com.splout.db.qnode.QNodeHandlerContext Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of splout-server Show documentation
Show all versions of splout-server Show documentation
Splout SQL is a read only, horizontally scalable and
partitioned SQL database that plays well with Hadoop.
The newest version!
package com.splout.db.qnode;
/*
* #%L
* Splout SQL Server
* %%
* Copyright (C) 2012 Datasalt Systems S.L.
* %%
* This program is free software: you can redistribute it and/or modify
* it under the terms of the GNU Affero General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU Affero General Public License
* along with this program. If not, see currentVersionsMap = new ConcurrentHashMap();
// The SploutConfiguration
private SploutConfiguration config;
// The coordination structures that use Hazelcast underneath
private CoordinationStructures coordinationStructures;
// Local map with all versions for a tablespace with the PartitionMap, ReplicationMap for each of them
private final Map tablespaceVersionsMap = new ConcurrentHashMap();
private ReplicaBalancer replicaBalancer;
// This flag is set to "false" after WARMING_TIME seconds (qnode.warming.time)
// Some actions will only be taken after warming time, just in case some nodes still didn't join the cluster.
private final AtomicBoolean isWarming = new AtomicBoolean(true);
// The per-DNode Thrift client pools
private ConcurrentMap> thriftClientCache = new ConcurrentHashMap>();
private ReentrantLock thriftClientCacheLock = new ReentrantLock();
private int thriftClientPoolSize;
public QNodeHandlerContext(SploutConfiguration config, CoordinationStructures coordinationStructures) {
this.config = config;
this.coordinationStructures = coordinationStructures;
this.thriftClientPoolSize = config.getInt(QNodeProperties.DNODE_POOL_SIZE);
this.replicaBalancer = new ReplicaBalancer(this);
initMetrics();
}
public static enum DNodeEvent {
LEAVE, ENTRY, UPDATE
}
private void initMetrics() {
Metrics.newGauge(QNodeHandlerContext.class, "thrift-total-connections-iddle", new Gauge() {
@Override
public Integer value() {
int count = 0;
for(Entry> queue : thriftClientCache.entrySet()) {
count += queue.getValue().size();
}
return count;
}
});
Metrics.newGauge(QNodeHandlerContext.class, "thrift-pools", new Gauge() {
@Override
public Integer value() {
return thriftClientCache.size();
}
});
}
@SuppressWarnings("serial")
public final static class TablespaceVersionInfoException extends Exception {
public TablespaceVersionInfoException(String msg) {
super(msg);
}
}
public Object tVLock = new Object();
/**
* Get the list of possible actions to take for balancing the cluster in case of under-replicated partitions.
*/
public List getBalanceActions() {
// we have this in this class to be able to use this lock (the same that recreats the in-memory TablespaceVersion map)
synchronized(tVLock) {
return replicaBalancer.scanPartitions();
}
}
/**
* Get the list of DNodes
*/
public List getDNodeList() {
List dNodeList = new ArrayList();
for(DNodeInfo dnode : getCoordinationStructures().getDNodes().values()) {
dNodeList.add(dnode.getAddress());
}
return dNodeList;
}
/**
* Update the in-memory map when a DNode joins, leaves or updates its DNodeINfo.
*/
public void updateTablespaceVersions(DNodeInfo dNodeInfo, DNodeEvent event)
throws TablespaceVersionInfoException {
synchronized(tVLock) {
Map tablespaceVersionMap = getTablespaceVersionsMap();
// First check if this DNode is not anymore serving a version that it used to serve (IMPLICIT leaving).
// This can happen for instance if a DNode removes an old version.
// In this case the version will eventually become empty here.
Iterator> iterator = tablespaceVersionMap.entrySet()
.iterator();
while(iterator.hasNext()) {
Map.Entry tablespaceVersion = iterator.next();
String tablespaceName = tablespaceVersion.getKey().getTablespace();
Long version = tablespaceVersion.getKey().getVersion();
// Is this DNode present in this version?
Tablespace tablespace = tablespaceVersion.getValue();
// We will rebuild the replication map to check if it became empty after the checkings or not
int nonEmptyReplicas = 0;
Iterator repIter = tablespace.getReplicationMap().getReplicationEntries()
.iterator();
while(repIter.hasNext()) {
ReplicationEntry entry = repIter.next();
int partition = entry.getShard();
if(entry.getNodes().contains(dNodeInfo.getAddress())) {
// Yes!
// So we have to check if this DNode is still serving this version/partition or not
if((dNodeInfo.getServingInfo().get(tablespaceName) == null)
|| (dNodeInfo.getServingInfo().get(tablespaceName).get(version) == null)
|| (dNodeInfo.getServingInfo().get(tablespaceName).get(version).get(partition) == null)) {
// NO! So we have to remove the DNode
entry.getNodes().remove(dNodeInfo.getAddress());
if(entry.getNodes().isEmpty()) {
repIter.remove();
// Remove also from PartitionMap
PartitionEntry pEntry = new PartitionEntry();
pEntry.setShard(entry.getShard());
tablespace.getPartitionMap().getPartitionEntries().remove(pEntry);
}
}
}
if(!entry.getNodes().isEmpty()) {
nonEmptyReplicas++;
}
}
if(nonEmptyReplicas == 0) {
// Delete TablespaceVersion
log.info("Removing empty tablespace version (implicit leaving from " + dNodeInfo.getAddress()
+ "): " + tablespaceName + ", " + version);
iterator.remove();
}
}
// Now iterate over all the tablespaces of this DNode to see new additions or EXPLICIT leavings
for(Map.Entry>> tablespaceEntry : dNodeInfo
.getServingInfo().entrySet()) {
String tablespaceName = tablespaceEntry.getKey();
// Iterate over all versions of this tablespace
for(Map.Entry> versionEntry : tablespaceEntry.getValue()
.entrySet()) {
Long versionName = versionEntry.getKey();
TablespaceVersion tablespaceVersion = new TablespaceVersion(tablespaceName, versionName);
Tablespace currentTablespace = tablespaceVersionMap.get(tablespaceVersion);
List partitionMap = new ArrayList();
List replicationMap = new ArrayList();
long deployDate = -1;
if(currentTablespace != null) {
// Not first time we see this tablespace. We do a copy of the partition map to be able to modify it without
// contention.
partitionMap.addAll(currentTablespace.getPartitionMap().getPartitionEntries());
replicationMap.addAll(currentTablespace.getReplicationMap().getReplicationEntries());
deployDate = currentTablespace.getCreationDate();
}
// Iterate over all partitions of this tablespace
for(Map.Entry partition : versionEntry.getValue().entrySet()) {
deployDate = deployDate == -1 ? partition.getValue().getDeploymentDate() : deployDate;
if(deployDate != -1 && (deployDate != partition.getValue().getDeploymentDate())) {
throw new TablespaceVersionInfoException(
"Inconsistent partition metadata within same node, deploy date was " + deployDate
+ " versus " + partition.getValue().getDeploymentDate());
}
PartitionMetadata metadata = partition.getValue();
Integer shard = partition.getKey();
// Create a PartitionEntry according to this PartitionMetadata
PartitionEntry myEntry = new PartitionEntry();
myEntry.setMax(metadata.getMaxKey());
myEntry.setMin(metadata.getMinKey());
myEntry.setShard(shard);
PartitionEntry existingPartitionEntry = null;
// Look for an existing PartitionEntry for the same shard in the PartitionMap
if(!partitionMap.contains(myEntry)) {
if(!event.equals(DNodeEvent.LEAVE)) {
// In this case all conditions are met for adding a new entry to the PartitionMap
partitionMap.add(myEntry);
// Note that now the PartitionMap is not necessarily sorted! let's sort it now
Collections.sort(partitionMap);
}
} else {
// Check consistency of this Partition Metadata
existingPartitionEntry = partitionMap.get(partitionMap.indexOf(myEntry));
if(existingPartitionEntry.getMax() == null || myEntry.getMax() == null) {
if(!(existingPartitionEntry.getMax() == null && myEntry.getMax() == null)) {
throw new TablespaceVersionInfoException(
"Inconsistent partition metadata between nodes: " + existingPartitionEntry
+ " versus " + myEntry);
}
} else {
if(!existingPartitionEntry.getMax().equals(myEntry.getMax())) {
throw new TablespaceVersionInfoException(
"Inconsistent partition metadata between nodes: " + existingPartitionEntry
+ " versus " + myEntry);
}
}
if(existingPartitionEntry.getMin() == null || myEntry.getMin() == null) {
if(!(existingPartitionEntry.getMin() == null && myEntry.getMin() == null)) {
throw new TablespaceVersionInfoException(
"Inconsistent partition metadata between nodes: " + existingPartitionEntry
+ " versus " + myEntry);
}
} else {
if(!existingPartitionEntry.getMin().equals(myEntry.getMin())) {
throw new TablespaceVersionInfoException(
"Inconsistent partition metadata between nodes: " + existingPartitionEntry
+ " versus " + myEntry);
}
}
}
// Create a ReplicationEntry according to this PartitionMetadata
// Will only contain this DNode as we don't know about the others yet
ReplicationEntry reEntry = new ReplicationEntry();
reEntry.setShard(shard);
reEntry.setExpectedReplicationFactor(metadata.getNReplicas());
reEntry.setNodes(new ArrayList());
// Look for an existing ReplicationEntry for the same shard in the ReplicationMap
if(replicationMap.contains(reEntry)) {
ReplicationEntry existingEntry = replicationMap.get(replicationMap.indexOf(reEntry));
if(event.equals(DNodeEvent.LEAVE)) {
// Remove it from replication map and partition map
existingEntry.getNodes().remove(dNodeInfo.getAddress());
if(existingEntry.getNodes().isEmpty()) {
replicationMap.remove(existingEntry);
if(existingPartitionEntry != null) {
partitionMap.remove(existingPartitionEntry);
} else {
throw new RuntimeException(
"ReplicationEntry for one shard with no associated PartitionEntry. This is very likely to be a software bug.");
}
}
} else {
if(!existingEntry.getNodes().contains(dNodeInfo.getAddress())) {
// Add it to replication map
existingEntry.getNodes().add(dNodeInfo.getAddress());
} else {
// We are adding / updating but the node already exists in the replication map.
}
}
} else if(!event.equals(DNodeEvent.LEAVE)) { // Otherwise just add and sort
// We check the DNodeEvent but although would be very weird if this DNode leaves and its ReplicationEntry
// wasn't present
reEntry.getNodes().add(dNodeInfo.getAddress());
replicationMap.add(reEntry);
Collections.sort(reEntry.getNodes());
Collections.sort(replicationMap);
}
}
// Delete tablespaceVersion if it is empty now
if(currentTablespace != null && replicationMap.size() == 0) {
log.info("Removing empty tablespaceVersion: " + tablespaceVersion
+ " due to explicit leaving from node " + dNodeInfo.getAddress());
tablespaceVersionMap.remove(tablespaceVersion);
} else {
// Update the info in memory
currentTablespace = new Tablespace(new PartitionMap(partitionMap), new ReplicationMap(
replicationMap), versionName, deployDate);
tablespaceVersionMap.put(tablespaceVersion, currentTablespace);
}
}
}
}
}
/**
* This method can be called to initialize a pool of connections to a dnode. This method may be called from multiple
* threads so it should be safe to call it concurrently.
*/
public void initializeThriftClientCacheFor(String dnode) throws TTransportException,
InterruptedException {
// this lock is on the whole cache but we would actually be interested in a per-DNode lock...
// there's only one lock for simplicity.
thriftClientCacheLock.lock();
try {
// initialize queue for this DNode
BlockingQueue dnodeQueue = thriftClientCache.get(dnode);
if(dnodeQueue == null) {
// this assures that the per-DNode queue is only created once and then reused.
dnodeQueue = new LinkedBlockingDeque(thriftClientPoolSize);
}
if(dnodeQueue.isEmpty()) {
try {
for(int i = dnodeQueue.size(); i < thriftClientPoolSize; i++) {
dnodeQueue.put(DNodeClient.get(dnode));
}
// we only put the queue if all connections have been populated
thriftClientCache.put(dnode, dnodeQueue);
} catch(TTransportException e) {
log.error("Error while trying to populate queue for " + dnode
+ ", will discard created connections.", e);
while(!dnodeQueue.isEmpty()) {
dnodeQueue.poll().getOutputProtocol().getTransport().close();
}
throw e;
}
} else {
// it should be safe to call this method from different places concurrently
// so we contemplate the case where another Thread already populated the queue
// and only populate it if it's really empty.
log.warn(Thread.currentThread().getName() + " : queue for [" + dnode
+ "] is not empty - it was populated before.");
}
} finally {
thriftClientCacheLock.unlock();
}
}
/**
* This method can be called by {@link QNodeHandler} to cancel the Thrift client cache when a DNode disconnects.
* Usually this happens when Hazelcast notifies it.
*/
public void discardThriftClientCacheFor(String dnode) throws InterruptedException {
thriftClientCacheLock.lock();
try {
// discarding all connections to a DNode who leaved
log.info(Thread.currentThread().getName() + " : trashing queue for [" + dnode + "] as it leaved.");
BlockingQueue dnodeQueue = thriftClientCache.get(dnode);
// release connections until empty
while(!dnodeQueue.isEmpty()) {
dnodeQueue.take().getOutputProtocol().getTransport().close();
}
thriftClientCache.remove(dnode); // to indicate that the DNode is not present
} finally {
thriftClientCacheLock.unlock();
}
}
/**
* Get the Thrift client for this DNode.
*/
public DNodeService.Client getDNodeClientFromPool(String dnode) throws TTransportException {
try {
BlockingQueue dnodeQueue = thriftClientCache.get(dnode);
if(dnodeQueue == null) {
// This shouldn't happen in real life because it is initialized by the QNode, but it is useful for unit
// testing.
// Under some rare race conditions the pool may be required before the QNode creates it, but this method
// assures that the queue will only be created once and, if it's not possible to create it, an exception
// will be thrown and nothing bad will happen.
initializeThriftClientCacheFor(dnode);
dnodeQueue = thriftClientCache.get(dnode);
}
DNodeService.Client client = dnodeQueue.take();
return client;
} catch(InterruptedException e) {
log.error("Interrupted", e);
return null;
}
}
/**
* Return a Thrift client to the pool. This method is a bit tricky since we may want to return a connection when a
* DNode already disconnected. Also, if the QNode is closing, we don't want to leave opened sockets around. To do it
* safely, we check in a loop whether 1) we are closing / cleaning the QNode or 2) the DNode has disconnected.
*/
public void returnDNodeClientToPool(String dnode, DNodeService.Client client, boolean renew) {
if(renew) {
// renew -> try to close if not properly close and set to null
client.getOutputProtocol().getTransport().close();
client = null;
}
do {
if(closing.get()) { // don't return to the pool if the system is already closing! we must close everything!
if(client != null) {
client.getOutputProtocol().getTransport().close();
}
return;
}
BlockingQueue dnodeQueue = thriftClientCache.get(dnode);
if(dnodeQueue == null) {
// dnode is not connected, so we exit.
if(client != null) {
client.getOutputProtocol().getTransport().close();
}
return;
}
if(client == null) { // we have to renew the connection
// endless loop, we need to get a new client otherwise the queue will not have the same size!
try {
client = DNodeClient.get(dnode);
} catch(TTransportException e) {
log.error(
"TTransportException while creating client to renew. Trying again, we can't exit here!", e);
}
} else { // we have a valid client so let's put it into the queue
try {
dnodeQueue.add(client);
} catch(IllegalStateException e) {
client.getOutputProtocol().getTransport().close();
log.error("Trying to return a connection for dnode ["
+ dnode
+ "] but the pool already has the maximum number of connections. This is likely a software bug!.");
}
}
} while(client == null);
// one last check to avoid not closing every socket.
// here we avoid leaking a socket in case a close has happened in parallel or a DNode disconnected right in the
// middle
if(closing.get() || thriftClientCache.get(dnode) == null) {
if(client != null) {
client.getOutputProtocol().getTransport().close();
}
}
}
/**
* Rotates the versions (deletes versions that are old or useless). To be executed at startup and after a deployment.
*/
public List synchronizeTablespaceVersions()
throws InterruptedException {
log.info("Starting to look for old tablespace versions to remove...");
int maxVersionsPerTablespace = config.getInt(QNodeProperties.VERSIONS_PER_TABLESPACE);
// Will contain the list of versions per each tablespace, sorted by creation date descendant
TreeMultimap tablespaces = TreeMultimap.create(Ordering.natural(),
new Comparator() {
@Override
public int compare(Tablespace tb1, Tablespace tb2) {
// reverse ordering. Older dates appears LAST. If same date, then version is compared.
int comp = -((Long) tb1.getCreationDate()).compareTo(tb2.getCreationDate());
if(comp == 0) {
return -((Long) tb1.getVersion()).compareTo(tb2.getVersion());
} else {
return comp;
}
}
});
Map myTablespaces = getTablespaceVersionsMap();
// We build a in memory version of tablespaces for analyzing it
// and prune old ones.
for(Entry entry : myTablespaces.entrySet()) {
tablespaces.put(entry.getKey().getTablespace(), entry.getValue());
}
log.info("Analyzing " + tablespaces.keySet().size() + " tablespaces with a total of " + tablespaces.size() + " versions...");
// We will remove only versions older than the one being served
Map hzVersionsBeingServed = coordinationStructures.getCopyVersionsBeingServed();
if(hzVersionsBeingServed == null) {
log.info("... No versions yet being served.");
return null; // nothing to do yet
}
log.info("Number of versions being served: " + hzVersionsBeingServed.size());
List tablespacesToRemove = new ArrayList();
for(Entry entry : hzVersionsBeingServed.entrySet()) {
String tablespace = entry.getKey();
Long versionBeingServed = entry.getValue();
// Tablespaces are sorted by creation date desc.
SortedSet allVersions = tablespaces.get(tablespace);
Iterator it = allVersions.iterator();
boolean foundVersionBeingServed = false;
int countVersionsAfter = 0;
while(it.hasNext()) {
Tablespace tb = it.next();
if(versionBeingServed.equals(tb.getVersion())) {
foundVersionBeingServed = true;
} else {
if(foundVersionBeingServed) {
countVersionsAfter++;
if(countVersionsAfter >= maxVersionsPerTablespace) {
// This is the case where we remove the version
// 1 - This tablespace has a version being served
// 2 - This version is older than the current tablespace being served
// 3 - We are already keeping maxVersionsPerTablespace versions
tablespacesToRemove.add(new com.splout.db.thrift.TablespaceVersion(tablespace, tb
.getVersion()));
log.info("Tablespace [" + tablespace + "] Version [" + tb.getVersion() + "] "
+ "created at [" + new Date(tb.getCreationDate())
+ "] REMOVED. We already keep younger versions.");
} else {
log.info("Tablespace [" + tablespace + "] Version [" + tb.getVersion() + "] "
+ "created at [" + new Date(tb.getCreationDate())
+ "] KEPT.");
}
} else {
log.info("Tablespace [" + tablespace + "] Version [" + tb.getVersion() + "] "
+ "created at [" + new Date(tb.getCreationDate()) + "] either younger than served one or without version being served. Keeping.");
}
}
}
if(!foundVersionBeingServed) {
log.info("Tablespace [" + tablespace
+ "] without any version being served. Please, have a look, and remove them if not used");
}
if(tablespacesToRemove.size() > 0) {
log.info("Sending [" + tablespacesToRemove + "] to all alive DNodes.");
for(DNodeInfo dnode : coordinationStructures.getDNodes().values()) {
DNodeService.Client client = null;
boolean renew = false;
try {
client = getDNodeClientFromPool(dnode.getAddress());
client.deleteOldVersions(tablespacesToRemove);
} catch(TTransportException e) {
renew = true;
log.warn("Failed sending delete TablespaceVersions order to (" + dnode
+ "). Not critical as they will be removed after other deployments.", e);
} catch(DNodeException e) {
log.warn("Failed sending delete TablespaceVersions order to (" + dnode
+ "). Not critical as they will be removed after other deployments.", e);
} catch(TException e) {
log.warn("Failed sending delete TablespaceVersions order to (" + dnode
+ "). Not critical as they will be removed after other deployments.", e);
} finally {
if(client != null) {
returnDNodeClientToPool(dnode.getAddress(), client, renew);
}
}
}
}
log.info("... done looking for old tablespace versions to remove...");
}
return tablespacesToRemove; // Return for unit test
}
private AtomicBoolean closing = new AtomicBoolean(false);
public void close() {
closing.set(true); // will indicate other parts of this code that things have to be closed!
for(Map.Entry> entry : thriftClientCache.entrySet()) {
while(entry.getValue().size() > 0) {
try {
entry.getValue().take().getOutputProtocol().getTransport().close();
} catch(InterruptedException e) {
log.error("Interrupted!", e);
}
}
}
}
public void maybeBalance() {
// do this only after warming
if(!isWarming.get() && config.getBoolean(QNodeProperties.REPLICA_BALANCE_ENABLE)) {
// check if we could balance some partitions
List balanceActions = getBalanceActions();
// we will only re-balance versions being served
// otherwise strange things may happen: to re-balance a version in the middle of its deployment...
Map versionsBeingServed = coordinationStructures.getCopyVersionsBeingServed();
for(ReplicaBalancer.BalanceAction action : balanceActions) {
if(versionsBeingServed != null && versionsBeingServed.get(action.getTablespace()) != null
&& versionsBeingServed.get(action.getTablespace()) == action.getVersion()) {
// put if absent + TTL
coordinationStructures.getDNodeReplicaBalanceActionsSet().putIfAbsent(action, "",
config.getLong(QNodeProperties.BALANCE_ACTIONS_TTL), TimeUnit.SECONDS);
}
}
}
}
// ---- Getters ---- //
public Map getCurrentVersionsMap() {
return currentVersionsMap;
}
public Map getTablespaceVersionsMap() {
return tablespaceVersionsMap;
}
public CoordinationStructures getCoordinationStructures() {
return coordinationStructures;
}
public SploutConfiguration getConfig() {
return config;
}
public ConcurrentMap> getThriftClientCache() {
return thriftClientCache;
}
public AtomicBoolean getIsWarming() {
return isWarming;
}
}