are awaiting completion. + */ + public Map<K, KafkaFutureImpl<V>> futures() { + return futures; + } + + /** + * Callback that is invoked when a `Call` returns a response successfully. + */ + public void onResponse( + long currentTimeMs, + RequestSpec<K> spec, + AbstractResponse response + ) { + clearInflightRequest(currentTimeMs, spec); + + if (spec.scope instanceof FulfillmentScope) { + int brokerId = ((FulfillmentScope) spec.scope).destinationBrokerId; + AdminApiHandler.ApiResult<K, V> result = handler.handleResponse( + brokerId, + spec.keys, + response + ); + result.completedKeys.forEach(this::complete); + result.failedKeys.forEach(this::completeExceptionally); + result.unmappedKeys.forEach(this::unmap); + } else { + AdminApiLookupStrategy.LookupResult<K> result = keys.lookupStrategy.handleResponse( + spec.keys, + response + ); + result.failedKeys.forEach(this::completeExceptionally); + result.mappedKeys.forEach(this::map); + } + } + + /** + * Callback that is invoked when a `Call` is failed. + */ + public void onFailure( + long currentTimeMs, + RequestSpec<K> spec, + Throwable t + ) { + clearInflightRequest(currentTimeMs, spec); + if (t instanceof DisconnectException) { + log.debug("Node disconnected before response could be received for request {}. " + + "Will attempt retry", spec.request); + + // After a disconnect, we want the driver to attempt to lookup the key + // again (if the key is dynamically mapped). This gives us a chance to + // find a new coordinator or partition leader for example. + spec.keys.stream() + .filter(keys.dynamicKeys::contains) + .forEach(this::unmap); + } else { + spec.keys.forEach(key -> completeExceptionally(key, t)); + } + } + + private void clearInflightRequest(long currentTimeMs, RequestSpec<K> spec) { + RequestState requestState = requestStates.get(spec.scope); + if (requestState != null) { + requestState.clearInflight(currentTimeMs); + } + } + + private <T extends ApiRequestScope> void collectRequests( + List<RequestSpec<K>> requests, + BiMultimap<T, K> multimap, + BiFunction<Set<K>, T, AbstractRequest.Builder<?>> buildRequest + ) { + for (Map.Entry<T, Set<K>> entry : multimap.entrySet()) { + T scope = entry.getKey(); + + Set<K> keys = entry.getValue(); + if (keys.isEmpty()) { + continue; + } + + RequestState requestState = requestStates.computeIfAbsent(scope, c -> new RequestState()); + if (requestState.hasInflight()) { + continue; + } + + AbstractRequest.Builder<?> request = buildRequest.apply(keys, scope); + RequestSpec<K> spec = new RequestSpec<>( + handler.apiName() + "(api=" + request.apiKey() + ")", + scope, + new HashSet<>(keys), // copy to avoid exposing mutable state + request, + requestState.nextAllowedRetryMs, + deadlineMs, + requestState.tries + ); + + requestState.setInflight(spec); + requests.add(spec); + } + } + + private void collectLookupRequests(List<RequestSpec<K>> requests) { + if (!keys.dynamicKeys.isEmpty()) { + collectRequests( + requests, + lookupMap, + (keys, scope) -> this.keys.lookupStrategy.buildRequest(keys) + ); + } + } + + private void collectFulfillmentRequests(List<RequestSpec<K>> requests) { + collectRequests( + requests, + fulfillmentMap, + (keys, scope) -> handler.buildRequest(scope.destinationBrokerId, keys) + ); + } + + /** + * This is a helper class which helps us to map requests that need to be sent + * to the internal `Call` implementation that is used internally in + * {@link org.apache.kafka.clients.admin.KafkaAdminClient}. + */ + public static class RequestSpec<K> { + public final String name; + public final ApiRequestScope scope; + public final Set<K> keys; + public final AbstractRequest.Builder<?> request; + public final long nextAllowedTryMs; + public final long deadlineMs; + public final int tries; + + public RequestSpec( + String name, + ApiRequestScope scope, + Set<K> keys, + AbstractRequest.Builder<?> request, + long nextAllowedTryMs, + long deadlineMs, + int tries + ) { + this.name = name; + this.scope = scope; + this.keys = keys; + this.request = request; + this.nextAllowedTryMs = nextAllowedTryMs; + this.deadlineMs = deadlineMs; + this.tries = tries; + } + + @Override + public String toString() { + return "RequestSpec(" + + "name=" + name + + ", scope=" + scope + + ", keys=" + keys + + ", request=" + request + + ", nextAllowedTryMs=" + nextAllowedTryMs + + ", deadlineMs=" + deadlineMs + + ", tries=" + tries + + ')'; + } + } + + /** + * Helper class used to track the request state within each request scope. + * This class enforces a maximum number of inflight request and keeps track + * of backoff/retry state. + */ + private class RequestState { + private Optional<RequestSpec<K>> inflightRequest = Optional.empty(); + private int tries = 0; Review comment: The lookup requests can often result in ping-pong behavior. For example, we use metadata to find partition leader, then we send a request to the leader and get a NOT_LEADER error, which sends us back to the metadata request, and so on. I was not too concerned that the number of tries would be reflected perfectly (as I noted in the javadoc, the behavior is ambigious for these cases), but I wanted to at least ensure that the count continued growing. So if a user does configure a `retries` limit, then eventually we'll hit it even in these ping-pong scenarios if we cannot satisfy the request. -- 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. For queries about this service, please contact Infrastructure at: us...@infra.apache.org
