public static void logFragmentTaskMessage(
FragmentTaskMessage ftask, long localHSId, long spHandle, boolean borrow) {
if (iv2log.isTraceEnabled()) {
String label = "rxFragMsg";
if (borrow) {
label = "rxBrrwMsg";
}
if (ftask.getSpHandle() != Long.MIN_VALUE && ftask.getSpHandle() != spHandle) {
iv2log.error(
"FragmentTaskMessage SP HANDLE conflict. Message: "
+ ftask.getSpHandle()
+ ", locally held: "
+ spHandle);
}
String logmsg = "%s %s from %s txnId %s spHandle %s trunc %s";
iv2log.trace(
String.format(
logmsg,
label,
CoreUtils.hsIdToString(localHSId),
CoreUtils.hsIdToString(ftask.m_sourceHSId),
txnIdToString(ftask.getTxnId()),
txnIdToString(spHandle),
txnIdToString(ftask.getTruncationHandle())));
}
}
@Override
public void run() {
initialize();
try {
while (m_shouldContinue) {
// Normal operation blocks the site thread on the sitetasker queue.
SiteTasker task = m_scheduler.take();
task.run(getSiteProcedureConnection());
}
} catch (OutOfMemoryError e) {
// Even though OOM should be caught by the Throwable section below,
// it sadly needs to be handled seperately. The goal here is to make
// sure VoltDB crashes.
String errmsg =
"Site: "
+ org.voltcore.utils.CoreUtils.hsIdToString(m_siteId)
+ " ran out of Java memory. "
+ "This node will shut down.";
VoltDB.crashLocalVoltDB(errmsg, true, e);
} catch (Throwable t) {
String errmsg =
"Site: "
+ org.voltcore.utils.CoreUtils.hsIdToString(m_siteId)
+ " encountered an "
+ "unexpected error and will die, taking this VoltDB node down.";
VoltDB.crashLocalVoltDB(errmsg, true, t);
}
shutdown();
}
public void handleDumpMessage() {
String who = CoreUtils.hsIdToString(m_mailbox.getHSId());
hostLog.warn("State dump for site: " + who);
hostLog.warn("" + who + ": partition: " + m_partitionId + ", isLeader: " + m_isLeader);
if (m_isLeader) {
hostLog.warn("" + who + ": replicas: " + CoreUtils.hsIdCollectionToString(m_replicaHSIds));
if (m_sendToHSIds.length > 0) {
m_mailbox.send(m_sendToHSIds, new DumpMessage());
}
}
hostLog.warn(
""
+ who
+ ": most recent SP handle: "
+ getCurrentTxnId()
+ " "
+ TxnEgo.txnIdToString(getCurrentTxnId()));
hostLog.warn(
""
+ who
+ ": outstanding txns: "
+ m_outstandingTxns.keySet()
+ " "
+ TxnEgo.txnIdCollectionToString(m_outstandingTxns.keySet()));
hostLog.warn("" + who + ": TransactionTaskQueue: " + m_pendingTasks.toString());
if (m_duplicateCounters.size() > 0) {
hostLog.warn("" + who + ": duplicate counters: ");
for (Entry<DuplicateCounterKey, DuplicateCounter> e : m_duplicateCounters.entrySet()) {
hostLog.warn("\t" + who + ": " + e.getKey().toString() + ": " + e.getValue().toString());
}
}
}
public static void logIv2InitiateTaskMessage(
Iv2InitiateTaskMessage itask, long localHSId, long txnid, long spHandle) {
if (iv2log.isTraceEnabled()) {
String logmsg = "rxInitMsg %s from %s ciHandle %s txnId %s spHandle %s trunc %s";
if (itask.getTxnId() != Long.MIN_VALUE && itask.getTxnId() != txnid) {
iv2log.error(
"Iv2InitiateTaskMessage TXN ID conflict. Message: "
+ itask.getTxnId()
+ ", locally held: "
+ txnid);
}
if (itask.getSpHandle() != Long.MIN_VALUE && itask.getSpHandle() != spHandle) {
iv2log.error(
"Iv2InitiateTaskMessage SP HANDLE conflict. Message: "
+ itask.getSpHandle()
+ ", locally held: "
+ spHandle);
}
iv2log.trace(
String.format(
logmsg,
CoreUtils.hsIdToString(localHSId),
CoreUtils.hsIdToString(itask.m_sourceHSId),
ClientInterfaceHandleManager.handleToString(itask.getClientInterfaceHandle()),
txnIdToString(txnid),
txnIdToString(spHandle),
txnIdToString(itask.getTruncationHandle())));
}
}
public static void logInitiatorRxMsg(VoltMessage msg, long localHSId) {
if (iv2log.isTraceEnabled()) {
if (msg instanceof InitiateResponseMessage) {
InitiateResponseMessage iresp = (InitiateResponseMessage) msg;
String logmsg = "rxInitRsp %s from %s ciHandle %s txnId %s spHandle %s status %s";
iv2log.trace(
String.format(
logmsg,
CoreUtils.hsIdToString(localHSId),
CoreUtils.hsIdToString(iresp.m_sourceHSId),
ClientInterfaceHandleManager.handleToString(iresp.getClientInterfaceHandle()),
txnIdToString(iresp.getTxnId()),
txnIdToString(iresp.getSpHandle()),
respStatusToString(iresp.getClientResponseData().getStatus())));
} else if (msg instanceof FragmentResponseMessage) {
FragmentResponseMessage fresp = (FragmentResponseMessage) msg;
String logmsg = "rxFragRsp %s from %s txnId %s spHandle %s status %s";
iv2log.trace(
String.format(
logmsg,
CoreUtils.hsIdToString(localHSId),
CoreUtils.hsIdToString(fresp.m_sourceHSId),
txnIdToString(fresp.getTxnId()),
txnIdToString(fresp.getSpHandle()),
fragStatusToString(fresp.getStatusCode())));
}
}
}
/** Runs when the RejoinCoordinator decides this site should start rejoin. */
void doInitiation(RejoinMessage message) {
m_coordinatorHsId = message.m_sourceHSId;
m_streamSnapshotMb = VoltDB.instance().getHostMessenger().createMailbox();
m_rejoinSiteProcessor = new StreamSnapshotSink(m_streamSnapshotMb);
// MUST choose the leader as the source.
long sourceSite = m_mailbox.getMasterHsId(m_partitionId);
long hsId =
m_rejoinSiteProcessor.initialize(
message.getSnapshotSourceCount(), message.getSnapshotBufferPool());
REJOINLOG.debug(
m_whoami
+ "received INITIATION message. Doing rejoin"
+ ". Source site is: "
+ CoreUtils.hsIdToString(sourceSite)
+ " and destination rejoin processor is: "
+ CoreUtils.hsIdToString(hsId)
+ " and snapshot nonce is: "
+ message.getSnapshotNonce());
registerSnapshotMonitor(message.getSnapshotNonce());
// Tell the RejoinCoordinator everything it will need to know to get us our snapshot stream.
RejoinMessage initResp = new RejoinMessage(m_mailbox.getHSId(), sourceSite, hsId);
m_mailbox.send(m_coordinatorHsId, initResp);
// Start waiting for snapshot data
m_taskQueue.offer(this);
}
public static void logTopology(long leaderHSId, List<Long> replicas, int partitionId) {
if (iv2log.isTraceEnabled()) {
String logmsg = "topology partition %d leader %s replicas (%s)";
iv2log.trace(
String.format(
logmsg,
partitionId,
CoreUtils.hsIdToString(leaderHSId),
CoreUtils.hsIdCollectionToString(replicas)));
}
}
/** Create a native VoltDB execution engine */
ExecutionEngine initializeEE(String serializedCatalog, final long timestamp) {
String hostname = CoreUtils.getHostnameOrAddress();
ExecutionEngine eeTemp = null;
try {
if (m_backend == BackendTarget.NATIVE_EE_JNI) {
eeTemp =
new ExecutionEngineJNI(
m_context.cluster.getRelativeIndex(),
m_siteId,
m_partitionId,
CoreUtils.getHostIdFromHSId(m_siteId),
hostname,
m_context
.cluster
.getDeployment()
.get("deployment")
.getSystemsettings()
.get("systemsettings")
.getMaxtemptablesize(),
m_numberOfPartitions);
eeTemp.loadCatalog(timestamp, serializedCatalog);
} else {
// set up the EE over IPC
eeTemp =
new ExecutionEngineIPC(
m_context.cluster.getRelativeIndex(),
m_siteId,
m_partitionId,
CoreUtils.getHostIdFromHSId(m_siteId),
hostname,
m_context
.cluster
.getDeployment()
.get("deployment")
.getSystemsettings()
.get("systemsettings")
.getMaxtemptablesize(),
m_backend,
VoltDB.instance().getConfig().m_ipcPorts.remove(0),
m_numberOfPartitions);
eeTemp.loadCatalog(timestamp, serializedCatalog);
}
}
// just print error info an bail if we run into an error here
catch (final Exception ex) {
hostLog.l7dlog(
Level.FATAL,
LogKeys.host_ExecutionSite_FailedConstruction.name(),
new Object[] {m_siteId, m_siteIndex},
ex);
VoltDB.crashLocalVoltDB(ex.getMessage(), true, ex);
}
return eeTemp;
}
public static void logFinishTransaction(InitiateResponseMessage msg, long localHSId) {
if (iv2log.isTraceEnabled()) {
String logmsg = "finishTxn %s ciHandle %s initHSId %s status %s";
iv2log.trace(
String.format(
logmsg,
CoreUtils.hsIdToString(localHSId),
ClientInterfaceHandleManager.handleToString(msg.getClientInterfaceHandle()),
CoreUtils.hsIdToString(msg.getCoordinatorHSId()),
respStatusToString(msg.getClientResponseData().getStatus())));
}
}
/** Process a new repair log response */
@Override
public void deliver(VoltMessage message) {
if (message instanceof Iv2RepairLogResponseMessage) {
Iv2RepairLogResponseMessage response = (Iv2RepairLogResponseMessage) message;
if (response.getRequestId() != m_requestId) {
tmLog.debug(
m_whoami
+ "rejecting stale repair response."
+ " Current request id is: "
+ m_requestId
+ " Received response for request id: "
+ response.getRequestId());
return;
}
ReplicaRepairStruct rrs = m_replicaRepairStructs.get(response.m_sourceHSId);
if (rrs.m_expectedResponses < 0) {
tmLog.debug(
m_whoami
+ "collecting "
+ response.getOfTotal()
+ " repair log entries from "
+ CoreUtils.hsIdToString(response.m_sourceHSId));
}
// Long.MAX_VALUE has rejoin semantics
if (response.getHandle() != Long.MAX_VALUE) {
m_maxSeenTxnId = Math.max(m_maxSeenTxnId, response.getHandle());
}
if (response.getPayload() != null) {
m_repairLogUnion.add(response);
if (tmLog.isTraceEnabled()) {
tmLog.trace(
m_whoami
+ " collected from "
+ CoreUtils.hsIdToString(response.m_sourceHSId)
+ ", message: "
+ response.getPayload());
}
}
if (rrs.update(response)) {
tmLog.debug(
m_whoami
+ "collected "
+ rrs.m_receivedResponses
+ " responses for "
+ rrs.m_expectedResponses
+ " repair log entries from "
+ CoreUtils.hsIdToString(response.m_sourceHSId));
if (areRepairLogsComplete()) {
repairSurvivors();
}
}
}
}
public static void logIv2MultipartSentinel(
MultiPartitionParticipantMessage message, long localHSId, long txnId) {
if (iv2log.isTraceEnabled()) {
String logmsg = "rxSntlMsg %s from %s txnId %s";
iv2log.trace(
String.format(
logmsg,
CoreUtils.hsIdToString(localHSId),
CoreUtils.hsIdToString(message.m_sourceHSId),
txnIdToString(txnId)));
}
}
public static void logCreateTransaction(Iv2InitiateTaskMessage msg) {
if (iv2log.isTraceEnabled()) {
String logmsg = "createTxn %s ciHandle %s initHSId %s proc %s";
iv2log.trace(
String.format(
logmsg,
CoreUtils.hsIdToString(msg.getInitiatorHSId()),
ClientInterfaceHandleManager.handleToString(msg.getClientInterfaceHandle()),
CoreUtils.hsIdToString(msg.getCoordinatorHSId()),
msg.getStoredProcedureInvocation().getProcName()));
}
}
public long getBuddySiteForMPI(long hsid) {
int host = CoreUtils.getHostIdFromHSId(hsid);
// We'll be lazy and get the map we'd feed to SiteTracker's
// constructor, then go looking for a matching host ID.
List<MailboxNodeContent> sitesList = getMailboxNodeContentList();
for (MailboxNodeContent site : sitesList) {
if (site.partitionId != MpInitiator.MP_INIT_PID
&& host == CoreUtils.getHostIdFromHSId(site.HSId)) {
return site.HSId;
}
}
throw new RuntimeException(
"Unable to find a buddy initiator for MPI with HSID: " + CoreUtils.hsIdToString(hsid));
}
public static void logCompleteTransactionMessage(
CompleteTransactionMessage ctask, long localHSId) {
if (iv2log.isTraceEnabled()) {
String logmsg = "rxCompMsg %s from %s txnId %s %s %s";
iv2log.trace(
String.format(
logmsg,
CoreUtils.hsIdToString(localHSId),
CoreUtils.hsIdToString(ctask.m_sourceHSId),
txnIdToString(ctask.getTxnId()),
ctask.isRollback() ? "ROLLBACK" : "COMMIT",
ctask.isRestart() ? "RESTART" : ""));
}
}
private long assignLeader(int partitionId, List<Long> children) {
// We used masterHostId = -1 as a way to force the leader choice to be
// the first replica in the list, if we don't have some other mechanism
// which has successfully overridden it.
int masterHostId = -1;
if (m_state.get() == AppointerState.CLUSTER_START) {
try {
// find master in topo
JSONArray parts = m_topo.getJSONArray("partitions");
for (int p = 0; p < parts.length(); p++) {
JSONObject aPartition = parts.getJSONObject(p);
int pid = aPartition.getInt("partition_id");
if (pid == partitionId) {
masterHostId = aPartition.getInt("master");
}
}
} catch (JSONException jse) {
tmLog.error("Failed to find master for partition " + partitionId + ", defaulting to 0");
jse.printStackTrace();
masterHostId = -1; // stupid default
}
} else {
// For now, if we're appointing a new leader as a result of a
// failure, just pick the first replica in the children list.
// Could eventually do something more complex here to try to keep a
// semi-balance, but it's unclear that this has much utility until
// we add rebalancing on rejoin as well.
masterHostId = -1;
}
long masterHSId = children.get(0);
for (Long child : children) {
if (CoreUtils.getHostIdFromHSId(child) == masterHostId) {
masterHSId = child;
break;
}
}
tmLog.info(
"Appointing HSId "
+ CoreUtils.hsIdToString(masterHSId)
+ " as leader for partition "
+ partitionId);
try {
m_iv2appointees.put(partitionId, masterHSId);
} catch (Exception e) {
VoltDB.crashLocalVoltDB("Unable to appoint new master for partition " + partitionId, true, e);
}
return masterHSId;
}
@Override
protected void updateStatsRow(Object rowKey, Object[] rowValues) {
long leader;
List<Long> sites = new ArrayList<Long>();
if (rowKey.equals(MpInitiator.MP_INIT_PID)) {
leader = getHSIdForMultiPartitionInitiator();
sites.add(leader);
} else {
leader = m_iv2Masters.pointInTimeCache().get((Integer) rowKey);
sites.addAll(getReplicasForPartition((Integer) rowKey));
}
rowValues[columnNameToIndex.get("Partition")] = rowKey;
rowValues[columnNameToIndex.get("Sites")] = CoreUtils.hsIdCollectionToString(sites);
rowValues[columnNameToIndex.get("Leader")] = CoreUtils.hsIdToString(leader);
}
@Override
public void run() {
Thread.currentThread().setName("Iv2ExecutionSite: " + CoreUtils.hsIdToString(m_siteId));
initialize(m_startupConfig.m_serializedCatalog, m_startupConfig.m_timestamp);
m_startupConfig = null; // release the serializedCatalog bytes.
try {
while (m_shouldContinue) {
if (m_rejoinState == kStateRunning) {
// Normal operation blocks the site thread on the sitetasker queue.
SiteTasker task = m_scheduler.take();
if (task instanceof TransactionTask) {
m_currentTxnId = ((TransactionTask) task).getTxnId();
m_lastTxnTime = EstTime.currentTimeMillis();
}
task.run(getSiteProcedureConnection());
} else {
// Rejoin operation poll and try to do some catchup work. Tasks
// are responsible for logging any rejoin work they might have.
SiteTasker task = m_scheduler.poll();
if (task != null) {
task.runForRejoin(getSiteProcedureConnection(), m_rejoinTaskLog);
}
replayFromTaskLog();
}
}
} catch (OutOfMemoryError e) {
// Even though OOM should be caught by the Throwable section below,
// it sadly needs to be handled seperately. The goal here is to make
// sure VoltDB crashes.
String errmsg =
"Site: "
+ org.voltcore.utils.CoreUtils.hsIdToString(m_siteId)
+ " ran out of Java memory. "
+ "This node will shut down.";
VoltDB.crashLocalVoltDB(errmsg, true, e);
} catch (Throwable t) {
String errmsg =
"Site: "
+ org.voltcore.utils.CoreUtils.hsIdToString(m_siteId)
+ " encountered an "
+ "unexpected error and will die, taking this VoltDB node down.";
VoltDB.crashLocalVoltDB(errmsg, true, t);
}
shutdown();
}
@Override
public String toString() {
StringBuilder sb = new StringBuilder();
sb.append("MpProcedureTask:");
sb.append(" TXN ID: ").append(TxnEgo.txnIdToString(getTxnId()));
sb.append(" SP HANDLE ID: ").append(TxnEgo.txnIdToString(getSpHandle()));
sb.append(" ON HSID: ").append(CoreUtils.hsIdToString(m_initiator.getHSId()));
return sb.toString();
}
/** Send missed-messages to survivors. */
public void repairSurvivors() {
// cancel() and repair() must be synchronized by the caller (the deliver lock,
// currently). If cancelled and the last repair message arrives, don't send
// out corrections!
if (this.m_promotionResult.isCancelled()) {
tmLog.debug(m_whoami + "Skipping repair message creation for cancelled Term.");
return;
}
int queued = 0;
tmLog.debug(m_whoami + "received all repair logs and is repairing surviving replicas.");
for (Iv2RepairLogResponseMessage li : m_repairLogUnion) {
List<Long> needsRepair = new ArrayList<Long>(5);
for (Entry<Long, ReplicaRepairStruct> entry : m_replicaRepairStructs.entrySet()) {
if (entry.getValue().needs(li.getHandle())) {
++queued;
tmLog.debug(
m_whoami
+ "repairing "
+ CoreUtils.hsIdToString(entry.getKey())
+ ". Max seen "
+ entry.getValue().m_maxSpHandleSeen
+ ". Repairing with "
+ li.getHandle());
needsRepair.add(entry.getKey());
}
}
if (!needsRepair.isEmpty()) {
if (tmLog.isTraceEnabled()) {
tmLog.trace(
m_whoami
+ "repairing: "
+ CoreUtils.hsIdCollectionToString(needsRepair)
+ " with message: "
+ li.getPayload());
}
m_mailbox.repairReplicasWith(needsRepair, li.getPayload());
}
}
tmLog.debug(m_whoami + "finished queuing " + queued + " replica repair messages.");
m_promotionResult.done(m_maxSeenTxnId);
}
/** Notify the coordinator that this site has received the first fragment message */
private void sendFirstFragResponse() {
if (JOINLOG.isDebugEnabled()) {
JOINLOG.debug(
"P"
+ m_partitionId
+ " sending first fragment response to coordinator "
+ CoreUtils.hsIdToString(m_coordinatorHsId));
}
RejoinMessage msg =
new RejoinMessage(m_mailbox.getHSId(), RejoinMessage.Type.FIRST_FRAGMENT_RECEIVED);
m_mailbox.send(m_coordinatorHsId, msg);
m_firstFragResponseSent = true;
}
@Override
public void run() {
REJOINLOG.debug(
m_whoami
+ "informing rejoinCoordinator "
+ CoreUtils.hsIdToString(m_coordinatorHsId)
+ " of REPLAY_FINISHED");
RejoinMessage replay_complete =
new RejoinMessage(m_mailbox.getHSId(), RejoinMessage.Type.REPLAY_FINISHED);
m_mailbox.send(m_coordinatorHsId, replay_complete);
m_currentlyRejoining.set(false);
SnapshotSaveAPI.recoveringSiteCount.decrementAndGet();
}
// This message used to be sent by the SP or MP initiator when they accepted a promotion.
// For dev speed, we'll detect mastership changes here and construct and send this message to the
// local client interface so we can keep the CIs implementation
private void sendLeaderChangeNotify(long hsId, int partitionId) {
try {
JSONStringer stringer = new JSONStringer();
stringer.object();
stringer.key(JSON_PARTITION_ID).value(partitionId);
stringer.key(JSON_INITIATOR_HSID).value(hsId);
stringer.endObject();
BinaryPayloadMessage bpm =
new BinaryPayloadMessage(new byte[0], stringer.toString().getBytes("UTF-8"));
int hostId = m_hostMessenger.getHostId();
m_hostMessenger.send(
CoreUtils.getHSIdFromHostAndSite(hostId, HostMessenger.CLIENT_INTERFACE_SITE_ID), bpm);
} catch (Exception e) {
VoltDB.crashLocalVoltDB("Unable to propogate leader promotion to client interface.", true, e);
}
}
/** Start fixing survivors: setup scoreboard and request repair logs. */
void prepareForFaultRecovery() {
for (Long hsid : m_survivors) {
m_replicaRepairStructs.put(hsid, new ReplicaRepairStruct());
}
tmLog.info(
m_whoami
+ "found (including self) "
+ m_survivors.size()
+ " surviving replicas to repair. "
+ " Survivors: "
+ CoreUtils.hsIdCollectionToString(m_survivors));
VoltMessage logRequest =
new Iv2RepairLogRequestMessage(m_requestId, Iv2RepairLogRequestMessage.SPREQUEST);
m_mailbox.send(com.google.common.primitives.Longs.toArray(m_survivors), logRequest);
}
public void dump(long hsId) {
final String who = CoreUtils.hsIdToString(hsId);
tmLog.info(
String.format(
"%s: REPLAY SEQUENCER DUMP, LAST POLLED FRAGMENT %d (%s), LAST SEEN TXNID %d (%s), %s%s",
who,
m_lastPolledFragmentTxnId,
TxnEgo.txnIdToString(m_lastPolledFragmentTxnId),
m_lastSeenTxnId,
TxnEgo.txnIdToString(m_lastSeenTxnId),
m_mpiEOLReached ? "MPI EOL, " : "",
m_mustDrain ? "MUST DRAIN" : ""));
for (Entry<Long, ReplayEntry> e : m_replayEntries.entrySet()) {
tmLog.info(String.format("%s: REPLAY ENTRY %s: %s", who, e.getKey(), e.getValue()));
}
}
public SimpleFileSnapshotDataTarget(File file, boolean needsFinalClose) throws IOException {
m_file = file;
m_tempFile = new File(m_file.getParentFile(), m_file.getName() + ".incomplete");
m_ras = new RandomAccessFile(m_tempFile, "rw");
m_fc = m_ras.getChannel();
m_needsFinalClose = needsFinalClose;
m_es = CoreUtils.getListeningSingleThreadExecutor("Snapshot write thread for " + m_file);
ScheduledFuture<?> syncTask = null;
syncTask =
DefaultSnapshotDataTarget.m_syncService.scheduleAtFixedRate(
new Runnable() {
private long syncedBytes = 0;
@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_bytesSinceLastSync.get() > 1024 * 1024 * 4) {
try {
final long syncStart = syncedBytes;
syncedBytes =
Bits.sync_file_range(
SNAP_LOG, m_ras.getFD(), m_fc, syncStart, m_fc.position());
} catch (IOException e) {
if (!(e instanceof java.nio.channels.AsynchronousCloseException)) {
SNAP_LOG.error("Error syncing snapshot", e);
} else {
SNAP_LOG.debug(
"Asynchronous close syncing snapshot data, presumably graceful", e);
}
}
// Blind setting to 0 means we could technically write more than
// 256 megabytes at a time but 512 is the worst case and that is fine
m_bytesSinceLastSync.set(0);
}
}
},
DefaultSnapshotDataTarget.SNAPSHOT_SYNC_FREQUENCY,
DefaultSnapshotDataTarget.SNAPSHOT_SYNC_FREQUENCY,
TimeUnit.MILLISECONDS);
m_syncTask = syncTask;
}
@Override
public Callable<Boolean> createSetup(
String file_path,
String file_nonce,
long txnId,
Map<Integer, Long> partitionTransactionIds,
JSONObject jsData,
SystemProcedureExecutionContext context,
VoltTable result,
Map<String, Map<Integer, Pair<Long, Long>>> exportSequenceNumbers,
SiteTracker tracker,
HashinatorSnapshotData hashinatorData,
long timestamp) {
assert SnapshotSiteProcessor.ExecutionSitesCurrentlySnapshotting.isEmpty();
final IndexSnapshotRequestConfig config =
new IndexSnapshotRequestConfig(jsData, context.getDatabase());
final Map<Integer, Long> pidToLocalHSIds = findLocalSources(config.partitionRanges, tracker);
// mark snapshot start in registry
final AtomicInteger numTables = new AtomicInteger(config.tables.length);
m_snapshotRecord =
SnapshotRegistry.startSnapshot(
txnId, context.getHostId(), file_path, file_nonce, SnapshotFormat.INDEX, config.tables);
// create table tasks
for (Table table : config.tables) {
createTasksForTable(
table, config.partitionRanges, pidToLocalHSIds, numTables, m_snapshotRecord);
result.addRow(
context.getHostId(),
CoreUtils.getHostnameOrAddress(),
table.getTypeName(),
"SUCCESS",
"");
}
return null;
}
@Override
public String toString() {
StringBuilder sb = new StringBuilder();
sb.append("COMPLETE_TRANSACTION (FROM COORD: ");
sb.append(CoreUtils.hsIdToString(m_coordinatorHSId));
sb.append(") FOR TXN ");
sb.append(m_txnId);
sb.append("\n FLAGS: ").append(m_flags);
sb.append("\n HASH: " + String.valueOf(m_hash));
if (isRollback()) sb.append("\n THIS IS AN ROLLBACK REQUEST");
if (requiresAck()) sb.append("\n THIS MESSAGE REQUIRES AN ACK");
if (isRestart()) {
sb.append("\n THIS IS A TRANSACTION RESTART");
}
return sb.toString();
}
/**
* LeaderAppointer handles centralized appointment of partition leaders across the partition. This
* is primarily so that the leaders can be evenly distributed throughout the cluster, reducing
* bottlenecks (at least at startup). As a side-effect, this service also controls the initial
* startup of the cluster, blocking operation until each partition has a k-safe set of replicas,
* each partition has a leader, and the MPI has started.
*/
public class LeaderAppointer implements Promotable {
private static final VoltLogger tmLog = new VoltLogger("TM");
private enum AppointerState {
INIT, // Initial start state, used to inhibit ZK callback actions
CLUSTER_START, // indicates that we're doing the initial cluster startup
DONE // indicates normal running conditions, including repair
}
private final HostMessenger m_hostMessenger;
private final ZooKeeper m_zk;
private final int m_partitionCount;
private final BabySitter[] m_partitionWatchers;
private final LeaderCache m_iv2appointees;
private final LeaderCache m_iv2masters;
private final PartitionCallback[] m_callbacks;
private final int m_kfactor;
private final JSONObject m_topo;
private final MpInitiator m_MPI;
private final AtomicReference<AppointerState> m_state =
new AtomicReference<AppointerState>(AppointerState.INIT);
private CountDownLatch m_startupLatch = null;
private final boolean m_partitionDetectionEnabled;
private boolean m_partitionDetected = false;
private boolean m_usingCommandLog = false;
private final AtomicBoolean m_replayComplete = new AtomicBoolean(false);
// Provide a single single-threaded executor service to all the BabySitters for each partition.
// This will guarantee that the ordering of events generated by ZooKeeper is preserved in the
// handling of callbacks in LeaderAppointer.
private final ExecutorService m_es =
CoreUtils.getCachedSingleThreadExecutor("LeaderAppointer-Babysitters", 15000);
private final SnapshotSchedule m_partSnapshotSchedule;
private final SnapshotResponseHandler m_snapshotHandler =
new SnapshotResponseHandler() {
@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 class PartitionCallback extends BabySitter.Callback {
final int m_partitionId;
final Set<Long> m_replicas;
long m_currentLeader;
/** Constructor used when we know (or think we know) who the leader for this partition is */
PartitionCallback(int partitionId, long currentLeader) {
this(partitionId);
// Try to be clever for repair. Create ourselves with the current leader set to
// whatever is in the LeaderCache, and claim that replica exists, then let the
// first run() call fix the world.
m_currentLeader = currentLeader;
m_replicas.add(currentLeader);
}
/** Constructor used at startup when there is no leader */
PartitionCallback(int partitionId) {
m_partitionId = partitionId;
// A bit of a hack, but we should never end up with an HSID as Long.MAX_VALUE
m_currentLeader = Long.MAX_VALUE;
m_replicas = new HashSet<Long>();
}
@Override
public void run(List<String> children) {
List<Long> updatedHSIds = VoltZK.childrenToReplicaHSIds(children);
// compute previously unseen HSId set in the callback list
Set<Long> newHSIds = new HashSet<Long>(updatedHSIds);
newHSIds.removeAll(m_replicas);
tmLog.debug("Newly seen replicas: " + CoreUtils.hsIdCollectionToString(newHSIds));
// compute previously seen but now vanished from the callback list HSId set
Set<Long> missingHSIds = new HashSet<Long>(m_replicas);
missingHSIds.removeAll(updatedHSIds);
tmLog.debug("Newly dead replicas: " + CoreUtils.hsIdCollectionToString(missingHSIds));
tmLog.debug(
"Handling babysitter callback for partition "
+ m_partitionId
+ ": children: "
+ CoreUtils.hsIdCollectionToString(updatedHSIds));
if (m_state.get() == AppointerState.CLUSTER_START) {
// We can't yet tolerate a host failure during startup. Crash it all
if (missingHSIds.size() > 0) {
VoltDB.crashGlobalVoltDB("Node failure detected during startup.", false, null);
}
// ENG-3166: Eventually we would like to get rid of the extra replicas beyond k_factor,
// but for now we just look to see how many replicas of this partition we actually expect
// and gate leader assignment on that many copies showing up.
int replicaCount = m_kfactor + 1;
JSONArray parts;
try {
parts = m_topo.getJSONArray("partitions");
for (int p = 0; p < parts.length(); p++) {
JSONObject aPartition = parts.getJSONObject(p);
int pid = aPartition.getInt("partition_id");
if (pid == m_partitionId) {
replicaCount = aPartition.getJSONArray("replicas").length();
}
}
} catch (JSONException e) {
// Ignore and just assume the normal number of replicas
}
if (children.size() == replicaCount) {
m_currentLeader = assignLeader(m_partitionId, updatedHSIds);
} else {
tmLog.info(
"Waiting on "
+ ((m_kfactor + 1) - children.size())
+ " more nodes "
+ "for k-safety before startup");
}
} else {
// Check for k-safety
if (!isClusterKSafe()) {
VoltDB.crashGlobalVoltDB(
"Some partitions have no replicas. Cluster has become unviable.", false, null);
}
// Check if replay has completed
if (m_replayComplete.get() == false) {
VoltDB.crashGlobalVoltDB(
"Detected node failure during command log replay. Cluster will shut down.",
false,
null);
}
// Check to see if there's been a possible network partition and we're not already handling
// it
if (m_partitionDetectionEnabled && !m_partitionDetected) {
doPartitionDetectionActivities();
}
// If we survived the above gauntlet of fail, appoint a new leader for this partition.
if (missingHSIds.contains(m_currentLeader)) {
m_currentLeader = assignLeader(m_partitionId, updatedHSIds);
}
}
m_replicas.clear();
m_replicas.addAll(updatedHSIds);
}
}
/* We'll use this callback purely for startup so we can discover when all
* the leaders we have appointed have completed their promotions and
* published themselves to Zookeeper */
LeaderCache.Callback m_masterCallback =
new LeaderCache.Callback() {
@Override
public void run(ImmutableMap<Integer, Long> cache) {
Set<Long> currentLeaders = new HashSet<Long>(cache.values());
tmLog.debug("Updated leaders: " + currentLeaders);
if (m_state.get() == AppointerState.CLUSTER_START) {
if (currentLeaders.size() == m_partitionCount) {
tmLog.debug("Leader appointment complete, promoting MPI and unblocking.");
m_state.set(AppointerState.DONE);
m_MPI.acceptPromotion();
m_startupLatch.countDown();
}
}
}
};
public LeaderAppointer(
HostMessenger hm,
int numberOfPartitions,
int kfactor,
boolean partitionDetectionEnabled,
SnapshotSchedule partitionSnapshotSchedule,
boolean usingCommandLog,
JSONObject topology,
MpInitiator mpi) {
m_hostMessenger = hm;
m_zk = hm.getZK();
m_kfactor = kfactor;
m_topo = topology;
m_MPI = mpi;
m_partitionCount = numberOfPartitions;
m_callbacks = new PartitionCallback[m_partitionCount];
m_partitionWatchers = new BabySitter[m_partitionCount];
m_iv2appointees = new LeaderCache(m_zk, VoltZK.iv2appointees);
m_iv2masters = new LeaderCache(m_zk, VoltZK.iv2masters, m_masterCallback);
m_partitionDetectionEnabled = partitionDetectionEnabled;
m_partSnapshotSchedule = partitionSnapshotSchedule;
m_usingCommandLog = usingCommandLog;
}
@Override
public void acceptPromotion() throws InterruptedException, ExecutionException, KeeperException {
// Crank up the leader caches. Use blocking startup so that we'll have valid point-in-time
// caches later.
m_iv2appointees.start(true);
m_iv2masters.start(true);
// Figure out what conditions we assumed leadership under.
if (m_iv2appointees.pointInTimeCache().size() == 0) {
tmLog.debug("LeaderAppointer in startup");
m_state.set(AppointerState.CLUSTER_START);
} else if ((m_iv2appointees.pointInTimeCache().size() != m_partitionCount)
|| (m_iv2masters.pointInTimeCache().size() != m_partitionCount)) {
// If we are promoted and the appointees or masters set is partial, the previous appointer
// failed
// during startup (at least for now, until we add add/remove a partition on the fly).
VoltDB.crashGlobalVoltDB("Detected failure during startup, unable to start", false, null);
} else {
tmLog.debug("LeaderAppointer in repair");
m_state.set(AppointerState.DONE);
}
if (m_state.get() == AppointerState.CLUSTER_START) {
// Need to block the return of acceptPromotion until after the MPI is promoted. Wait for this
// latch
// to countdown after appointing all the partition leaders. The
// LeaderCache callback will count it down once it has seen all the
// appointed leaders publish themselves as the actual leaders.
m_startupLatch = new CountDownLatch(1);
writeKnownLiveNodes(m_hostMessenger.getLiveHostIds());
for (int i = 0; i < m_partitionCount; i++) {
String dir = LeaderElector.electionDirForPartition(i);
// Race along with all of the replicas for this partition to create the ZK parent node
try {
m_zk.create(dir, null, Ids.OPEN_ACL_UNSAFE, CreateMode.PERSISTENT);
} catch (KeeperException.NodeExistsException e) {
// expected on all nodes that don't start() first.
}
m_callbacks[i] = new PartitionCallback(i);
Pair<BabySitter, List<String>> sitterstuff =
BabySitter.blockingFactory(m_zk, dir, m_callbacks[i], m_es);
m_partitionWatchers[i] = sitterstuff.getFirst();
}
m_startupLatch.await();
} else {
// If we're taking over for a failed LeaderAppointer, we know when
// we get here that every partition had a leader at some point in
// time. We'll seed each of the PartitionCallbacks for each
// partition with the HSID of the last published leader. The
// blocking startup of the BabySitter watching that partition will
// call our callback, get the current full set of replicas, and
// appoint a new leader if the seeded one has actually failed
Map<Integer, Long> masters = m_iv2masters.pointInTimeCache();
tmLog.info("LeaderAppointer repairing with master set: " + masters);
for (Entry<Integer, Long> master : masters.entrySet()) {
int partId = master.getKey();
String dir = LeaderElector.electionDirForPartition(partId);
m_callbacks[partId] = new PartitionCallback(partId, master.getValue());
Pair<BabySitter, List<String>> sitterstuff =
BabySitter.blockingFactory(m_zk, dir, m_callbacks[partId], m_es);
m_partitionWatchers[partId] = sitterstuff.getFirst();
}
// just go ahead and promote our MPI
m_MPI.acceptPromotion();
}
}
private long assignLeader(int partitionId, List<Long> children) {
// We used masterHostId = -1 as a way to force the leader choice to be
// the first replica in the list, if we don't have some other mechanism
// which has successfully overridden it.
int masterHostId = -1;
if (m_state.get() == AppointerState.CLUSTER_START) {
try {
// find master in topo
JSONArray parts = m_topo.getJSONArray("partitions");
for (int p = 0; p < parts.length(); p++) {
JSONObject aPartition = parts.getJSONObject(p);
int pid = aPartition.getInt("partition_id");
if (pid == partitionId) {
masterHostId = aPartition.getInt("master");
}
}
} catch (JSONException jse) {
tmLog.error("Failed to find master for partition " + partitionId + ", defaulting to 0");
jse.printStackTrace();
masterHostId = -1; // stupid default
}
} else {
// For now, if we're appointing a new leader as a result of a
// failure, just pick the first replica in the children list.
// Could eventually do something more complex here to try to keep a
// semi-balance, but it's unclear that this has much utility until
// we add rebalancing on rejoin as well.
masterHostId = -1;
}
long masterHSId = children.get(0);
for (Long child : children) {
if (CoreUtils.getHostIdFromHSId(child) == masterHostId) {
masterHSId = child;
break;
}
}
tmLog.info(
"Appointing HSId "
+ CoreUtils.hsIdToString(masterHSId)
+ " as leader for partition "
+ partitionId);
try {
m_iv2appointees.put(partitionId, masterHSId);
} catch (Exception e) {
VoltDB.crashLocalVoltDB("Unable to appoint new master for partition " + partitionId, true, e);
}
return masterHSId;
}
private void writeKnownLiveNodes(List<Integer> liveNodes) {
try {
if (m_zk.exists(VoltZK.lastKnownLiveNodes, null) == null) {
// VoltZK.createPersistentZKNodes should have done this
m_zk.create(VoltZK.lastKnownLiveNodes, null, Ids.OPEN_ACL_UNSAFE, CreateMode.PERSISTENT);
}
JSONStringer stringer = new JSONStringer();
stringer.object();
stringer.key("liveNodes").array();
for (Integer node : liveNodes) {
stringer.value(node);
}
stringer.endArray();
stringer.endObject();
JSONObject obj = new JSONObject(stringer.toString());
tmLog.debug("Writing live nodes to ZK: " + obj.toString(4));
m_zk.setData(VoltZK.lastKnownLiveNodes, obj.toString(4).getBytes("UTF-8"), -1);
} catch (Exception e) {
VoltDB.crashLocalVoltDB(
"Unable to update known live nodes at ZK path: " + VoltZK.lastKnownLiveNodes, true, e);
}
}
private Set<Integer> readPriorKnownLiveNodes() {
Set<Integer> nodes = new HashSet<Integer>();
try {
byte[] data = m_zk.getData(VoltZK.lastKnownLiveNodes, false, null);
String jsonString = new String(data, "UTF-8");
tmLog.debug("Read prior known live nodes: " + jsonString);
JSONObject jsObj = new JSONObject(jsonString);
JSONArray jsonNodes = jsObj.getJSONArray("liveNodes");
for (int ii = 0; ii < jsonNodes.length(); ii++) {
nodes.add(jsonNodes.getInt(ii));
}
} catch (Exception e) {
VoltDB.crashLocalVoltDB(
"Unable to read prior known live nodes at ZK path: " + VoltZK.lastKnownLiveNodes,
true,
e);
}
return nodes;
}
/**
* Given a set of the known host IDs before a fault, and the known host IDs in the post-fault
* cluster, determine whether or not we think a network partition may have happened. NOTE: this
* assumes that we have already done the k-safety validation for every partition and already
* failed if we weren't a viable cluster. ALSO NOTE: not private so it may be unit-tested.
*/
static boolean makePPDDecision(Set<Integer> previousHosts, Set<Integer> currentHosts) {
// Real partition detection stuff would go here
// find the lowest hostId between the still-alive hosts and the
// failed hosts. Which set contains the lowest hostId?
int blessedHostId = Integer.MAX_VALUE;
boolean blessedHostIdInFailedSet = true;
// This should be all the pre-partition hosts IDs. Any new host IDs
// (say, if this was triggered by rejoin), will be greater than any surviving
// host ID, so don't worry about including it in this search.
for (Integer hostId : previousHosts) {
if (hostId < blessedHostId) {
blessedHostId = hostId;
}
}
for (Integer hostId : currentHosts) {
if (hostId.equals(blessedHostId)) {
blessedHostId = hostId;
blessedHostIdInFailedSet = false;
}
}
// Evaluate PPD triggers.
boolean partitionDetectionTriggered = false;
// Exact 50-50 splits. The set with the lowest survivor host doesn't trigger PPD
// If the blessed host is in the failure set, this set is not blessed.
if (currentHosts.size() * 2 == previousHosts.size()) {
if (blessedHostIdInFailedSet) {
tmLog.info(
"Partition detection triggered for 50/50 cluster failure. "
+ "This survivor set is shutting down.");
partitionDetectionTriggered = true;
} else {
tmLog.info(
"Partition detected for 50/50 failure. "
+ "This survivor set is continuing execution.");
}
}
// A strict, viable minority is always a partition.
if (currentHosts.size() * 2 < previousHosts.size()) {
tmLog.info(
"Partition detection triggered. " + "This minority survivor set is shutting down.");
partitionDetectionTriggered = true;
}
return partitionDetectionTriggered;
}
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 boolean isClusterKSafe() {
boolean retval = true;
for (int i = 0; i < m_partitionCount; i++) {
String dir = LeaderElector.electionDirForPartition(i);
try {
List<String> replicas = m_zk.getChildren(dir, null, null);
if (replicas.isEmpty()) {
tmLog.fatal("K-Safety violation: No replicas found for partition: " + i);
retval = false;
}
} catch (Exception e) {
VoltDB.crashLocalVoltDB("Unable to read replicas in ZK dir: " + dir, true, e);
}
}
return retval;
}
public void onReplayCompletion() {
m_replayComplete.set(true);
}
public void shutdown() {
try {
m_iv2appointees.shutdown();
m_iv2masters.shutdown();
for (BabySitter watcher : m_partitionWatchers) {
watcher.shutdown();
}
} catch (Exception e) {
// don't care, we're going down
}
}
}
@Override
public void run(List<String> children) {
List<Long> updatedHSIds = VoltZK.childrenToReplicaHSIds(children);
// compute previously unseen HSId set in the callback list
Set<Long> newHSIds = new HashSet<Long>(updatedHSIds);
newHSIds.removeAll(m_replicas);
tmLog.debug("Newly seen replicas: " + CoreUtils.hsIdCollectionToString(newHSIds));
// compute previously seen but now vanished from the callback list HSId set
Set<Long> missingHSIds = new HashSet<Long>(m_replicas);
missingHSIds.removeAll(updatedHSIds);
tmLog.debug("Newly dead replicas: " + CoreUtils.hsIdCollectionToString(missingHSIds));
tmLog.debug(
"Handling babysitter callback for partition "
+ m_partitionId
+ ": children: "
+ CoreUtils.hsIdCollectionToString(updatedHSIds));
if (m_state.get() == AppointerState.CLUSTER_START) {
// We can't yet tolerate a host failure during startup. Crash it all
if (missingHSIds.size() > 0) {
VoltDB.crashGlobalVoltDB("Node failure detected during startup.", false, null);
}
// ENG-3166: Eventually we would like to get rid of the extra replicas beyond k_factor,
// but for now we just look to see how many replicas of this partition we actually expect
// and gate leader assignment on that many copies showing up.
int replicaCount = m_kfactor + 1;
JSONArray parts;
try {
parts = m_topo.getJSONArray("partitions");
for (int p = 0; p < parts.length(); p++) {
JSONObject aPartition = parts.getJSONObject(p);
int pid = aPartition.getInt("partition_id");
if (pid == m_partitionId) {
replicaCount = aPartition.getJSONArray("replicas").length();
}
}
} catch (JSONException e) {
// Ignore and just assume the normal number of replicas
}
if (children.size() == replicaCount) {
m_currentLeader = assignLeader(m_partitionId, updatedHSIds);
} else {
tmLog.info(
"Waiting on "
+ ((m_kfactor + 1) - children.size())
+ " more nodes "
+ "for k-safety before startup");
}
} else {
// Check for k-safety
if (!isClusterKSafe()) {
VoltDB.crashGlobalVoltDB(
"Some partitions have no replicas. Cluster has become unviable.", false, null);
}
// Check if replay has completed
if (m_replayComplete.get() == false) {
VoltDB.crashGlobalVoltDB(
"Detected node failure during command log replay. Cluster will shut down.",
false,
null);
}
// Check to see if there's been a possible network partition and we're not already handling
// it
if (m_partitionDetectionEnabled && !m_partitionDetected) {
doPartitionDetectionActivities();
}
// If we survived the above gauntlet of fail, appoint a new leader for this partition.
if (missingHSIds.contains(m_currentLeader)) {
m_currentLeader = assignLeader(m_partitionId, updatedHSIds);
}
}
m_replicas.clear();
m_replicas.addAll(updatedHSIds);
}
/** Export data from a single catalog version and database instance. */
public class ExportGeneration {
/** Processors also log using this facility. */
private static final VoltLogger exportLog = new VoltLogger("EXPORT");
public Long m_timestamp;
public final File m_directory;
private String m_leadersZKPath;
private String m_mailboxesZKPath;
/**
* Data sources, one per table per site, provide the interface to poll() and ack() Export data
* from the execution engines. Data sources are configured by the Export manager at initialization
* time. partitionid : <tableid : datasource>.
*/
public final HashMap<Integer, HashMap<String, ExportDataSource>> m_dataSourcesByPartition =
new HashMap<Integer, HashMap<String, ExportDataSource>>();
private int m_numSources = 0;
private final AtomicInteger m_drainedSources = new AtomicInteger(0);
private final Runnable m_onAllSourcesDrained;
private final Runnable m_onSourceDrained =
new Runnable() {
@Override
public void run() {
int numSourcesDrained = m_drainedSources.incrementAndGet();
exportLog.info(
"Drained source in generation "
+ m_timestamp
+ " with "
+ numSourcesDrained
+ " of "
+ m_numSources
+ " drained");
if (numSourcesDrained == m_numSources) {
if (m_partitionLeaderZKName.isEmpty()) {
m_onAllSourcesDrained.run();
} else {
ListenableFuture<?> removeLeadership =
m_childUpdatingThread.submit(
new Runnable() {
@Override
public void run() {
for (Map.Entry<Integer, String> entry :
m_partitionLeaderZKName.entrySet()) {
m_zk.delete(
m_leadersZKPath + "/" + entry.getKey() + "/" + entry.getValue(),
-1,
new AsyncCallback.VoidCallback() {
@Override
public void processResult(int rc, String path, Object ctx) {
KeeperException.Code code = KeeperException.Code.get(rc);
if (code != KeeperException.Code.OK) {
VoltDB.crashLocalVoltDB(
"Error in export leader election giving up leadership of "
+ path,
true,
KeeperException.create(code));
}
}
},
null);
}
}
},
null);
removeLeadership.addListener(
m_onAllSourcesDrained, MoreExecutors.sameThreadExecutor());
}
;
}
}
};
private Mailbox m_mbox;
private ZooKeeper m_zk;
private volatile boolean shutdown = false;
private static final ListeningExecutorService m_childUpdatingThread =
CoreUtils.getListeningExecutorService("Export ZK Watcher", 1);
private final Map<Integer, String> m_partitionLeaderZKName = new HashMap<Integer, String>();
private final Set<Integer> m_partitionsIKnowIAmTheLeader = new HashSet<Integer>();
/*
* Set to true if this export generation was initialized from disk
* instead of being fed data from the current live system
*/
private boolean m_diskBased = false;
/**
* Constructor to create a new generation of export data
*
* @param exportOverflowDirectory
* @throws IOException
*/
public ExportGeneration(long txnId, Runnable onAllSourcesDrained, File exportOverflowDirectory)
throws IOException {
m_onAllSourcesDrained = onAllSourcesDrained;
m_timestamp = txnId;
m_directory = new File(exportOverflowDirectory, Long.toString(txnId));
if (!m_directory.mkdirs()) {
throw new IOException("Could not create " + m_directory);
}
exportLog.info("Creating new export generation " + m_timestamp);
}
/**
* Constructor to create a generation based on one that has been persisted to disk
*
* @param generationDirectory
* @param generationTimestamp
* @throws IOException
*/
public ExportGeneration(Runnable onAllSourcesDrained, File generationDirectory)
throws IOException {
m_onAllSourcesDrained = onAllSourcesDrained;
m_directory = generationDirectory;
}
public boolean isDiskBased() {
return m_diskBased;
}
boolean initializeGenerationFromDisk(final Connector conn, HostMessenger messenger) {
m_diskBased = true;
Set<Integer> partitions = new HashSet<Integer>();
/*
* Find all the advertisements. Once one is found, extract the nonce
* and check for any data files related to the advertisement. If no data files
* exist ignore the advertisement.
*/
boolean hadValidAd = false;
for (File f : m_directory.listFiles()) {
if (f.getName().endsWith(".ad")) {
boolean haveDataFiles = false;
String nonce = f.getName().substring(0, f.getName().length() - 3);
for (File dataFile : m_directory.listFiles()) {
if (dataFile.getName().startsWith(nonce) && !dataFile.getName().equals(f.getName())) {
haveDataFiles = true;
break;
}
}
if (haveDataFiles) {
try {
addDataSource(f, partitions);
hadValidAd = true;
} catch (IOException e) {
VoltDB.crashLocalVoltDB("Error intializing export datasource " + f, true, e);
}
} else {
// Delete ads that have no data
f.delete();
}
}
}
createAndRegisterAckMailboxes(partitions, messenger);
exportLog.info("Restoring export generation " + m_timestamp);
return hadValidAd;
}
/*
* Run a leader election for every partition to determine who will
* start consuming the export data.
*
*/
public void kickOffLeaderElection() {
m_childUpdatingThread.submit(
new Runnable() {
@Override
public void run() {
try {
/*
* The path where leaders will register for this generation
*/
m_leadersZKPath = VoltZK.exportGenerations + "/" + m_timestamp + "/" + "leaders";
/*
* Create a directory for each partition
*/
for (Integer partition : m_dataSourcesByPartition.keySet()) {
ZKUtil.asyncMkdirs(m_zk, m_leadersZKPath + "/" + partition);
}
/*
* Queue the creation of our ephemeral sequential and then queue
* a task to retrieve the children to find the result of the election
*/
List<ZKUtil.ChildrenCallback> callbacks = new ArrayList<ZKUtil.ChildrenCallback>();
for (final Integer partition : m_dataSourcesByPartition.keySet()) {
m_zk.create(
m_leadersZKPath + "/" + partition + "/leader",
null,
Ids.OPEN_ACL_UNSAFE,
CreateMode.EPHEMERAL_SEQUENTIAL,
new org.apache.zookeeper_voltpatches.AsyncCallback.StringCallback() {
@Override
public void processResult(int rc, String path, Object ctx, String name) {
KeeperException.Code code = KeeperException.Code.get(rc);
if (code != KeeperException.Code.OK) {
VoltDB.crashLocalVoltDB(
"Error in export leader election",
true,
KeeperException.create(code));
}
String splitName[] = name.split("/");
m_partitionLeaderZKName.put(partition, splitName[splitName.length - 1]);
}
},
null);
ZKUtil.ChildrenCallback cb = new ZKUtil.ChildrenCallback();
callbacks.add(cb);
m_zk.getChildren(
m_leadersZKPath + "/" + partition,
constructLeaderChildWatcher(partition),
cb,
null);
}
/*
* Process the result of the per partition elections.
* No worries about ordering with the watcher because the watcher tasks
* all get funneled through this thread
*/
Iterator<ZKUtil.ChildrenCallback> iter = callbacks.iterator();
for (Integer partition : m_dataSourcesByPartition.keySet()) {
ZKUtil.ChildrenCallback cb = iter.next();
handleLeaderChildrenUpdate(partition, cb.getChildren());
}
} catch (Throwable t) {
VoltDB.crashLocalVoltDB("Error in export leader election", true, t);
}
}
});
}
private Watcher constructLeaderChildWatcher(final Integer partition) {
return new Watcher() {
@Override
public void process(final WatchedEvent event) {
final Runnable processRunnable =
new Runnable() {
@Override
public void run() {
if (m_drainedSources.get() == m_numSources) {
return;
}
final AsyncCallback.ChildrenCallback childrenCallback =
new org.apache.zookeeper_voltpatches.AsyncCallback.ChildrenCallback() {
@Override
public void processResult(
final int rc,
final String path,
Object ctx,
final List<String> children) {
KeeperException.Code code = KeeperException.Code.get(rc);
if (code != KeeperException.Code.OK) {
VoltDB.crashLocalVoltDB(
"Error in export leader election",
true,
KeeperException.create(code));
}
m_childUpdatingThread.execute(
new Runnable() {
@Override
public void run() {
try {
handleLeaderChildrenUpdate(partition, children);
} catch (Throwable t) {
VoltDB.crashLocalVoltDB(
"Error in export leader election", true, t);
}
}
});
}
};
m_zk.getChildren(
m_leadersZKPath + "/" + partition,
constructLeaderChildWatcher(partition),
childrenCallback,
null);
}
};
m_childUpdatingThread.execute(processRunnable);
}
};
}
private void handleLeaderChildrenUpdate(Integer partition, List<String> children) {
if (m_drainedSources.get() == m_numSources || children.isEmpty()) {
return;
}
String leader = Collections.min(children);
if (m_partitionLeaderZKName.get(partition).equals(leader)) {
if (m_partitionsIKnowIAmTheLeader.add(partition)) {
for (ExportDataSource eds : m_dataSourcesByPartition.get(partition).values()) {
try {
eds.acceptMastership();
} catch (Exception e) {
exportLog.error("Unable to start exporting", e);
}
}
}
}
}
void initializeGenerationFromCatalog(
final Connector conn,
int hostId,
HostMessenger messenger,
List<Pair<Integer, Long>> partitions) {
/*
* Now create datasources based on the catalog
*/
Iterator<ConnectorTableInfo> tableInfoIt = conn.getTableinfo().iterator();
// Only populate partitions in use if export is actually happening
Set<Integer> partitionsInUse = new HashSet<Integer>();
while (tableInfoIt.hasNext()) {
ConnectorTableInfo next = tableInfoIt.next();
Table table = next.getTable();
addDataSources(table, hostId, partitions);
for (Pair<Integer, Long> p : partitions) {
partitionsInUse.add(p.getFirst());
}
}
createAndRegisterAckMailboxes(partitionsInUse, messenger);
}
private void createAndRegisterAckMailboxes(
final Set<Integer> localPartitions, HostMessenger messenger) {
m_zk = messenger.getZK();
m_mailboxesZKPath = VoltZK.exportGenerations + "/" + m_timestamp + "/" + "mailboxes";
m_mbox =
new LocalMailbox(messenger) {
@Override
public void deliver(VoltMessage message) {
if (message instanceof BinaryPayloadMessage) {
BinaryPayloadMessage bpm = (BinaryPayloadMessage) message;
ByteBuffer buf = ByteBuffer.wrap(bpm.m_payload);
final int partition = buf.getInt();
final int length = buf.getInt();
byte stringBytes[] = new byte[length];
buf.get(stringBytes);
String signature = new String(stringBytes, Constants.UTF8ENCODING);
final long ackUSO = buf.getLong();
final HashMap<String, ExportDataSource> partitionSources =
m_dataSourcesByPartition.get(partition);
if (partitionSources == null) {
exportLog.error(
"Received an export ack for partition "
+ partition
+ " which does not exist on this node");
return;
}
final ExportDataSource eds = partitionSources.get(signature);
if (eds == null) {
exportLog.error(
"Received an export ack for partition "
+ partition
+ " source signature "
+ signature
+ " which does not exist on this node");
return;
}
try {
eds.ack(ackUSO);
} catch (RejectedExecutionException ignoreIt) {
// ignore it: as it is already shutdown
}
} else {
exportLog.error("Receive unexpected message " + message + " in export subsystem");
}
}
};
messenger.createMailbox(null, m_mbox);
for (Integer partition : localPartitions) {
final String partitionDN = m_mailboxesZKPath + "/" + partition;
ZKUtil.asyncMkdirs(m_zk, partitionDN);
ZKUtil.StringCallback cb = new ZKUtil.StringCallback();
m_zk.create(
partitionDN + "/" + m_mbox.getHSId(),
null,
Ids.OPEN_ACL_UNSAFE,
CreateMode.EPHEMERAL,
cb,
null);
}
ListenableFuture<?> fut =
m_childUpdatingThread.submit(
new Runnable() {
@Override
public void run() {
List<Pair<Integer, ZKUtil.ChildrenCallback>> callbacks =
new ArrayList<Pair<Integer, ZKUtil.ChildrenCallback>>();
for (Integer partition : localPartitions) {
ZKUtil.ChildrenCallback callback = new ZKUtil.ChildrenCallback();
m_zk.getChildren(
m_mailboxesZKPath + "/" + partition,
constructMailboxChildWatcher(),
callback,
null);
callbacks.add(Pair.of(partition, callback));
}
for (Pair<Integer, ZKUtil.ChildrenCallback> p : callbacks) {
final Integer partition = p.getFirst();
List<String> children = null;
try {
children = p.getSecond().getChildren();
} catch (InterruptedException e) {
Throwables.propagate(e);
} catch (KeeperException e) {
Throwables.propagate(e);
}
ImmutableList.Builder<Long> mailboxes = ImmutableList.builder();
for (String child : children) {
if (child.equals(Long.toString(m_mbox.getHSId()))) continue;
mailboxes.add(Long.valueOf(child));
}
ImmutableList<Long> mailboxHsids = mailboxes.build();
for (ExportDataSource eds : m_dataSourcesByPartition.get(partition).values()) {
eds.updateAckMailboxes(Pair.of(m_mbox, mailboxHsids));
}
}
}
});
try {
fut.get();
} catch (Throwable t) {
Throwables.propagate(t);
}
}
private Watcher constructMailboxChildWatcher() {
return new Watcher() {
@Override
public void process(final WatchedEvent event) {
m_childUpdatingThread.submit(
new Runnable() {
@Override
public void run() {
try {
handleChildUpdate(event);
} catch (Throwable t) {
VoltDB.crashLocalVoltDB("Error in export ack handling", true, t);
}
}
});
}
};
}
private void handleChildUpdate(final WatchedEvent event) {
m_zk.getChildren(
event.getPath(), constructMailboxChildWatcher(), constructChildRetrievalCallback(), null);
}
private AsyncCallback.ChildrenCallback constructChildRetrievalCallback() {
return new AsyncCallback.ChildrenCallback() {
@Override
public void processResult(
final int rc, final String path, Object ctx, final List<String> children) {
m_childUpdatingThread.submit(
new Runnable() {
@Override
public void run() {
try {
if (shutdown) return;
KeeperException.Code code = KeeperException.Code.get(rc);
if (code != KeeperException.Code.OK) {
throw KeeperException.create(code);
}
final String split[] = path.split("/");
final int partition = Integer.valueOf(split[split.length - 1]);
ImmutableList.Builder<Long> mailboxes = ImmutableList.builder();
for (String child : children) {
if (child.equals(Long.toString(m_mbox.getHSId()))) continue;
mailboxes.add(Long.valueOf(child));
}
ImmutableList<Long> mailboxHsids = mailboxes.build();
for (ExportDataSource eds : m_dataSourcesByPartition.get(partition).values()) {
eds.updateAckMailboxes(Pair.of(m_mbox, mailboxHsids));
}
} catch (Throwable t) {
VoltDB.crashLocalVoltDB("Error in export ack handling", true, t);
}
}
});
}
};
}
public long getQueuedExportBytes(int partitionId, String signature) {
// assert(m_dataSourcesByPartition.containsKey(partitionId));
// assert(m_dataSourcesByPartition.get(partitionId).containsKey(delegateId));
HashMap<String, ExportDataSource> sources = m_dataSourcesByPartition.get(partitionId);
if (sources == null) {
/*
* This is fine. If the table is dropped it won't have an entry in the generation created
* after the table was dropped.
*/
// exportLog.error("Could not find export data sources for generation " +
// m_timestamp + " partition "
// + partitionId);
return 0;
}
ExportDataSource source = sources.get(signature);
if (source == null) {
/*
* This is fine. If the table is dropped it won't have an entry in the generation created
* after the table was dropped.
*/
// exportLog.error("Could not find export data source for generation " + m_timestamp + "
// partition " + partitionId +
// " signature " + signature);
return 0;
}
return source.sizeInBytes();
}
/*
* Create a datasource based on an ad file
*/
private void addDataSource(File adFile, Set<Integer> partitions) throws IOException {
m_numSources++;
ExportDataSource source = new ExportDataSource(m_onSourceDrained, adFile);
partitions.add(source.getPartitionId());
m_timestamp = source.getGeneration();
exportLog.info(
"Creating ExportDataSource for "
+ adFile
+ " table "
+ source.getTableName()
+ " signature "
+ source.getSignature()
+ " partition id "
+ source.getPartitionId()
+ " bytes "
+ source.sizeInBytes());
HashMap<String, ExportDataSource> dataSourcesForPartition =
m_dataSourcesByPartition.get(source.getPartitionId());
if (dataSourcesForPartition == null) {
dataSourcesForPartition = new HashMap<String, ExportDataSource>();
m_dataSourcesByPartition.put(source.getPartitionId(), dataSourcesForPartition);
}
dataSourcesForPartition.put(source.getSignature(), source);
}
/*
* An unfortunate test only method for supplying a mock source
*/
public void addDataSource(ExportDataSource source) {
HashMap<String, ExportDataSource> dataSourcesForPartition =
m_dataSourcesByPartition.get(source.getPartitionId());
if (dataSourcesForPartition == null) {
dataSourcesForPartition = new HashMap<String, ExportDataSource>();
m_dataSourcesByPartition.put(source.getPartitionId(), dataSourcesForPartition);
}
dataSourcesForPartition.put(source.getSignature(), source);
}
// silly helper to add datasources for a table catalog object
private void addDataSources(Table table, int hostId, List<Pair<Integer, Long>> partitions) {
for (Pair<Integer, Long> p : partitions) {
Integer partition = p.getFirst();
Long site = p.getSecond();
/*
* IOException can occur if there is a problem
* with the persistent aspects of the datasource storage
*/
try {
HashMap<String, ExportDataSource> dataSourcesForPartition =
m_dataSourcesByPartition.get(partition);
if (dataSourcesForPartition == null) {
dataSourcesForPartition = new HashMap<String, ExportDataSource>();
m_dataSourcesByPartition.put(partition, dataSourcesForPartition);
}
ExportDataSource exportDataSource =
new ExportDataSource(
m_onSourceDrained,
"database",
table.getTypeName(),
partition,
site,
table.getSignature(),
m_timestamp,
table.getColumns(),
m_directory.getPath());
m_numSources++;
exportLog.info(
"Creating ExportDataSource for table "
+ table.getTypeName()
+ " signature "
+ table.getSignature()
+ " partition id "
+ partition);
dataSourcesForPartition.put(table.getSignature(), exportDataSource);
} catch (IOException e) {
VoltDB.crashLocalVoltDB(
"Error creating datasources for table " + table.getTypeName() + " host id " + hostId,
true,
e);
}
}
}
public void pushExportBuffer(
int partitionId,
String signature,
long uso,
long bufferPtr,
ByteBuffer buffer,
boolean sync,
boolean endOfStream) {
// System.out.println("In generation " + m_timestamp + " partition " + partitionId + "
// signature " + signature + (buffer == null ? " null buffer " : (" buffer length " +
// buffer.remaining())));
// for (Integer i : m_dataSourcesByPartition.keySet()) {
// System.out.println("Have partition " + i);
// }
assert (m_dataSourcesByPartition.containsKey(partitionId));
assert (m_dataSourcesByPartition.get(partitionId).containsKey(signature));
HashMap<String, ExportDataSource> sources = m_dataSourcesByPartition.get(partitionId);
if (sources == null) {
exportLog.error(
"Could not find export data sources for partition "
+ partitionId
+ " generation "
+ m_timestamp
+ " the export data is being discarded");
DBBPool.deleteCharArrayMemory(bufferPtr);
return;
}
ExportDataSource source = sources.get(signature);
if (source == null) {
exportLog.error(
"Could not find export data source for partition "
+ partitionId
+ " signature "
+ signature
+ " generation "
+ m_timestamp
+ " the export data is being discarded");
DBBPool.deleteCharArrayMemory(bufferPtr);
return;
}
source.pushExportBuffer(uso, bufferPtr, buffer, sync, endOfStream);
}
public void closeAndDelete() throws IOException {
List<ListenableFuture<?>> tasks = new ArrayList<ListenableFuture<?>>();
for (HashMap<String, ExportDataSource> map : m_dataSourcesByPartition.values()) {
for (ExportDataSource source : map.values()) {
tasks.add(source.closeAndDelete());
}
}
try {
Futures.allAsList(tasks).get();
} catch (Exception e) {
Throwables.propagateIfPossible(e, IOException.class);
}
shutdown = true;
VoltFile.recursivelyDelete(m_directory);
}
/*
* Returns true if the generatino was completely truncated away
*/
public boolean truncateExportToTxnId(long txnId, long[] perPartitionTxnIds) {
// create an easy partitionId:txnId lookup.
HashMap<Integer, Long> partitionToTxnId = new HashMap<Integer, Long>();
for (long tid : perPartitionTxnIds) {
partitionToTxnId.put(TxnEgo.getPartitionId(tid), tid);
}
List<ListenableFuture<?>> tasks = new ArrayList<ListenableFuture<?>>();
// pre-iv2, the truncation point is the snapshot transaction id.
// In iv2, truncation at the per-partition txn id recorded in the snapshot.
for (HashMap<String, ExportDataSource> dataSources : m_dataSourcesByPartition.values()) {
for (ExportDataSource source : dataSources.values()) {
if (VoltDB.instance().isIV2Enabled()) {
Long truncationPoint = partitionToTxnId.get(source.getPartitionId());
if (truncationPoint == null) {
exportLog.error(
"Snapshot "
+ txnId
+ " does not include truncation point for partition "
+ source.getPartitionId());
} else {
tasks.add(source.truncateExportToTxnId(truncationPoint));
}
} else {
tasks.add(source.truncateExportToTxnId(txnId));
}
}
}
try {
Futures.allAsList(tasks).get();
} catch (Exception e) {
VoltDB.crashLocalVoltDB(
"Unexpected exception truncating export data during snapshot restore. "
+ "You can back up export overflow data and start the "
+ "DB without it to get past this error",
true,
e);
}
return m_drainedSources.get() == m_numSources;
}
public void close() {
List<ListenableFuture<?>> tasks = new ArrayList<ListenableFuture<?>>();
for (HashMap<String, ExportDataSource> sources : m_dataSourcesByPartition.values()) {
for (ExportDataSource source : sources.values()) {
tasks.add(source.close());
}
}
try {
Futures.allAsList(tasks).get();
} catch (Exception e) {
// Logging of errors is done inside the tasks so nothing to do here
// intentionally not failing if there is an issue with close
exportLog.error("Error closing export data sources", e);
}
shutdown = true;
}
/**
* Indicate to all associated {@link ExportDataSource}to assume mastership role for the given
* partition id
*
* @param partitionId
*/
public void acceptMastershipTask(int partitionId) {
HashMap<String, ExportDataSource> partitionDataSourceMap =
m_dataSourcesByPartition.get(partitionId);
exportLog.info(
"Export generation " + m_timestamp + " accepting mastership for partition " + partitionId);
for (ExportDataSource eds : partitionDataSourceMap.values()) {
try {
eds.acceptMastership();
} catch (Exception e) {
exportLog.error("Unable to start exporting", e);
}
}
}
@Override
public String toString() {
return "Export Generation - " + m_timestamp.toString();
}
}
