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));
126 void setSuccess(final int index) {
127 results.get(index).success(Optional.empty());
131 public String toString() {
132 return MoreObjects.toStringHelper(this).add("requests", requests).toString();
136 private static final class DeleteMessagesTo extends AsyncMessage<Void> {
137 final long toSequenceNr;
139 DeleteMessagesTo(final long toSequenceNr) {
140 this.toSequenceNr = toSequenceNr;
144 public String toString() {
145 return MoreObjects.toStringHelper(this).add("toSequenceNr", toSequenceNr).toString();
149 // responses == null on success, Exception on failure
150 record WrittenMessages(WriteMessages message, List<Object> responses, long writtenBytes) {
152 verify(responses.size() == message.size(), "Mismatched %s and %s", message, responses);
153 verify(writtenBytes >= 0, "Unexpected length %s", writtenBytes);
156 private void complete() {
157 for (int i = 0, size = responses.size(); i < size; ++i) {
158 if (responses.get(i) instanceof Exception ex) {
159 message.setFailure(i, ex);
161 message.setSuccess(i);
168 * A {@link SegmentedJournalActor} which delays issuing a flush operation until a watermark is reached or when the
172 * The problem we are addressing is that there is a queue sitting in from of the actor, which we have no direct
173 * access to. Since a flush involves committing data to durable storage, that operation can easily end up dominating
177 * We solve this by having an additional queue in which we track which messages were written and trigger a flush
178 * only when the number of bytes we have written exceeds specified limit. The other part is that each time this
179 * queue becomes non-empty, we send a dedicated message to self. This acts as a actor queue probe -- when we receive
180 * it, we know we have processed all messages that were in the queue when we first delayed the write.
183 * The combination of these mechanisms ensure we use a minimal delay while also ensuring we take advantage of
184 * batching opportunities.
186 private static final class Delayed extends SegmentedJournalActor {
187 private static final class Flush extends AsyncMessage<Void> {
190 Flush(final long batch) {
195 private record UnflushedWrite(WrittenMessages message, Stopwatch start, long count) {
197 requireNonNull(message);
198 requireNonNull(start);
202 private final ArrayDeque<UnflushedWrite> unflushedWrites = new ArrayDeque<>();
203 private final Stopwatch unflushedDuration = Stopwatch.createUnstarted();
204 private final long maxUnflushedBytes;
206 private long batch = 0;
207 private long unflushedBytes = 0;
209 Delayed(final String persistenceId, final File directory, final StorageLevel storage,
210 final int maxEntrySize, final int maxSegmentSize, final int maxUnflushedBytes) {
211 super(persistenceId, directory, storage, maxEntrySize, maxSegmentSize);
212 this.maxUnflushedBytes = maxUnflushedBytes;
216 ReceiveBuilder addMessages(final ReceiveBuilder builder) {
217 return super.addMessages(builder).match(Flush.class, this::handleFlush);
220 private void handleFlush(final Flush message) {
221 if (message.batch == batch) {
224 LOG.debug("{}: batch {} not flushed by {}", persistenceId(), batch, message.batch);
229 void onWrittenMessages(final WrittenMessages message, final Stopwatch started, final long count) {
230 boolean first = unflushedWrites.isEmpty();
232 unflushedDuration.start();
234 unflushedWrites.addLast(new UnflushedWrite(message, started, count));
235 unflushedBytes = unflushedBytes + message.writtenBytes;
236 if (unflushedBytes >= maxUnflushedBytes) {
237 LOG.debug("{}: reached {} unflushed journal bytes", persistenceId(), unflushedBytes);
240 LOG.debug("{}: deferring journal flush", persistenceId());
241 self().tell(new Flush(++batch), ActorRef.noSender());
247 final var unsyncedSize = unflushedWrites.size();
248 if (unsyncedSize == 0) {
253 LOG.debug("{}: flushing {} journal writes after {}", persistenceId(), unsyncedSize,
254 unflushedDuration.stop());
255 flushJournal(unflushedBytes, unsyncedSize);
257 final var sw = Stopwatch.createStarted();
258 unflushedWrites.forEach(write -> completeWriteMessages(write.message, write.start, write.count));
259 unflushedWrites.clear();
261 unflushedDuration.reset();
262 LOG.debug("{}: completed {} flushed journal writes in {}", persistenceId(), unsyncedSize, sw);
266 private static final class Immediate extends SegmentedJournalActor {
267 Immediate(final String persistenceId, final File directory, final StorageLevel storage,
268 final int maxEntrySize, final int maxSegmentSize) {
269 super(persistenceId, directory, storage, maxEntrySize, maxSegmentSize);
273 void onWrittenMessages(final WrittenMessages message, final Stopwatch started, final long count) {
274 flushJournal(message.writtenBytes, 1);
275 completeWriteMessages(message, started, count);
284 private static final Logger LOG = LoggerFactory.getLogger(SegmentedJournalActor.class);
285 private static final JournalSerdes DELETE_NAMESPACE = JournalSerdes.builder()
286 .register(LongEntrySerdes.LONG_ENTRY_SERDES, Long.class)
288 private static final int DELETE_SEGMENT_SIZE = 64 * 1024;
290 private final String persistenceId;
291 private final StorageLevel storage;
292 private final int maxSegmentSize;
293 private final int maxEntrySize;
294 private final File directory;
296 // Tracks the time it took us to write a batch of messages
297 private Timer batchWriteTime;
298 // Tracks the number of individual messages written
299 private Meter messageWriteCount;
300 // Tracks the size distribution of messages
301 private Histogram messageSize;
302 // Tracks the number of messages completed for each flush
303 private Histogram flushMessages;
304 // Tracks the number of bytes completed for each flush
305 private Histogram flushBytes;
306 // Tracks the duration of flush operations
307 private Timer flushTime;
309 private DataJournal dataJournal;
310 private SegmentedJournal<Long> deleteJournal;
311 private long lastDelete;
313 private SegmentedJournalActor(final String persistenceId, final File directory, final StorageLevel storage,
314 final int maxEntrySize, final int maxSegmentSize) {
315 this.persistenceId = requireNonNull(persistenceId);
316 this.directory = requireNonNull(directory);
317 this.storage = requireNonNull(storage);
318 this.maxEntrySize = maxEntrySize;
319 this.maxSegmentSize = maxSegmentSize;
322 static Props props(final String persistenceId, final File directory, final StorageLevel storage,
323 final int maxEntrySize, final int maxSegmentSize, final int maxUnflushedBytes) {
324 final var pid = requireNonNull(persistenceId);
325 return maxUnflushedBytes > 0
326 ? Props.create(Delayed.class, pid, directory, storage, maxEntrySize, maxSegmentSize, maxUnflushedBytes)
327 : Props.create(Immediate.class, pid, directory, storage, maxEntrySize, maxSegmentSize);
330 final String persistenceId() {
331 return persistenceId;
334 final void flushJournal(final long bytes, final int messages) {
335 final var sw = Stopwatch.createStarted();
337 LOG.debug("{}: journal flush completed in {}", persistenceId, sw.stop());
338 flushBytes.update(bytes);
339 flushMessages.update(messages);
340 flushTime.update(sw.elapsed(TimeUnit.NANOSECONDS), TimeUnit.NANOSECONDS);
344 public Receive createReceive() {
345 return addMessages(receiveBuilder())
346 .matchAny(this::handleUnknown)
350 ReceiveBuilder addMessages(final ReceiveBuilder builder) {
352 .match(DeleteMessagesTo.class, this::handleDeleteMessagesTo)
353 .match(ReadHighestSequenceNr.class, this::handleReadHighestSequenceNr)
354 .match(ReplayMessages.class, this::handleReplayMessages)
355 .match(WriteMessages.class, this::handleWriteMessages);
359 public void preStart() throws Exception {
360 LOG.debug("{}: actor starting", persistenceId);
363 final var registry = MetricsReporter.getInstance(MeteringBehavior.DOMAIN).getMetricsRegistry();
364 final var actorName = self().path().parent().toStringWithoutAddress() + '/' + directory.getName();
366 batchWriteTime = registry.timer(MetricRegistry.name(actorName, "batchWriteTime"));
367 messageWriteCount = registry.meter(MetricRegistry.name(actorName, "messageWriteCount"));
368 messageSize = registry.histogram(MetricRegistry.name(actorName, "messageSize"));
369 flushBytes = registry.histogram(MetricRegistry.name(actorName, "flushBytes"));
370 flushMessages = registry.histogram(MetricRegistry.name(actorName, "flushMessages"));
371 flushTime = registry.timer(MetricRegistry.name(actorName, "flushTime"));
375 public void postStop() throws Exception {
376 LOG.debug("{}: actor stopping", persistenceId);
377 if (dataJournal != null) {
379 LOG.debug("{}: data journal closed", persistenceId);
382 if (deleteJournal != null) {
383 deleteJournal.close();
384 LOG.debug("{}: delete journal closed", persistenceId);
385 deleteJournal = null;
387 LOG.debug("{}: actor stopped", persistenceId);
391 static AsyncMessage<Void> deleteMessagesTo(final long toSequenceNr) {
392 return new DeleteMessagesTo(toSequenceNr);
395 static AsyncMessage<Long> readHighestSequenceNr(final long fromSequenceNr) {
396 return new ReadHighestSequenceNr(fromSequenceNr);
399 static AsyncMessage<Void> replayMessages(final long fromSequenceNr, final long toSequenceNr, final long max,
400 final Consumer<PersistentRepr> replayCallback) {
401 return new ReplayMessages(fromSequenceNr, toSequenceNr, max, replayCallback);
404 private void handleDeleteMessagesTo(final DeleteMessagesTo message) {
407 LOG.debug("{}: delete messages {}", persistenceId, message);
410 final long to = Long.min(dataJournal.lastWrittenSequenceNr(), message.toSequenceNr);
411 LOG.debug("{}: adjusted delete to {}", persistenceId, to);
413 if (lastDelete < to) {
414 LOG.debug("{}: deleting entries up to {}", persistenceId, to);
417 final var deleteWriter = deleteJournal.writer();
418 final var entry = deleteWriter.append(lastDelete);
419 deleteWriter.commit(entry.index());
420 dataJournal.deleteTo(lastDelete);
422 LOG.debug("{}: compaction started", persistenceId);
423 dataJournal.compactTo(lastDelete);
424 deleteJournal.compact(entry.index());
425 LOG.debug("{}: compaction finished", persistenceId);
427 LOG.debug("{}: entries up to {} already deleted", persistenceId, lastDelete);
430 message.promise.success(null);
433 private void handleReadHighestSequenceNr(final ReadHighestSequenceNr message) {
434 LOG.debug("{}: looking for highest sequence on {}", persistenceId, message);
436 if (directory.isDirectory()) {
439 sequence = dataJournal.lastWrittenSequenceNr();
444 LOG.debug("{}: highest sequence is {}", message, sequence);
445 message.promise.success(sequence);
448 private void handleReplayMessages(final ReplayMessages message) {
449 LOG.debug("{}: replaying messages {}", persistenceId, message);
453 final long from = Long.max(lastDelete + 1, message.fromSequenceNr);
454 LOG.debug("{}: adjusted fromSequenceNr to {}", persistenceId, from);
456 dataJournal.handleReplayMessages(message, from);
459 private void handleWriteMessages(final WriteMessages message) {
462 final var started = Stopwatch.createStarted();
463 final long start = dataJournal.lastWrittenSequenceNr();
464 final var writtenMessages = dataJournal.handleWriteMessages(message);
466 onWrittenMessages(writtenMessages, started, dataJournal.lastWrittenSequenceNr() - start);
469 final void completeWriteMessages(final WrittenMessages message, final Stopwatch started, final long count) {
470 batchWriteTime.update(started.stop().elapsed(TimeUnit.NANOSECONDS), TimeUnit.NANOSECONDS);
471 messageWriteCount.mark(count);
472 // log message after statistics are updated
473 LOG.debug("{}: write of {} bytes completed in {}", persistenceId, message.writtenBytes, started);
478 * Handle a check of written messages.
480 * @param message Messages which were written
481 * @param started Stopwatch started when the write started
482 * @param count number of writes
484 abstract void onWrittenMessages(WrittenMessages message, Stopwatch started, long count);
486 private void handleUnknown(final Object message) {
487 LOG.error("{}: Received unknown message {}", persistenceId, message);
490 private void ensureOpen() {
491 if (dataJournal != null) {
492 verifyNotNull(deleteJournal);
496 final var sw = Stopwatch.createStarted();
497 deleteJournal = SegmentedJournal.<Long>builder().withDirectory(directory).withName("delete")
498 .withNamespace(DELETE_NAMESPACE).withMaxSegmentSize(DELETE_SEGMENT_SIZE).build();
499 final var lastDeleteRecovered = deleteJournal.openReader(deleteJournal.writer().getLastIndex())
500 .tryNext((index, value, length) -> value);
501 lastDelete = lastDeleteRecovered == null ? 0 : lastDeleteRecovered.longValue();
503 dataJournal = new DataJournalV0(persistenceId, messageSize, context().system(), storage, directory,
504 maxEntrySize, maxSegmentSize);
505 dataJournal.deleteTo(lastDelete);
506 LOG.debug("{}: journal open in {} with last index {}, deleted to {}", persistenceId, sw,
507 dataJournal.lastWrittenSequenceNr(), lastDelete);
510 abstract void flushWrites();