public static Pair<byte[], Integer> retrieveChunksAsBytes(
ZooKeeper zk, String path, String prefix, boolean getCRC) throws Exception {
TreeSet<String> chunks = new TreeSet<String>();
while (true) {
boolean allUploadsComplete = true;
if (!chunks.contains(path + "/" + prefix + "_complete")) {
allUploadsComplete = false;
}
if (allUploadsComplete) {
break;
}
chunks = new TreeSet<String>(zk.getChildren(path, false));
for (String chunk : chunks) {
for (int ii = 0; ii < chunks.size(); ii++) {
if (chunk.startsWith(path + "/" + prefix)) {
chunks.add(chunk);
}
}
}
}
byte resultBuffers[][] = new byte[chunks.size() - 1][];
int ii = 0;
PureJavaCrc32 crc = getCRC ? new PureJavaCrc32() : null;
for (String chunk : chunks) {
if (chunk.endsWith("_complete")) continue;
resultBuffers[ii] = zk.getData(chunk, false, null);
if (crc != null) {
crc.update(resultBuffers[ii]);
}
ii++;
}
return Pair.of(decompressBytes(resultBuffers), crc != null ? (int) crc.getValue() : null);
}
public DefaultSnapshotDataTarget(
final File file,
final int hostId,
final String clusterName,
final String databaseName,
final String tableName,
final int numPartitions,
final boolean isReplicated,
final List<Integer> partitionIds,
final VoltTable schemaTable,
final long txnId,
final long timestamp,
int version[])
throws IOException {
String hostname = CoreUtils.getHostnameOrAddress();
m_file = file;
m_tableName = tableName;
m_fos = new FileOutputStream(file);
m_channel = m_fos.getChannel();
m_needsFinalClose = !isReplicated;
final FastSerializer fs = new FastSerializer();
fs.writeInt(0); // CRC
fs.writeInt(0); // Header length placeholder
fs.writeByte(
1); // Indicate the snapshot was not completed, set to true for the CRC calculation, false
// later
for (int ii = 0; ii < 4; ii++) {
fs.writeInt(version[ii]); // version
}
JSONStringer stringer = new JSONStringer();
byte jsonBytes[] = null;
try {
stringer.object();
stringer.key("txnId").value(txnId);
stringer.key("hostId").value(hostId);
stringer.key("hostname").value(hostname);
stringer.key("clusterName").value(clusterName);
stringer.key("databaseName").value(databaseName);
stringer.key("tableName").value(tableName.toUpperCase());
stringer.key("isReplicated").value(isReplicated);
stringer.key("isCompressed").value(true);
stringer.key("checksumType").value("CRC32C");
stringer.key("timestamp").value(timestamp);
/*
* The timestamp string is for human consumption, automated stuff should use
* the actual timestamp
*/
stringer.key("timestampString").value(SnapshotUtil.formatHumanReadableDate(timestamp));
if (!isReplicated) {
stringer.key("partitionIds").array();
for (int partitionId : partitionIds) {
stringer.value(partitionId);
}
stringer.endArray();
stringer.key("numPartitions").value(numPartitions);
}
stringer.endObject();
String jsonString = stringer.toString();
JSONObject jsonObj = new JSONObject(jsonString);
jsonString = jsonObj.toString(4);
jsonBytes = jsonString.getBytes("UTF-8");
} catch (Exception e) {
throw new IOException(e);
}
fs.writeInt(jsonBytes.length);
fs.write(jsonBytes);
final BBContainer container = fs.getBBContainer();
container.b.position(4);
container.b.putInt(container.b.remaining() - 4);
container.b.position(0);
final byte schemaBytes[] = PrivateVoltTableFactory.getSchemaBytes(schemaTable);
final PureJavaCrc32 crc = new PureJavaCrc32();
ByteBuffer aggregateBuffer = ByteBuffer.allocate(container.b.remaining() + schemaBytes.length);
aggregateBuffer.put(container.b);
aggregateBuffer.put(schemaBytes);
aggregateBuffer.flip();
crc.update(aggregateBuffer.array(), 4, aggregateBuffer.capacity() - 4);
final int crcValue = (int) crc.getValue();
aggregateBuffer.putInt(crcValue).position(8);
aggregateBuffer.put((byte) 0).position(0); // Haven't actually finished writing file
if (m_simulateFullDiskWritingHeader) {
m_writeException = new IOException("Disk full");
m_writeFailed = true;
m_fos.close();
throw m_writeException;
}
/*
* Be completely sure the write succeeded. If it didn't
* the disk is probably full or the path is bunk etc.
*/
m_acceptOneWrite = true;
ListenableFuture<?> writeFuture =
write(Callables.returning((BBContainer) DBBPool.wrapBB(aggregateBuffer)), false);
try {
writeFuture.get();
} catch (InterruptedException e) {
m_fos.close();
throw new java.io.InterruptedIOException();
} catch (ExecutionException e) {
m_fos.close();
throw m_writeException;
}
if (m_writeFailed) {
m_fos.close();
throw m_writeException;
}
ScheduledFuture<?> syncTask = null;
syncTask =
m_syncService.scheduleAtFixedRate(
new Runnable() {
@Override
public void run() {
// Only sync for at least 4 megabyte of data, enough to amortize the cost of seeking
// on ye olden platters. Since we are appending to a file it's actually 2 seeks.
while (m_bytesWrittenSinceLastSync.get() > (1024 * 1024 * 4)) {
final int bytesSinceLastSync = m_bytesWrittenSinceLastSync.getAndSet(0);
try {
m_channel.force(false);
} catch (IOException e) {
if (!(e instanceof java.nio.channels.AsynchronousCloseException)) {
SNAP_LOG.error("Error syncing snapshot", e);
} else {
SNAP_LOG.debug(
"Asynchronous close syncing snasphot data, presumably graceful", e);
}
}
m_bytesAllowedBeforeSync.release(bytesSinceLastSync);
}
}
},
SNAPSHOT_SYNC_FREQUENCY,
SNAPSHOT_SYNC_FREQUENCY,
TimeUnit.MILLISECONDS);
m_syncTask = syncTask;
}
// XXX maybe consider an IOException subclass at some point
public TableSaveFile(
FileChannel dataIn,
int readAheadChunks,
Integer[] relevantPartitionIds,
boolean continueOnCorruptedChunk)
throws IOException {
try {
EELibraryLoader.loadExecutionEngineLibrary(true);
if (relevantPartitionIds == null) {
m_relevantPartitionIds = null;
} else {
m_relevantPartitionIds = new HashSet<Integer>();
for (Integer i : relevantPartitionIds) {
m_relevantPartitionIds.add(i);
}
}
m_chunkReads = new Semaphore(readAheadChunks);
m_saveFile = dataIn;
m_continueOnCorruptedChunk = continueOnCorruptedChunk;
final PureJavaCrc32 crc = new PureJavaCrc32();
/*
* If the CRC check fails because the file wasn't completed
*/
final PureJavaCrc32 secondCRC = new PureJavaCrc32();
/*
* Get the header with the save restore specific information
*/
final ByteBuffer lengthBuffer = ByteBuffer.allocate(8);
while (lengthBuffer.hasRemaining()) {
final int read = m_saveFile.read(lengthBuffer);
if (read == -1) {
throw new EOFException();
}
}
lengthBuffer.flip();
final int originalCRC = lengthBuffer.getInt();
int length = lengthBuffer.getInt();
crc.update(lengthBuffer.array(), 4, 4);
secondCRC.update(lengthBuffer.array(), 4, 4);
if (length < 0) {
throw new IOException("Corrupted save file has negative header length");
}
if (length > 2097152) {
throw new IOException("Corrupted save file has unreasonable header length > 2 megs");
}
final ByteBuffer saveRestoreHeader = ByteBuffer.allocate(length);
while (saveRestoreHeader.hasRemaining()) {
final int read = m_saveFile.read(saveRestoreHeader);
if (read == -1 || read < length) {
throw new EOFException();
}
}
saveRestoreHeader.flip();
crc.update(saveRestoreHeader.array());
secondCRC.update(new byte[] {1});
secondCRC.update(saveRestoreHeader.array(), 1, saveRestoreHeader.array().length - 1);
/*
* Get the template for the VoltTable serialization header.
* It will have an extra length value preceded to it so that
* it can be sucked straight into a buffer. This will not
* contain a row count since that varies from chunk to chunk
* and is supplied by the chunk
*/
lengthBuffer.clear();
lengthBuffer.limit(4);
/*
* Why this stupidity and no while loop?
* Because java is broken and complains about a random final
* elsewhere if you do.
*/
{
final int read = m_saveFile.read(lengthBuffer);
if (read == -1) {
throw new EOFException();
}
}
crc.update(lengthBuffer.array(), 0, 4);
secondCRC.update(lengthBuffer.array(), 0, 4);
lengthBuffer.flip();
length = lengthBuffer.getInt();
if (length < 4) {
throw new IOException(
"Corrupted save file has negative length or too small length for VoltTable header");
}
if (length > 2097152) {
throw new IOException(
"Corrupted save file has unreasonable VoltTable header length > 2 megs");
}
m_tableHeader = ByteBuffer.allocate(length + 4);
m_tableHeader.putInt(length);
while (m_tableHeader.hasRemaining()) {
final int read = m_saveFile.read(m_tableHeader);
if (read == -1) {
throw new EOFException();
}
}
crc.update(m_tableHeader.array(), 4, length);
secondCRC.update(m_tableHeader.array(), 4, length);
boolean failedCRCDueToNotCompleted = false;
final int actualCRC = (int) crc.getValue();
if (originalCRC != actualCRC) {
/*
* Check if the CRC mismatch is due to the snapshot not being completed
*/
final int secondCRCValue = (int) secondCRC.getValue();
if (secondCRCValue == originalCRC) {
failedCRCDueToNotCompleted = true;
} else {
throw new IOException("Checksum mismatch");
}
}
FastDeserializer fd = new FastDeserializer(saveRestoreHeader);
byte completedByte = fd.readByte();
m_completed = failedCRCDueToNotCompleted ? false : (completedByte == 1 ? true : false);
for (int ii = 0; ii < 4; ii++) {
m_versionNum[ii] = fd.readInt();
}
/*
* Support the original pre 1.3 header format as well as a new JSON format.
* JSON will make it possible to add info to a snapshot header without
* breaking backwards compatibility.
*/
if (m_versionNum[3] == 0) {
m_txnId = fd.readLong();
m_timestamp = TransactionIdManager.getTimestampFromTransactionId(m_txnId);
m_hostId = fd.readInt();
m_hostname = fd.readString();
m_clusterName = fd.readString();
m_databaseName = fd.readString();
m_tableName = fd.readString();
m_isReplicated = fd.readBoolean();
m_isCompressed = false;
m_checksumType = ChecksumType.CRC32;
if (!m_isReplicated) {
m_partitionIds = (int[]) fd.readArray(int.class);
if (!m_completed) {
for (Integer partitionId : m_partitionIds) {
m_corruptedPartitions.add(partitionId);
}
}
m_totalPartitions = fd.readInt();
} else {
m_partitionIds = new int[] {0};
m_totalPartitions = 1;
if (!m_completed) {
m_corruptedPartitions.add(0);
}
}
m_hasVersion2FormatChunks = false;
} else {
assert (m_versionNum[3] == 1 || m_versionNum[3] == 2);
if (m_versionNum[3] >= 2) {
m_hasVersion2FormatChunks = true;
} else {
m_hasVersion2FormatChunks = false;
}
int numJSONBytes = fd.readInt();
byte jsonBytes[] = new byte[numJSONBytes];
fd.readFully(jsonBytes);
String jsonString = new String(jsonBytes, "UTF-8");
JSONObject obj = new JSONObject(jsonString);
m_txnId = obj.getLong("txnId");
// Timestamp field added for 3.0, might not be there
if (obj.has("timestamp")) {
m_timestamp = obj.getLong("timestamp");
} else {
// Pre 3.0/IV2 the timestamp was in the transactionid
m_timestamp = TransactionIdManager.getTimestampFromTransactionId(m_txnId);
}
m_hostId = obj.getInt("hostId");
m_hostname = obj.getString("hostname");
m_clusterName = obj.getString("clusterName");
m_databaseName = obj.getString("databaseName");
m_tableName = obj.getString("tableName");
m_isReplicated = obj.getBoolean("isReplicated");
m_isCompressed = obj.optBoolean("isCompressed", false);
m_checksumType = ChecksumType.valueOf(obj.optString("checksumType", "CRC32"));
if (!m_isReplicated) {
JSONArray partitionIds = obj.getJSONArray("partitionIds");
m_partitionIds = new int[partitionIds.length()];
for (int ii = 0; ii < m_partitionIds.length; ii++) {
m_partitionIds[ii] = partitionIds.getInt(ii);
}
if (!m_completed) {
for (Integer partitionId : m_partitionIds) {
m_corruptedPartitions.add(partitionId);
}
}
m_totalPartitions = obj.getInt("numPartitions");
} else {
m_partitionIds = new int[] {0};
m_totalPartitions = 1;
if (!m_completed) {
m_corruptedPartitions.add(0);
}
}
}
/*
* Several runtime exceptions can be thrown in valid failure cases where
* a corrupt save file is being detected.
*/
} catch (BufferUnderflowException e) {
throw new IOException(e);
} catch (BufferOverflowException e) {
throw new IOException(e);
} catch (IndexOutOfBoundsException e) {
throw new IOException(e);
} catch (JSONException e) {
throw new IOException(e);
}
}
