2 * Copyright (c) 2019 Pantheon Technologies, s.r.o. and others. All rights reserved.
4 * This program and the accompanying materials are made available under the
5 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
6 * and is available at http://www.eclipse.org/legal/epl-v10.html
8 package org.opendaylight.controller.akka.segjournal;
10 import static com.google.common.base.Verify.verify;
11 import static com.google.common.base.Verify.verifyNotNull;
12 import static java.util.Objects.requireNonNull;
14 import akka.actor.AbstractActor;
15 import akka.actor.ActorRef;
16 import akka.actor.Props;
17 import akka.japi.pf.ReceiveBuilder;
18 import akka.persistence.AtomicWrite;
19 import akka.persistence.PersistentRepr;
20 import com.codahale.metrics.Histogram;
21 import com.codahale.metrics.Meter;
22 import com.codahale.metrics.MetricRegistry;
23 import com.codahale.metrics.Timer;
24 import com.google.common.base.MoreObjects;
25 import com.google.common.base.Stopwatch;
26 import io.atomix.storage.journal.Indexed;
27 import io.atomix.storage.journal.JournalSerdes;
28 import io.atomix.storage.journal.SegmentedJournal;
29 import io.atomix.storage.journal.StorageLevel;
31 import java.util.ArrayDeque;
32 import java.util.ArrayList;
33 import java.util.List;
34 import java.util.Optional;
35 import java.util.concurrent.TimeUnit;
36 import java.util.function.Consumer;
37 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
38 import org.opendaylight.controller.cluster.reporting.MetricsReporter;
39 import org.slf4j.Logger;
40 import org.slf4j.LoggerFactory;
41 import scala.concurrent.Future;
42 import scala.concurrent.Promise;
45 * This actor handles a single PersistentActor's journal. The journal is split into two {@link SegmentedJournal}s:
47 * <li>A memory-mapped data journal, containing actual data entries</li>
48 * <li>A simple file journal, containing sequence numbers of last deleted entry</li>
52 * This is a conscious design decision to minimize the amount of data that is being stored in the data journal while
53 * speeding up normal operations. Since the SegmentedJournal is an append-only linear log and Akka requires the ability
54 * to delete persistence entries, we need ability to mark a subset of a SegmentedJournal as deleted. While we could
55 * treat such delete requests as normal events, this leads to a mismatch between SegmentedJournal indices (as exposed by
56 * {@link Indexed}) and Akka sequence numbers -- requiring us to potentially perform costly deserialization to find the
57 * index corresponding to a particular sequence number, or maintain moderately-complex logic and data structures to
58 * perform that mapping in sub-linear time complexity.
61 * Split-file approach allows us to treat sequence numbers and indices as equivalent, without maintaining any explicit
62 * mapping information. The only additional information we need to maintain is the last deleted sequence number.
64 abstract sealed class SegmentedJournalActor extends AbstractActor {
65 abstract static sealed class AsyncMessage<T> {
66 final Promise<T> promise = Promise.apply();
69 private static final class ReadHighestSequenceNr extends AsyncMessage<Long> {
70 private final long fromSequenceNr;
72 ReadHighestSequenceNr(final long fromSequenceNr) {
73 this.fromSequenceNr = fromSequenceNr;
77 public String toString() {
78 return MoreObjects.toStringHelper(this).add("fromSequenceNr", fromSequenceNr).toString();
82 static final class ReplayMessages extends AsyncMessage<Void> {
83 private final long fromSequenceNr;
84 final long toSequenceNr;
86 final Consumer<PersistentRepr> replayCallback;
88 ReplayMessages(final long fromSequenceNr,
89 final long toSequenceNr, final long max, final Consumer<PersistentRepr> replayCallback) {
90 this.fromSequenceNr = fromSequenceNr;
91 this.toSequenceNr = toSequenceNr;
93 this.replayCallback = requireNonNull(replayCallback);
97 public String toString() {
98 return MoreObjects.toStringHelper(this).add("fromSequenceNr", fromSequenceNr)
99 .add("toSequenceNr", toSequenceNr).add("max", max).toString();
103 static final class WriteMessages {
104 private final List<AtomicWrite> requests = new ArrayList<>();
105 private final List<Promise<Optional<Exception>>> results = new ArrayList<>();
107 Future<Optional<Exception>> add(final AtomicWrite write) {
108 final var promise = Promise.<Optional<Exception>>apply();
110 results.add(promise);
111 return promise.future();
115 return requests.size();
118 AtomicWrite getRequest(final int index) {
119 return requests.get(index);
122 void setFailure(final int index, final Exception cause) {
123 results.get(index).success(Optional.of(cause));
127 void setSuccess(final int index) {
128 results.get(index).success(Optional.empty());
132 public String toString() {
133 return MoreObjects.toStringHelper(this).add("requests", requests).toString();
137 private static final class DeleteMessagesTo extends AsyncMessage<Void> {
138 final long toSequenceNr;
140 DeleteMessagesTo(final long toSequenceNr) {
141 this.toSequenceNr = toSequenceNr;
145 public String toString() {
146 return MoreObjects.toStringHelper(this).add("toSequenceNr", toSequenceNr).toString();
150 // responses == null on success, Exception on failure
151 record WrittenMessages(WriteMessages message, List<Object> responses, long writtenBytes) {
153 verify(responses.size() == message.size(), "Mismatched %s and %s", message, responses);
154 verify(writtenBytes >= 0, "Unexpected length %s", writtenBytes);
157 private void complete() {
158 for (int i = 0, size = responses.size(); i < size; ++i) {
159 if (responses.get(i) instanceof Exception ex) {
160 message.setFailure(i, ex);
162 message.setSuccess(i);
169 * A {@link SegmentedJournalActor} which delays issuing a flush operation until a watermark is reached or when the
173 * The problem we are addressing is that there is a queue sitting in from of the actor, which we have no direct
174 * access to. Since a flush involves committing data to durable storage, that operation can easily end up dominating
178 * We solve this by having an additional queue in which we track which messages were written and trigger a flush
179 * only when the number of bytes we have written exceeds specified limit. The other part is that each time this
180 * queue becomes non-empty, we send a dedicated message to self. This acts as a actor queue probe -- when we receive
181 * it, we know we have processed all messages that were in the queue when we first delayed the write.
184 * The combination of these mechanisms ensure we use a minimal delay while also ensuring we take advantage of
185 * batching opportunities.
187 private static final class Delayed extends SegmentedJournalActor {
188 private static final class Flush extends AsyncMessage<Void> {
191 Flush(final long batch) {
196 private final ArrayDeque<WrittenMessages> unflushedWrites = new ArrayDeque<>();
197 private final Stopwatch unflushedDuration = Stopwatch.createUnstarted();
198 private final long maxUnflushedBytes;
200 private long batch = 0;
201 private long unflushedBytes = 0;
203 Delayed(final String persistenceId, final File directory, final StorageLevel storage,
204 final int maxEntrySize, final int maxSegmentSize, final int maxUnflushedBytes) {
205 super(persistenceId, directory, storage, maxEntrySize, maxSegmentSize);
206 this.maxUnflushedBytes = maxUnflushedBytes;
210 ReceiveBuilder addMessages(final ReceiveBuilder builder) {
211 return super.addMessages(builder).match(Flush.class, this::handleFlush);
214 private void handleFlush(final Flush message) {
215 if (message.batch == batch) {
218 LOG.debug("{}: batch {} not flushed by {}", persistenceId(), batch, message.batch);
223 void onWrittenMessages(final WrittenMessages message) {
224 boolean first = unflushedWrites.isEmpty();
226 unflushedDuration.start();
228 unflushedWrites.addLast(message);
229 unflushedBytes = unflushedBytes + message.writtenBytes;
230 if (unflushedBytes >= maxUnflushedBytes) {
231 LOG.debug("{}: reached {} unflushed journal bytes", persistenceId(), unflushedBytes);
234 LOG.debug("{}: deferring journal flush", persistenceId());
235 self().tell(new Flush(++batch), ActorRef.noSender());
241 final var unsyncedSize = unflushedWrites.size();
242 if (unsyncedSize == 0) {
247 LOG.debug("{}: flushing {} journal writes after {}", persistenceId(), unsyncedSize,
248 unflushedDuration.stop());
249 flushJournal(unflushedBytes, unsyncedSize);
251 final var sw = Stopwatch.createStarted();
252 unflushedWrites.forEach(WrittenMessages::complete);
253 unflushedWrites.clear();
255 unflushedDuration.reset();
256 LOG.debug("{}: completed {} flushed journal writes in {}", persistenceId(), unsyncedSize, sw);
260 private static final class Immediate extends SegmentedJournalActor {
261 Immediate(final String persistenceId, final File directory, final StorageLevel storage,
262 final int maxEntrySize, final int maxSegmentSize) {
263 super(persistenceId, directory, storage, maxEntrySize, maxSegmentSize);
267 void onWrittenMessages(final WrittenMessages message) {
268 flushJournal(message.writtenBytes, 1);
278 private static final Logger LOG = LoggerFactory.getLogger(SegmentedJournalActor.class);
279 private static final JournalSerdes DELETE_NAMESPACE = JournalSerdes.builder()
280 .register(LongEntrySerdes.LONG_ENTRY_SERDES, Long.class)
282 private static final int DELETE_SEGMENT_SIZE = 64 * 1024;
284 private final String persistenceId;
285 private final StorageLevel storage;
286 private final int maxSegmentSize;
287 private final int maxEntrySize;
288 private final File directory;
290 // Tracks the time it took us to write a batch of messages
291 private Timer batchWriteTime;
292 // Tracks the number of individual messages written
293 private Meter messageWriteCount;
294 // Tracks the size distribution of messages
295 private Histogram messageSize;
296 // Tracks the number of messages completed for each flush
297 private Histogram flushMessages;
298 // Tracks the number of bytes completed for each flush
299 private Histogram flushBytes;
300 // Tracks the duration of flush operations
301 private Timer flushTime;
303 private DataJournal dataJournal;
304 private SegmentedJournal<Long> deleteJournal;
305 private long lastDelete;
307 private SegmentedJournalActor(final String persistenceId, final File directory, final StorageLevel storage,
308 final int maxEntrySize, final int maxSegmentSize) {
309 this.persistenceId = requireNonNull(persistenceId);
310 this.directory = requireNonNull(directory);
311 this.storage = requireNonNull(storage);
312 this.maxEntrySize = maxEntrySize;
313 this.maxSegmentSize = maxSegmentSize;
316 static Props props(final String persistenceId, final File directory, final StorageLevel storage,
317 final int maxEntrySize, final int maxSegmentSize, final int maxUnflushedBytes) {
318 final var pid = requireNonNull(persistenceId);
319 return maxUnflushedBytes > 0
320 ? Props.create(Delayed.class, pid, directory, storage, maxEntrySize, maxSegmentSize, maxUnflushedBytes)
321 : Props.create(Immediate.class, pid, directory, storage, maxEntrySize, maxSegmentSize);
324 final String persistenceId() {
325 return persistenceId;
328 final void flushJournal(final long bytes, final int messages) {
329 final var sw = Stopwatch.createStarted();
331 LOG.debug("{}: journal flush completed in {}", persistenceId, sw.stop());
332 flushBytes.update(bytes);
333 flushMessages.update(messages);
334 flushTime.update(sw.elapsed(TimeUnit.NANOSECONDS), TimeUnit.NANOSECONDS);
338 public Receive createReceive() {
339 return addMessages(receiveBuilder())
340 .matchAny(this::handleUnknown)
344 ReceiveBuilder addMessages(final ReceiveBuilder builder) {
346 .match(DeleteMessagesTo.class, this::handleDeleteMessagesTo)
347 .match(ReadHighestSequenceNr.class, this::handleReadHighestSequenceNr)
348 .match(ReplayMessages.class, this::handleReplayMessages)
349 .match(WriteMessages.class, this::handleWriteMessages);
353 public void preStart() throws Exception {
354 LOG.debug("{}: actor starting", persistenceId);
357 final var registry = MetricsReporter.getInstance(MeteringBehavior.DOMAIN).getMetricsRegistry();
358 final var actorName = self().path().parent().toStringWithoutAddress() + '/' + directory.getName();
360 batchWriteTime = registry.timer(MetricRegistry.name(actorName, "batchWriteTime"));
361 messageWriteCount = registry.meter(MetricRegistry.name(actorName, "messageWriteCount"));
362 messageSize = registry.histogram(MetricRegistry.name(actorName, "messageSize"));
363 flushBytes = registry.histogram(MetricRegistry.name(actorName, "flushBytes"));
364 flushMessages = registry.histogram(MetricRegistry.name(actorName, "flushMessages"));
365 flushTime = registry.timer(MetricRegistry.name(actorName, "flushTime"));
369 public void postStop() throws Exception {
370 LOG.debug("{}: actor stopping", persistenceId);
371 if (dataJournal != null) {
373 LOG.debug("{}: data journal closed", persistenceId);
376 if (deleteJournal != null) {
377 deleteJournal.close();
378 LOG.debug("{}: delete journal closed", persistenceId);
379 deleteJournal = null;
381 LOG.debug("{}: actor stopped", persistenceId);
385 static AsyncMessage<Void> deleteMessagesTo(final long toSequenceNr) {
386 return new DeleteMessagesTo(toSequenceNr);
389 static AsyncMessage<Long> readHighestSequenceNr(final long fromSequenceNr) {
390 return new ReadHighestSequenceNr(fromSequenceNr);
393 static AsyncMessage<Void> replayMessages(final long fromSequenceNr, final long toSequenceNr, final long max,
394 final Consumer<PersistentRepr> replayCallback) {
395 return new ReplayMessages(fromSequenceNr, toSequenceNr, max, replayCallback);
398 private void handleDeleteMessagesTo(final DeleteMessagesTo message) {
401 LOG.debug("{}: delete messages {}", persistenceId, message);
404 final long to = Long.min(dataJournal.lastWrittenSequenceNr(), message.toSequenceNr);
405 LOG.debug("{}: adjusted delete to {}", persistenceId, to);
407 if (lastDelete < to) {
408 LOG.debug("{}: deleting entries up to {}", persistenceId, to);
411 final var deleteWriter = deleteJournal.writer();
412 final var entry = deleteWriter.append(lastDelete);
413 deleteWriter.commit(entry.index());
414 dataJournal.deleteTo(lastDelete);
416 LOG.debug("{}: compaction started", persistenceId);
417 dataJournal.compactTo(lastDelete);
418 deleteJournal.compact(entry.index());
419 LOG.debug("{}: compaction finished", persistenceId);
421 LOG.debug("{}: entries up to {} already deleted", persistenceId, lastDelete);
424 message.promise.success(null);
427 private void handleReadHighestSequenceNr(final ReadHighestSequenceNr message) {
428 LOG.debug("{}: looking for highest sequence on {}", persistenceId, message);
430 if (directory.isDirectory()) {
433 sequence = dataJournal.lastWrittenSequenceNr();
438 LOG.debug("{}: highest sequence is {}", message, sequence);
439 message.promise.success(sequence);
442 private void handleReplayMessages(final ReplayMessages message) {
443 LOG.debug("{}: replaying messages {}", persistenceId, message);
447 final long from = Long.max(lastDelete + 1, message.fromSequenceNr);
448 LOG.debug("{}: adjusted fromSequenceNr to {}", persistenceId, from);
450 dataJournal.handleReplayMessages(message, from);
453 private void handleWriteMessages(final WriteMessages message) {
456 final var sw = Stopwatch.createStarted();
457 final long start = dataJournal.lastWrittenSequenceNr();
458 final var writtenMessages = dataJournal.handleWriteMessages(message);
461 batchWriteTime.update(sw.elapsed(TimeUnit.NANOSECONDS), TimeUnit.NANOSECONDS);
462 messageWriteCount.mark(dataJournal.lastWrittenSequenceNr() - start);
464 // log message after statistics are updated
465 LOG.debug("{}: write of {} bytes completed in {}", persistenceId, writtenMessages.writtenBytes, sw);
466 onWrittenMessages(writtenMessages);
470 * Handle a check of written messages.
472 * @param message Messages which were written
474 abstract void onWrittenMessages(WrittenMessages message);
476 private void handleUnknown(final Object message) {
477 LOG.error("{}: Received unknown message {}", persistenceId, message);
480 private void ensureOpen() {
481 if (dataJournal != null) {
482 verifyNotNull(deleteJournal);
486 final var sw = Stopwatch.createStarted();
487 deleteJournal = SegmentedJournal.<Long>builder().withDirectory(directory).withName("delete")
488 .withNamespace(DELETE_NAMESPACE).withMaxSegmentSize(DELETE_SEGMENT_SIZE).build();
489 final var lastEntry = deleteJournal.writer().getLastEntry();
490 lastDelete = lastEntry == null ? 0 : lastEntry.entry();
492 dataJournal = new DataJournalV0(persistenceId, messageSize, context().system(), storage, directory,
493 maxEntrySize, maxSegmentSize);
494 dataJournal.deleteTo(lastDelete);
495 LOG.debug("{}: journal open in {} with last index {}, deleted to {}", persistenceId, sw,
496 dataJournal.lastWrittenSequenceNr(), lastDelete);
499 abstract void flushWrites();