@Override
public void handleResponse(ClientResponse resp) {
if (resp == null) {
VoltDB.crashLocalVoltDB(
"Received a null response to a snapshot initiation request. "
+ "This should be impossible.",
true,
null);
} else if (resp.getStatus() != ClientResponse.SUCCESS) {
tmLog.info(
"Failed to complete partition detection snapshot, status: "
+ resp.getStatus()
+ ", reason: "
+ resp.getStatusString());
tmLog.info("Retrying partition detection snapshot...");
SnapshotUtil.requestSnapshot(
0L,
m_partSnapshotSchedule.getPath(),
m_partSnapshotSchedule.getPrefix() + System.currentTimeMillis(),
true,
SnapshotFormat.NATIVE,
null,
m_snapshotHandler,
true);
} else if (!SnapshotUtil.didSnapshotRequestSucceed(resp.getResults())) {
VoltDB.crashGlobalVoltDB(
"Unable to complete partition detection snapshot: " + resp.getResults()[0],
false,
null);
} else {
VoltDB.crashGlobalVoltDB(
"Partition detection snapshot completed. Shutting down.", false, null);
}
}
private void doPartitionDetectionActivities() {
// We should never re-enter here once we've decided we're partitioned and doomed
assert (!m_partitionDetected);
// After everything is resolved, write the new surviving set to ZK
List<Integer> currentNodes = null;
try {
currentNodes = m_hostMessenger.getLiveHostIds();
} catch (Exception e) {
}
Set<Integer> currentHosts = new HashSet<Integer>(currentNodes);
Set<Integer> previousHosts = readPriorKnownLiveNodes();
boolean partitionDetectionTriggered = makePPDDecision(previousHosts, currentHosts);
if (partitionDetectionTriggered) {
m_partitionDetected = true;
if (m_usingCommandLog) {
// Just shut down immediately
VoltDB.crashGlobalVoltDB(
"Use of command logging detected, no additional database snapshot will "
+ "be generated. Please use the 'recover' action to restore the database if necessary.",
false,
null);
} else {
SnapshotUtil.requestSnapshot(
0L,
m_partSnapshotSchedule.getPath(),
m_partSnapshotSchedule.getPrefix() + System.currentTimeMillis(),
true,
SnapshotFormat.NATIVE,
null,
m_snapshotHandler,
true);
}
}
// If the cluster host set has changed, then write the new set to ZK
// NOTE: we don't want to update the known live nodes if we've decided that our subcluster is
// dying, otherwise a poorly timed subsequent failure might reverse this decision. Any future
// promoted
// LeaderAppointer should make their partition detection decision based on the pre-partition
// cluster state.
else if (!currentHosts.equals(previousHosts)) {
writeKnownLiveNodes(currentNodes);
}
}
private void createSetupIv2(
String file_path,
final String pathType,
final String file_nonce,
SnapshotFormat format,
final long txnId,
final Map<Integer, Long> partitionTransactionIds,
JSONObject jsData,
final SystemProcedureExecutionContext context,
final VoltTable result,
ExtensibleSnapshotDigestData extraSnapshotData,
SiteTracker tracker,
HashinatorSnapshotData hashinatorData,
long timestamp) {
SnapshotWritePlan plan;
if (format == SnapshotFormat.NATIVE) {
plan = new NativeSnapshotWritePlan();
} else if (format == SnapshotFormat.CSV) {
plan = new CSVSnapshotWritePlan();
} else if (format == SnapshotFormat.STREAM) {
plan = new StreamSnapshotWritePlan();
} else if (format == SnapshotFormat.INDEX) {
plan = new IndexSnapshotWritePlan();
} else {
throw new RuntimeException("BAD BAD BAD");
}
file_path = SnapshotUtil.getRealPath(SnapshotPathType.valueOf(pathType), file_path);
final Callable<Boolean> deferredSetup =
plan.createSetup(
file_path,
pathType,
file_nonce,
txnId,
partitionTransactionIds,
jsData,
context,
result,
extraSnapshotData,
tracker,
hashinatorData,
timestamp);
m_deferredSetupFuture =
VoltDB.instance()
.submitSnapshotIOWork(
new DeferredSnapshotSetup(plan, deferredSetup, txnId, partitionTransactionIds));
synchronized (m_createLock) {
// Seems like this should be cleared out just in case
// Log if there is actually anything to clear since it is unexpected
if (!m_taskListsForHSIds.isEmpty()) {
SNAP_LOG.warn("Found lingering snapshot tasks while setting up a snapshot");
}
m_taskListsForHSIds.clear();
m_createSuccess.set(true);
m_createResult.set(result);
m_taskListsForHSIds.putAll(plan.getTaskListsForHSIds());
// HACK HACK HACK. If the task list is empty, this host has no work to do for
// this snapshot. We're going to create an empty list of tasks for one of the sites to do
// so that we'll have a SnapshotSiteProcessor which will do the logSnapshotCompleteToZK.
if (m_taskListsForHSIds.isEmpty()) {
SNAP_LOG.debug(
"Node had no snapshot work to do. Creating a null task to drive completion.");
m_taskListsForHSIds.put(context.getSiteId(), new ArrayDeque<SnapshotTableTask>());
}
SNAP_LOG.debug(
"Planned tasks: "
+ CoreUtils.hsIdCollectionToString(plan.getTaskListsForHSIds().keySet()));
SNAP_LOG.debug(
"Created tasks for HSIds: "
+ CoreUtils.hsIdCollectionToString(m_taskListsForHSIds.keySet()));
}
}
/**
* The only public method: do all the work to start a snapshot. Assumes that a snapshot is
* feasible, that the caller has validated it can be accomplished, that the caller knows this is a
* consistent or useful transaction point at which to snapshot.
*
* @param file_path
* @param file_nonce
* @param format
* @param block
* @param txnId
* @param data
* @param context
* @param hostname
* @return VoltTable describing the results of the snapshot attempt
*/
public VoltTable startSnapshotting(
final String file_path,
final String file_nonce,
final SnapshotFormat format,
final byte block,
final long multiPartTxnId,
final long partitionTxnId,
final long legacyPerPartitionTxnIds[],
final String data,
final SystemProcedureExecutionContext context,
final String hostname,
final HashinatorSnapshotData hashinatorData,
final long timestamp) {
TRACE_LOG.trace("Creating snapshot target and handing to EEs");
final VoltTable result = SnapshotUtil.constructNodeResultsTable();
final int numLocalSites =
context.getCluster().getDeployment().get("deployment").getSitesperhost();
// One site wins the race to create the snapshot targets, populating
// m_taskListsForSites for the other sites and creating an appropriate
// number of snapshot permits.
synchronized (SnapshotSiteProcessor.m_snapshotCreateLock) {
SnapshotSiteProcessor.m_snapshotCreateSetupBarrierActualAction.set(
new Runnable() {
@Override
public void run() {
Map<Integer, Long> partitionTransactionIds = new HashMap<Integer, Long>();
partitionTransactionIds = m_partitionLastSeenTransactionIds;
SNAP_LOG.debug("Last seen partition transaction ids " + partitionTransactionIds);
m_partitionLastSeenTransactionIds = new HashMap<Integer, Long>();
partitionTransactionIds.put(TxnEgo.getPartitionId(multiPartTxnId), multiPartTxnId);
/*
* Do a quick sanity check that the provided IDs
* don't conflict with currently active partitions. If they do
* it isn't fatal we can just skip it.
*/
for (long txnId : legacyPerPartitionTxnIds) {
final int legacyPartition = TxnEgo.getPartitionId(txnId);
if (partitionTransactionIds.containsKey(legacyPartition)) {
SNAP_LOG.warn(
"While saving a snapshot and propagating legacy "
+ "transaction ids found an id that matches currently active partition"
+ partitionTransactionIds.get(legacyPartition));
} else {
partitionTransactionIds.put(legacyPartition, txnId);
}
}
exportSequenceNumbers = SnapshotSiteProcessor.getExportSequenceNumbers();
createSetupIv2(
file_path,
file_nonce,
format,
multiPartTxnId,
partitionTransactionIds,
data,
context,
result,
exportSequenceNumbers,
context.getSiteTrackerForSnapshot(),
hashinatorData,
timestamp);
}
});
// Create a barrier to use with the current number of sites to wait for
// or if the barrier is already set up check if it is broken and reset if necessary
SnapshotSiteProcessor.readySnapshotSetupBarriers(numLocalSites);
// From within this EE, record the sequence numbers as of the start of the snapshot (now)
// so that the info can be put in the digest.
SnapshotSiteProcessor.populateExportSequenceNumbersForExecutionSite(context);
SNAP_LOG.debug(
"Registering transaction id "
+ partitionTxnId
+ " for "
+ TxnEgo.getPartitionId(partitionTxnId));
m_partitionLastSeenTransactionIds.put(TxnEgo.getPartitionId(partitionTxnId), partitionTxnId);
}
boolean runPostTasks = false;
VoltTable earlyResultTable = null;
try {
SnapshotSiteProcessor.m_snapshotCreateSetupBarrier.await();
try {
synchronized (m_createLock) {
SNAP_LOG.debug(
"Found tasks for HSIds: "
+ CoreUtils.hsIdCollectionToString(m_taskListsForHSIds.keySet()));
SNAP_LOG.debug("Looking for local HSID: " + CoreUtils.hsIdToString(context.getSiteId()));
Deque<SnapshotTableTask> taskList = m_taskListsForHSIds.remove(context.getSiteId());
// If createSetup failed, then the first site to reach here is going
// to send the results table generated by createSetup, and then empty out the table.
// All other sites to reach here will send the appropriate empty table.
// If createSetup was a success but the taskList is null, then we'll use the block
// switch to figure out what flavor of empty SnapshotSave result table to return.
if (!m_createSuccess.get()) {
// There shouldn't be any work for any site if we failed
assert (m_taskListsForHSIds.isEmpty());
VoltTable finalresult = m_createResult.get();
if (finalresult != null) {
m_createResult.set(null);
earlyResultTable = finalresult;
} else {
// We returned a non-empty NodeResultsTable with the failures in it,
// every other site needs to return a NodeResultsTable as well.
earlyResultTable = SnapshotUtil.constructNodeResultsTable();
}
} else if (taskList == null) {
SNAP_LOG.debug("No task for this site, block " + block);
// This node is participating in the snapshot but this site has nothing to do.
// Send back an appropriate empty table based on the block flag
if (block != 0) {
runPostTasks = true;
earlyResultTable = SnapshotUtil.constructPartitionResultsTable();
earlyResultTable.addRow(
context.getHostId(),
hostname,
CoreUtils.getSiteIdFromHSId(context.getSiteId()),
"SUCCESS",
"");
} else {
earlyResultTable = SnapshotUtil.constructNodeResultsTable();
}
} else {
context
.getSiteSnapshotConnection()
.initiateSnapshots(format, taskList, multiPartTxnId, exportSequenceNumbers);
}
if (m_deferredSetupFuture != null) {
// Add a listener to the deferred setup so that it can kick off the snapshot
// task once the setup is done.
m_deferredSetupFuture.addListener(
new Runnable() {
@Override
public void run() {
DeferredSnapshotSetup deferredSnapshotSetup = null;
try {
deferredSnapshotSetup = m_deferredSetupFuture.get();
} catch (Exception e) {
// it doesn't throw
}
assert deferredSnapshotSetup != null;
context
.getSiteSnapshotConnection()
.startSnapshotWithTargets(
deferredSnapshotSetup.getPlan().getSnapshotDataTargets());
}
},
CoreUtils.SAMETHREADEXECUTOR);
}
}
} finally {
SnapshotSiteProcessor.m_snapshotCreateFinishBarrier.await(120, TimeUnit.SECONDS);
}
} catch (TimeoutException e) {
VoltDB.crashLocalVoltDB(
"Timed out waiting 120 seconds for all threads to arrive and start snapshot", true, null);
} catch (InterruptedException e) {
result.addRow(context.getHostId(), hostname, "", "FAILURE", CoreUtils.throwableToString(e));
earlyResultTable = result;
} catch (BrokenBarrierException e) {
result.addRow(context.getHostId(), hostname, "", "FAILURE", CoreUtils.throwableToString(e));
earlyResultTable = result;
}
// If earlyResultTable is set, return here
if (earlyResultTable != null) {
if (runPostTasks) {
// Need to run post-snapshot tasks before finishing
SnapshotSiteProcessor.runPostSnapshotTasks(context);
}
return earlyResultTable;
}
if (block != 0) {
HashSet<Exception> failures = Sets.newHashSet();
String status = "SUCCESS";
String err = "";
try {
// For blocking snapshot, propogate the error from deferred setup back to the client
final DeferredSnapshotSetup deferredSnapshotSetup = m_deferredSetupFuture.get();
if (deferredSnapshotSetup != null && deferredSnapshotSetup.getError() != null) {
status = "FAILURE";
err = deferredSnapshotSetup.getError().toString();
failures.add(deferredSnapshotSetup.getError());
}
failures.addAll(context.getSiteSnapshotConnection().completeSnapshotWork());
SnapshotSiteProcessor.runPostSnapshotTasks(context);
} catch (Exception e) {
status = "FAILURE";
err = e.toString();
failures.add(e);
}
final VoltTable blockingResult = SnapshotUtil.constructPartitionResultsTable();
if (failures.isEmpty()) {
blockingResult.addRow(
context.getHostId(),
hostname,
CoreUtils.getSiteIdFromHSId(context.getSiteId()),
status,
err);
} else {
status = "FAILURE";
for (Exception e : failures) {
err = e.toString();
}
blockingResult.addRow(
context.getHostId(),
hostname,
CoreUtils.getSiteIdFromHSId(context.getSiteId()),
status,
err);
}
return blockingResult;
}
return result;
}
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;
}
