@Override
@Test
public void testGetVersions() throws Exception {
List<ByteArray> keys = getKeys(2);
ByteArray key = keys.get(0);
byte[] value = getValue();
VectorClock vc = getClock(0, 0);
Store<ByteArray, byte[], byte[]> store = getStore();
store.put(key, Versioned.value(value, vc), null);
List<Versioned<byte[]>> versioneds = store.get(key, null);
List<Version> versions = store.getVersions(key);
assertEquals(1, versioneds.size());
assertTrue(versions.size() > 0);
for (int i = 0; i < versions.size(); i++)
assertEquals(versioneds.get(0).getVersion(), versions.get(i));
assertEquals(0, store.getVersions(keys.get(1)).size());
}
@SuppressWarnings("unchecked")
public class StreamingSlopPusherJob implements Runnable {
private static final Logger logger = Logger.getLogger(StreamingSlopPusherJob.class.getName());
public static final String TYPE_NAME = "streaming";
private static final Versioned<Slop> END = Versioned.value(null);
private final MetadataStore metadataStore;
private final StoreRepository storeRepo;
private final FailureDetector failureDetector;
private final ConcurrentMap<Integer, SynchronousQueue<Versioned<Slop>>> slopQueues;
private final ExecutorService consumerExecutor;
private final EventThrottler readThrottler;
private AdminClient adminClient;
private final Cluster cluster;
private final List<Future> consumerResults;
private final VoldemortConfig voldemortConfig;
private final Map<Integer, Set<Integer>> zoneMapping;
private final ConcurrentHashMap<Integer, Long> attemptedByNode, succeededByNode;
private final Semaphore repairPermits;
public StreamingSlopPusherJob(
StoreRepository storeRepo,
MetadataStore metadataStore,
FailureDetector failureDetector,
VoldemortConfig voldemortConfig,
Semaphore repairPermits) {
this.storeRepo = storeRepo;
this.metadataStore = metadataStore;
this.failureDetector = failureDetector;
this.voldemortConfig = voldemortConfig;
this.repairPermits = Utils.notNull(repairPermits);
this.cluster = metadataStore.getCluster();
this.slopQueues =
new ConcurrentHashMap<Integer, SynchronousQueue<Versioned<Slop>>>(
cluster.getNumberOfNodes());
this.consumerExecutor =
Executors.newFixedThreadPool(
cluster.getNumberOfNodes(),
new ThreadFactory() {
public Thread newThread(Runnable r) {
Thread thread = new Thread(r);
thread.setName("slop-pusher");
return thread;
}
});
this.readThrottler = new EventThrottler(voldemortConfig.getSlopMaxReadBytesPerSec());
this.adminClient = null;
this.consumerResults = Lists.newArrayList();
this.attemptedByNode = new ConcurrentHashMap<Integer, Long>(cluster.getNumberOfNodes());
this.succeededByNode = new ConcurrentHashMap<Integer, Long>(cluster.getNumberOfNodes());
this.zoneMapping = Maps.newHashMap();
}
public void run() {
// don't try to run slop pusher job when rebalancing
if (metadataStore
.getServerState()
.equals(MetadataStore.VoldemortState.REBALANCING_MASTER_SERVER)) {
logger.error("Cannot run slop pusher job since Voldemort server is rebalancing");
return;
}
boolean terminatedEarly = false;
Date startTime = new Date();
logger.info("Started streaming slop pusher job at " + startTime);
SlopStorageEngine slopStorageEngine = storeRepo.getSlopStore();
ClosableIterator<Pair<ByteArray, Versioned<Slop>>> iterator = null;
if (adminClient == null) {
adminClient =
new AdminClient(
cluster,
new AdminClientConfig()
.setMaxThreads(cluster.getNumberOfNodes())
.setMaxConnectionsPerNode(1));
}
if (voldemortConfig.getSlopZonesDownToTerminate() > 0) {
// Populating the zone mapping for early termination
zoneMapping.clear();
for (Node n : cluster.getNodes()) {
if (failureDetector.isAvailable(n)) {
Set<Integer> nodes = zoneMapping.get(n.getZoneId());
if (nodes == null) {
nodes = Sets.newHashSet();
zoneMapping.put(n.getZoneId(), nodes);
}
nodes.add(n.getId());
}
}
// Check how many zones are down
int zonesDown = 0;
for (Zone zone : cluster.getZones()) {
if (zoneMapping.get(zone.getId()) == null || zoneMapping.get(zone.getId()).size() == 0)
zonesDown++;
}
// Terminate early
if (voldemortConfig.getSlopZonesDownToTerminate() <= zoneMapping.size()
&& zonesDown >= voldemortConfig.getSlopZonesDownToTerminate()) {
logger.info(
"Completed streaming slop pusher job at "
+ startTime
+ " early because "
+ zonesDown
+ " zones are down");
stopAdminClient();
return;
}
}
// Clearing the statistics
AtomicLong attemptedPushes = new AtomicLong(0);
for (Node node : cluster.getNodes()) {
attemptedByNode.put(node.getId(), 0L);
succeededByNode.put(node.getId(), 0L);
}
acquireRepairPermit();
try {
StorageEngine<ByteArray, Slop, byte[]> slopStore = slopStorageEngine.asSlopStore();
iterator = slopStore.entries();
while (iterator.hasNext()) {
Pair<ByteArray, Versioned<Slop>> keyAndVal;
try {
keyAndVal = iterator.next();
Versioned<Slop> versioned = keyAndVal.getSecond();
// Retrieve the node
int nodeId = versioned.getValue().getNodeId();
Node node = cluster.getNodeById(nodeId);
attemptedPushes.incrementAndGet();
Long attempted = attemptedByNode.get(nodeId);
attemptedByNode.put(nodeId, attempted + 1L);
if (attemptedPushes.get() % 10000 == 0)
logger.info("Attempted pushing " + attemptedPushes + " slops");
if (logger.isTraceEnabled())
logger.trace(
"Pushing slop for "
+ versioned.getValue().getNodeId()
+ " and store "
+ versioned.getValue().getStoreName());
if (failureDetector.isAvailable(node)) {
SynchronousQueue<Versioned<Slop>> slopQueue = slopQueues.get(nodeId);
if (slopQueue == null) {
// No previous slop queue, add one
slopQueue = new SynchronousQueue<Versioned<Slop>>();
slopQueues.put(nodeId, slopQueue);
consumerResults.add(
consumerExecutor.submit(new SlopConsumer(nodeId, slopQueue, slopStorageEngine)));
}
boolean offered =
slopQueue.offer(
versioned, voldemortConfig.getClientRoutingTimeoutMs(), TimeUnit.MILLISECONDS);
if (!offered) {
if (logger.isDebugEnabled())
logger.debug(
"No consumer appeared for slop in "
+ voldemortConfig.getClientConnectionTimeoutMs()
+ " ms");
}
readThrottler.maybeThrottle(nBytesRead(keyAndVal));
} else {
logger.trace(node + " declared down, won't push slop");
}
} catch (RejectedExecutionException e) {
throw new VoldemortException("Ran out of threads in executor", e);
}
}
} catch (InterruptedException e) {
logger.warn("Interrupted exception", e);
terminatedEarly = true;
} catch (Exception e) {
logger.error(e, e);
terminatedEarly = true;
} finally {
try {
if (iterator != null) iterator.close();
} catch (Exception e) {
logger.warn("Failed to close iterator cleanly as database might be closed", e);
}
// Adding the poison pill
for (SynchronousQueue<Versioned<Slop>> slopQueue : slopQueues.values()) {
try {
slopQueue.put(END);
} catch (InterruptedException e) {
logger.warn("Error putting poison pill", e);
}
}
for (Future result : consumerResults) {
try {
result.get();
} catch (Exception e) {
logger.warn("Exception in consumer", e);
}
}
// Only if exception didn't take place do we update the counts
if (!terminatedEarly) {
Map<Integer, Long> outstanding =
Maps.newHashMapWithExpectedSize(cluster.getNumberOfNodes());
for (int nodeId : succeededByNode.keySet()) {
logger.info(
"Slops to node "
+ nodeId
+ " - Succeeded - "
+ succeededByNode.get(nodeId)
+ " - Attempted - "
+ attemptedByNode.get(nodeId));
outstanding.put(nodeId, attemptedByNode.get(nodeId) - succeededByNode.get(nodeId));
}
slopStorageEngine.resetStats(outstanding);
logger.info("Completed streaming slop pusher job which started at " + startTime);
} else {
for (int nodeId : succeededByNode.keySet()) {
logger.info(
"Slops to node "
+ nodeId
+ " - Succeeded - "
+ succeededByNode.get(nodeId)
+ " - Attempted - "
+ attemptedByNode.get(nodeId));
}
logger.info("Completed early streaming slop pusher job which started at " + startTime);
}
// Shut down admin client as not to waste connections
consumerResults.clear();
slopQueues.clear();
stopAdminClient();
this.repairPermits.release();
}
}
private void stopAdminClient() {
if (adminClient != null) {
adminClient.stop();
adminClient = null;
}
}
private int nBytesRead(Pair<ByteArray, Versioned<Slop>> keyAndVal) {
return keyAndVal.getFirst().length() + slopSize(keyAndVal.getSecond());
}
/**
* Returns the approximate size of slop to help in throttling
*
* @param slopVersioned The versioned slop whose size we want
* @return Size in bytes
*/
private int slopSize(Versioned<Slop> slopVersioned) {
int nBytes = 0;
Slop slop = slopVersioned.getValue();
nBytes += slop.getKey().length();
nBytes += ((VectorClock) slopVersioned.getVersion()).sizeInBytes();
switch (slop.getOperation()) {
case PUT:
{
nBytes += slop.getValue().length;
break;
}
case DELETE:
{
break;
}
default:
logger.error("Unknown slop operation: " + slop.getOperation());
}
return nBytes;
}
/** Smart slop iterator which keeps two previous batches of data */
private class SlopIterator extends AbstractIterator<Versioned<Slop>> {
private final SynchronousQueue<Versioned<Slop>> slopQueue;
private final List<Pair<ByteArray, Version>> deleteBatch;
private final EventThrottler writeThrottler;
private int writtenLast = 0;
private long slopsDone = 0L;
private boolean shutDown = false, isComplete = false;
public SlopIterator(
SynchronousQueue<Versioned<Slop>> slopQueue, List<Pair<ByteArray, Version>> deleteBatch) {
this.slopQueue = slopQueue;
this.deleteBatch = deleteBatch;
this.writeThrottler = new EventThrottler(voldemortConfig.getSlopMaxWriteBytesPerSec());
}
public boolean isComplete() {
return isComplete;
}
@Override
protected Versioned<Slop> computeNext() {
try {
Versioned<Slop> head = null;
if (!shutDown) {
head = slopQueue.take();
if (head.equals(END)) {
shutDown = true;
isComplete = true;
} else {
slopsDone++;
if (slopsDone % voldemortConfig.getSlopBatchSize() == 0) {
shutDown = true;
}
writeThrottler.maybeThrottle(writtenLast);
writtenLast = slopSize(head);
deleteBatch.add(Pair.create(head.getValue().makeKey(), head.getVersion()));
return head;
}
}
return endOfData();
} catch (Exception e) {
logger.error("Got an exception " + e);
return endOfData();
}
}
}
private void acquireRepairPermit() {
logger.info("Acquiring lock to perform streaming slop pusher job ");
try {
this.repairPermits.acquire();
logger.info("Acquired lock to perform streaming slop pusher job ");
} catch (InterruptedException e) {
stopAdminClient();
throw new IllegalStateException(
"Streaming slop pusher job interrupted while waiting for permit.", e);
}
}
private class SlopConsumer implements Runnable {
private final int nodeId;
private SynchronousQueue<Versioned<Slop>> slopQueue;
private long startTime;
private SlopStorageEngine slopStorageEngine;
// Keep two lists to track deleted items
private List<Pair<ByteArray, Version>> previous, current;
public SlopConsumer(
int nodeId,
SynchronousQueue<Versioned<Slop>> slopQueue,
SlopStorageEngine slopStorageEngine) {
this.nodeId = nodeId;
this.slopQueue = slopQueue;
this.slopStorageEngine = slopStorageEngine;
this.previous = Lists.newArrayList();
this.current = Lists.newArrayList();
}
public void run() {
try {
SlopIterator iterator = null;
do {
if (!current.isEmpty()) {
if (!previous.isEmpty()) {
for (Pair<ByteArray, Version> entry : previous) {
slopStorageEngine.delete(entry.getFirst(), entry.getSecond());
}
Long succeeded = succeededByNode.get(nodeId);
succeeded += previous.size();
succeededByNode.put(nodeId, succeeded);
previous.clear();
}
previous = null;
previous = current;
current = Lists.newArrayList();
}
this.startTime = System.currentTimeMillis();
iterator = new SlopIterator(slopQueue, current);
adminClient.updateSlopEntries(nodeId, iterator);
} while (!iterator.isComplete());
// Clear up both previous and current
if (!previous.isEmpty()) {
for (Pair<ByteArray, Version> entry : previous)
slopStorageEngine.delete(entry.getFirst(), entry.getSecond());
Long succeeded = succeededByNode.get(nodeId);
succeeded += previous.size();
succeededByNode.put(nodeId, succeeded);
previous.clear();
}
if (!current.isEmpty()) {
for (Pair<ByteArray, Version> entry : current)
slopStorageEngine.delete(entry.getFirst(), entry.getSecond());
Long succeeded = succeededByNode.get(nodeId);
succeeded += current.size();
succeededByNode.put(nodeId, succeeded);
current.clear();
}
} catch (UnreachableStoreException e) {
failureDetector.recordException(
metadataStore.getCluster().getNodeById(nodeId),
System.currentTimeMillis() - this.startTime,
e);
throw e;
} finally {
// Clean the slop queue and remove the queue from the global
// queue
slopQueue.clear();
slopQueues.remove(nodeId);
}
}
}
}