private void snapshotTranslog(Translog.Snapshot snapshot, CommitPoint.FileInfo fileInfo)
throws IOException {
blobContainer.writeBlob(fileInfo.name(), snapshot.stream(), snapshot.lengthInBytes());
//
// long chunkBytes = Long.MAX_VALUE;
// if (chunkSize != null) {
// chunkBytes = chunkSize.bytes();
// }
//
// long totalLength = fileInfo.length();
// long numberOfChunks = totalLength / chunkBytes;
// if (totalLength % chunkBytes > 0) {
// numberOfChunks++;
// }
// if (numberOfChunks == 0) {
// numberOfChunks++;
// }
//
// if (numberOfChunks == 1) {
// blobContainer.writeBlob(fileInfo.name(), snapshot.stream(),
// snapshot.lengthInBytes());
// } else {
// InputStream translogStream = snapshot.stream();
// long totalLengthLeftToWrite = totalLength;
// for (int i = 0; i < numberOfChunks; i++) {
// long lengthToWrite = chunkBytes;
// if (totalLengthLeftToWrite < chunkBytes) {
// lengthToWrite = totalLengthLeftToWrite;
// }
// blobContainer.writeBlob(fileInfo.name() + ".part" + i, new
// LimitInputStream(translogStream, lengthToWrite), lengthToWrite);
// totalLengthLeftToWrite -= lengthToWrite;
// }
// }
}
private CommitPoints buildCommitPoints(ImmutableMap<String, BlobMetaData> blobs) {
List<CommitPoint> commitPoints = Lists.newArrayList();
for (String name : blobs.keySet()) {
if (name.startsWith("commit-")) {
try {
commitPoints.add(CommitPoints.fromXContent(blobContainer.readBlobFully(name)));
} catch (Exception e) {
logger.warn("failed to read commit point [{}]", e, name);
}
}
}
return new CommitPoints(commitPoints);
}
private void snapshotFile(
Directory dir,
final CommitPoint.FileInfo fileInfo,
final CountDownLatch latch,
final List<Throwable> failures)
throws IOException {
long chunkBytes = Long.MAX_VALUE;
if (chunkSize != null) {
chunkBytes = chunkSize.bytes();
}
long totalLength = fileInfo.length();
long numberOfChunks = totalLength / chunkBytes;
if (totalLength % chunkBytes > 0) {
numberOfChunks++;
}
if (numberOfChunks == 0) {
numberOfChunks++;
}
final long fNumberOfChunks = numberOfChunks;
final AtomicLong counter = new AtomicLong(numberOfChunks);
for (long i = 0; i < fNumberOfChunks; i++) {
final long partNumber = i;
IndexInput indexInput = null;
try {
indexInput = dir.openInput(fileInfo.physicalName());
indexInput.seek(partNumber * chunkBytes);
InputStreamIndexInput is = new ThreadSafeInputStreamIndexInput(indexInput, chunkBytes);
String blobName = fileInfo.name();
if (fNumberOfChunks > 1) {
// if we do chunks, then all of them are in the form of "[xxx].part[N]".
blobName += ".part" + partNumber;
}
final IndexInput fIndexInput = indexInput;
blobContainer.writeBlob(
blobName,
is,
is.actualSizeToRead(),
new ImmutableBlobContainer.WriterListener() {
@Override
public void onCompleted() {
try {
fIndexInput.close();
} catch (IOException e) {
// ignore
}
if (counter.decrementAndGet() == 0) {
latch.countDown();
}
}
@Override
public void onFailure(Throwable t) {
try {
fIndexInput.close();
} catch (IOException e) {
// ignore
}
failures.add(t);
if (counter.decrementAndGet() == 0) {
latch.countDown();
}
}
});
} catch (Exception e) {
if (indexInput != null) {
try {
indexInput.close();
} catch (IOException e1) {
// ignore
}
}
failures.add(e);
latch.countDown();
}
}
}
private void recoverFile(
final CommitPoint.FileInfo fileInfo,
final ImmutableMap<String, BlobMetaData> blobs,
final CountDownLatch latch,
final List<Throwable> failures) {
final IndexOutput indexOutput;
try {
// we create an output with no checksum, this is because the pure binary data of the file is
// not
// the checksum (because of seek). We will create the checksum file once copying is done
indexOutput = store.createOutputWithNoChecksum(fileInfo.physicalName());
} catch (IOException e) {
failures.add(e);
latch.countDown();
return;
}
String firstFileToRecover = fileInfo.name();
if (!blobs.containsKey(fileInfo.name())) {
// chunking, append part0 to it
firstFileToRecover = fileInfo.name() + ".part0";
}
if (!blobs.containsKey(firstFileToRecover)) {
// no file, what to do, what to do?
logger.warn(
"no file [{}]/[{}] to recover, ignoring it", fileInfo.name(), fileInfo.physicalName());
latch.countDown();
return;
}
final AtomicInteger partIndex = new AtomicInteger();
blobContainer.readBlob(
firstFileToRecover,
new BlobContainer.ReadBlobListener() {
@Override
public synchronized void onPartial(byte[] data, int offset, int size) throws IOException {
recoveryStatus.index().addCurrentFilesSize(size);
indexOutput.writeBytes(data, offset, size);
}
@Override
public synchronized void onCompleted() {
int part = partIndex.incrementAndGet();
String partName = fileInfo.name() + ".part" + part;
if (blobs.containsKey(partName)) {
// continue with the new part
blobContainer.readBlob(partName, this);
return;
} else {
// we are done...
try {
indexOutput.close();
// write the checksum
if (fileInfo.checksum() != null) {
store.writeChecksum(fileInfo.physicalName(), fileInfo.checksum());
}
store.directory().sync(Collections.singleton(fileInfo.physicalName()));
} catch (IOException e) {
onFailure(e);
return;
}
}
latch.countDown();
}
@Override
public void onFailure(Throwable t) {
failures.add(t);
latch.countDown();
}
});
}
private void recoverTranslog(CommitPoint commitPoint, ImmutableMap<String, BlobMetaData> blobs)
throws IndexShardGatewayRecoveryException {
if (commitPoint.translogFiles().isEmpty()) {
// no translog files, bail
indexShard.start("post recovery from gateway, no translog");
return;
}
try {
indexShard.performRecoveryPrepareForTranslog();
final AtomicReference<Throwable> failure = new AtomicReference<Throwable>();
final CountDownLatch latch = new CountDownLatch(1);
final Iterator<CommitPoint.FileInfo> transIt = commitPoint.translogFiles().iterator();
blobContainer.readBlob(
transIt.next().name(),
new BlobContainer.ReadBlobListener() {
FastByteArrayOutputStream bos = new FastByteArrayOutputStream();
boolean ignore = false;
@Override
public synchronized void onPartial(byte[] data, int offset, int size)
throws IOException {
if (ignore) {
return;
}
bos.write(data, offset, size);
// if we don't have enough to read the header size of the first translog, bail and
// wait for the next one
if (bos.size() < 4) {
return;
}
BytesStreamInput si = new BytesStreamInput(bos.unsafeByteArray(), 0, bos.size());
int position;
while (true) {
try {
position = si.position();
if (position + 4 > bos.size()) {
break;
}
int opSize = si.readInt();
int curPos = si.position();
if ((si.position() + opSize) > bos.size()) {
break;
}
Translog.Operation operation = TranslogStreams.readTranslogOperation(si);
if ((si.position() - curPos) != opSize) {
logger.warn(
"mismatch in size, expected [{}], got [{}]",
opSize,
si.position() - curPos);
}
recoveryStatus.translog().addTranslogOperations(1);
indexShard.performRecoveryOperation(operation);
if (si.position() >= bos.size()) {
position = si.position();
break;
}
} catch (Exception e) {
logger.warn(
"failed to retrieve translog after [{}] operations, ignoring the rest, considered corrupted",
e,
recoveryStatus.translog().currentTranslogOperations());
ignore = true;
latch.countDown();
return;
}
}
FastByteArrayOutputStream newBos = new FastByteArrayOutputStream();
int leftOver = bos.size() - position;
if (leftOver > 0) {
newBos.write(bos.unsafeByteArray(), position, leftOver);
}
bos = newBos;
}
@Override
public synchronized void onCompleted() {
if (ignore) {
return;
}
if (!transIt.hasNext()) {
latch.countDown();
return;
}
blobContainer.readBlob(transIt.next().name(), this);
}
@Override
public void onFailure(Throwable t) {
failure.set(t);
latch.countDown();
}
});
latch.await();
if (failure.get() != null) {
throw failure.get();
}
indexShard.performRecoveryFinalization(true);
} catch (Throwable e) {
throw new IndexShardGatewayRecoveryException(shardId, "Failed to recover translog", e);
}
}
@Override
public void recover(RecoveryStatus recoveryStatus) throws IndexShardGatewayRecoveryException {
this.recoveryStatus = recoveryStatus;
final ImmutableMap<String, BlobMetaData> blobs;
try {
blobs = blobContainer.listBlobs();
} catch (IOException e) {
throw new IndexShardGatewayRecoveryException(shardId, "Failed to list content of gateway", e);
}
List<CommitPoint> commitPointsList = Lists.newArrayList();
boolean atLeastOneCommitPointExists = false;
for (String name : blobs.keySet()) {
if (name.startsWith("commit-")) {
atLeastOneCommitPointExists = true;
try {
commitPointsList.add(CommitPoints.fromXContent(blobContainer.readBlobFully(name)));
} catch (Exception e) {
logger.warn("failed to read commit point [{}]", e, name);
}
}
}
if (atLeastOneCommitPointExists && commitPointsList.isEmpty()) {
// no commit point managed to load, bail so we won't corrupt the index, will require manual
// intervention
throw new IndexShardGatewayRecoveryException(
shardId, "Commit points exists but none could be loaded", null);
}
CommitPoints commitPoints = new CommitPoints(commitPointsList);
if (commitPoints.commits().isEmpty()) {
// no commit points, clean the store just so we won't recover wrong files
try {
indexShard.store().deleteContent();
} catch (IOException e) {
logger.warn("failed to clean store before starting shard", e);
}
recoveryStatus.index().startTime(System.currentTimeMillis());
recoveryStatus.index().time(System.currentTimeMillis() - recoveryStatus.index().startTime());
recoveryStatus.translog().startTime(System.currentTimeMillis());
recoveryStatus
.translog()
.time(System.currentTimeMillis() - recoveryStatus.index().startTime());
return;
}
for (CommitPoint commitPoint : commitPoints) {
if (!commitPointExistsInBlobs(commitPoint, blobs)) {
logger.warn(
"listed commit_point [{}]/[{}], but not all files exists, ignoring",
commitPoint.name(),
commitPoint.version());
continue;
}
try {
recoveryStatus.index().startTime(System.currentTimeMillis());
recoveryStatus.updateStage(RecoveryStatus.Stage.INDEX);
recoverIndex(commitPoint, blobs);
recoveryStatus
.index()
.time(System.currentTimeMillis() - recoveryStatus.index().startTime());
recoveryStatus.translog().startTime(System.currentTimeMillis());
recoveryStatus.updateStage(RecoveryStatus.Stage.TRANSLOG);
recoverTranslog(commitPoint, blobs);
recoveryStatus
.translog()
.time(System.currentTimeMillis() - recoveryStatus.index().startTime());
return;
} catch (Exception e) {
throw new IndexShardGatewayRecoveryException(
shardId,
"failed to recover commit_point ["
+ commitPoint.name()
+ "]/["
+ commitPoint.version()
+ "]",
e);
}
}
throw new IndexShardGatewayRecoveryException(
shardId, "No commit point data is available in gateway", null);
}
private void doSnapshot(final Snapshot snapshot) throws IndexShardGatewaySnapshotFailedException {
ImmutableMap<String, BlobMetaData> blobs;
try {
blobs = blobContainer.listBlobs();
} catch (IOException e) {
throw new IndexShardGatewaySnapshotFailedException(shardId, "failed to list blobs", e);
}
long generation = findLatestFileNameGeneration(blobs);
CommitPoints commitPoints = buildCommitPoints(blobs);
currentSnapshotStatus.index().startTime(System.currentTimeMillis());
currentSnapshotStatus.updateStage(SnapshotStatus.Stage.INDEX);
final SnapshotIndexCommit snapshotIndexCommit = snapshot.indexCommit();
final Translog.Snapshot translogSnapshot = snapshot.translogSnapshot();
final CountDownLatch indexLatch = new CountDownLatch(snapshotIndexCommit.getFiles().length);
final CopyOnWriteArrayList<Throwable> failures = new CopyOnWriteArrayList<Throwable>();
final List<CommitPoint.FileInfo> indexCommitPointFiles = Lists.newArrayList();
int indexNumberOfFiles = 0;
long indexTotalFilesSize = 0;
for (final String fileName : snapshotIndexCommit.getFiles()) {
StoreFileMetaData md;
try {
md = store.metaData(fileName);
} catch (IOException e) {
throw new IndexShardGatewaySnapshotFailedException(
shardId, "Failed to get store file metadata", e);
}
boolean snapshotRequired = false;
if (snapshot.indexChanged() && fileName.equals(snapshotIndexCommit.getSegmentsFileName())) {
snapshotRequired = true; // we want to always snapshot the segment file if the index changed
}
CommitPoint.FileInfo fileInfo = commitPoints.findPhysicalIndexFile(fileName);
if (fileInfo == null
|| !fileInfo.isSame(md)
|| !commitPointFileExistsInBlobs(fileInfo, blobs)) {
// commit point file does not exists in any commit point, or has different length, or does
// not fully exists in the listed blobs
snapshotRequired = true;
}
if (snapshotRequired) {
indexNumberOfFiles++;
indexTotalFilesSize += md.length();
// create a new FileInfo
try {
CommitPoint.FileInfo snapshotFileInfo =
new CommitPoint.FileInfo(
fileNameFromGeneration(++generation), fileName, md.length(), md.checksum());
indexCommitPointFiles.add(snapshotFileInfo);
snapshotFile(snapshotIndexCommit.getDirectory(), snapshotFileInfo, indexLatch, failures);
} catch (IOException e) {
failures.add(e);
indexLatch.countDown();
}
} else {
indexCommitPointFiles.add(fileInfo);
indexLatch.countDown();
}
}
currentSnapshotStatus.index().files(indexNumberOfFiles, indexTotalFilesSize);
try {
indexLatch.await();
} catch (InterruptedException e) {
failures.add(e);
}
if (!failures.isEmpty()) {
throw new IndexShardGatewaySnapshotFailedException(
shardId(), "Failed to perform snapshot (index files)", failures.get(failures.size() - 1));
}
currentSnapshotStatus
.index()
.time(System.currentTimeMillis() - currentSnapshotStatus.index().startTime());
currentSnapshotStatus.updateStage(SnapshotStatus.Stage.TRANSLOG);
currentSnapshotStatus.translog().startTime(System.currentTimeMillis());
// Note, we assume the snapshot is always started from "base 0". We need to seek forward if we
// want to lastTranslogPosition if we want the delta
List<CommitPoint.FileInfo> translogCommitPointFiles = Lists.newArrayList();
int expectedNumberOfOperations = 0;
boolean snapshotRequired = false;
if (snapshot.newTranslogCreated()) {
if (translogSnapshot.lengthInBytes() > 0) {
snapshotRequired = true;
expectedNumberOfOperations = translogSnapshot.estimatedTotalOperations();
}
} else {
// if we have a commit point, check that we have all the files listed in it in the blob store
if (!commitPoints.commits().isEmpty()) {
CommitPoint commitPoint = commitPoints.commits().get(0);
boolean allTranslogFilesExists = true;
for (CommitPoint.FileInfo fileInfo : commitPoint.translogFiles()) {
if (!commitPointFileExistsInBlobs(fileInfo, blobs)) {
allTranslogFilesExists = false;
break;
}
}
// if everything exists, we can seek forward in case there are new operations, otherwise, we
// copy over all again...
if (allTranslogFilesExists) {
translogCommitPointFiles.addAll(commitPoint.translogFiles());
if (snapshot.sameTranslogNewOperations()) {
translogSnapshot.seekForward(snapshot.lastTranslogLength());
if (translogSnapshot.lengthInBytes() > 0) {
snapshotRequired = true;
expectedNumberOfOperations =
translogSnapshot.estimatedTotalOperations()
- snapshot.lastTotalTranslogOperations();
}
} // else (no operations, nothing to snapshot)
} else {
// a full translog snapshot is required
if (translogSnapshot.lengthInBytes() > 0) {
expectedNumberOfOperations = translogSnapshot.estimatedTotalOperations();
snapshotRequired = true;
}
}
} else {
// no commit point, snapshot all the translog
if (translogSnapshot.lengthInBytes() > 0) {
expectedNumberOfOperations = translogSnapshot.estimatedTotalOperations();
snapshotRequired = true;
}
}
}
currentSnapshotStatus.translog().expectedNumberOfOperations(expectedNumberOfOperations);
if (snapshotRequired) {
CommitPoint.FileInfo addedTranslogFileInfo =
new CommitPoint.FileInfo(
fileNameFromGeneration(++generation),
"translog-" + translogSnapshot.translogId(),
translogSnapshot.lengthInBytes(),
null /* no need for checksum in translog */);
translogCommitPointFiles.add(addedTranslogFileInfo);
try {
snapshotTranslog(translogSnapshot, addedTranslogFileInfo);
} catch (Exception e) {
throw new IndexShardGatewaySnapshotFailedException(
shardId, "Failed to snapshot translog", e);
}
}
currentSnapshotStatus
.translog()
.time(System.currentTimeMillis() - currentSnapshotStatus.translog().startTime());
// now create and write the commit point
currentSnapshotStatus.updateStage(SnapshotStatus.Stage.FINALIZE);
long version = 0;
if (!commitPoints.commits().isEmpty()) {
version = commitPoints.commits().iterator().next().version() + 1;
}
String commitPointName = "commit-" + Long.toString(version, Character.MAX_RADIX);
CommitPoint commitPoint =
new CommitPoint(
version,
commitPointName,
CommitPoint.Type.GENERATED,
indexCommitPointFiles,
translogCommitPointFiles);
try {
byte[] commitPointData = CommitPoints.toXContent(commitPoint);
blobContainer.writeBlob(
commitPointName, new FastByteArrayInputStream(commitPointData), commitPointData.length);
} catch (Exception e) {
throw new IndexShardGatewaySnapshotFailedException(
shardId, "Failed to write commit point", e);
}
// delete all files that are not referenced by any commit point
// build a new CommitPoint, that includes this one and all the saved ones
List<CommitPoint> newCommitPointsList = Lists.newArrayList();
newCommitPointsList.add(commitPoint);
for (CommitPoint point : commitPoints) {
if (point.type() == CommitPoint.Type.SAVED) {
newCommitPointsList.add(point);
}
}
CommitPoints newCommitPoints = new CommitPoints(newCommitPointsList);
// first, go over and delete all the commit points
for (String blobName : blobs.keySet()) {
if (!blobName.startsWith("commit-")) {
continue;
}
long checkedVersion =
Long.parseLong(blobName.substring("commit-".length()), Character.MAX_RADIX);
if (!newCommitPoints.hasVersion(checkedVersion)) {
try {
blobContainer.deleteBlob(blobName);
} catch (IOException e) {
// ignore
}
}
}
// now go over all the blobs, and if they don't exists in a commit point, delete them
for (String blobName : blobs.keySet()) {
String name = blobName;
if (!name.startsWith("__")) {
continue;
}
if (blobName.contains(".part")) {
name = blobName.substring(0, blobName.indexOf(".part"));
}
if (newCommitPoints.findNameFile(name) == null) {
try {
blobContainer.deleteBlob(blobName);
} catch (IOException e) {
// ignore, will delete it laters
}
}
}
}