Examples with ReplicationPeerImpl org.apache.hadoop.hbase.replication.ReplicationPeerImpl used on opensource projects

Example 1 with ReplicationPeerImpl

use of org.apache.hadoop.hbase.replication.ReplicationPeerImpl in project hbase by apache.

the class SplitLogWorker method processSyncReplicationWAL.

// returns whether we need to continue the split work
private static boolean processSyncReplicationWAL(String name, Configuration conf, RegionServerServices server, FileSystem fs, Path walDir) throws IOException {
    Path walFile = new Path(walDir, name);
    String filename = walFile.getName();
    Optional<String> optSyncPeerId = SyncReplicationWALProvider.getSyncReplicationPeerIdFromWALName(filename);
    if (!optSyncPeerId.isPresent()) {
        return true;
    }
    String peerId = optSyncPeerId.get();
    ReplicationPeerImpl peer = server.getReplicationSourceService().getReplicationPeers().getPeer(peerId);
    if (peer == null || !peer.getPeerConfig().isSyncReplication()) {
        return true;
    }
    Pair<SyncReplicationState, SyncReplicationState> stateAndNewState = peer.getSyncReplicationStateAndNewState();
    if (stateAndNewState.getFirst().equals(SyncReplicationState.ACTIVE) && stateAndNewState.getSecond().equals(SyncReplicationState.NONE)) {
        // copy the file to remote and overwrite the previous one
        String remoteWALDir = peer.getPeerConfig().getRemoteWALDir();
        Path remoteWALDirForPeer = ReplicationUtils.getPeerRemoteWALDir(remoteWALDir, peerId);
        Path tmpRemoteWAL = new Path(remoteWALDirForPeer, filename + ".tmp");
        FileSystem remoteFs = ReplicationUtils.getRemoteWALFileSystem(conf, remoteWALDir);
        try (FSDataInputStream in = fs.open(walFile);
            FSDataOutputStream out = remoteFs.createNonRecursive(tmpRemoteWAL, true, CommonFSUtils.getDefaultBufferSize(remoteFs), remoteFs.getDefaultReplication(tmpRemoteWAL), remoteFs.getDefaultBlockSize(tmpRemoteWAL), null)) {
            IOUtils.copy(in, out);
        }
        Path toCommitRemoteWAL = new Path(remoteWALDirForPeer, filename + ReplicationUtils.RENAME_WAL_SUFFIX);
        // Some FileSystem implementations may not support atomic rename so we need to do it in two
        // phases
        FSUtils.renameFile(remoteFs, tmpRemoteWAL, toCommitRemoteWAL);
        FSUtils.renameFile(remoteFs, toCommitRemoteWAL, new Path(remoteWALDirForPeer, filename));
    } else if ((stateAndNewState.getFirst().equals(SyncReplicationState.ACTIVE) && stateAndNewState.getSecond().equals(SyncReplicationState.STANDBY)) || stateAndNewState.getFirst().equals(SyncReplicationState.STANDBY)) {
        // check whether we still need to process this file
        // actually we only write wal file which name is ended with .syncrep in A state, and after
        // transiting to a state other than A, we will reopen all the regions so the data in the wal
        // will be flushed so the wal file will be archived soon. But it is still possible that there
        // is a server crash when we are transiting from A to S, to simplify the logic of the transit
        // procedure, here we will also check the remote snapshot directory in state S, so that we do
        // not need wait until all the wal files with .syncrep suffix to be archived before finishing
        // the procedure.
        String remoteWALDir = peer.getPeerConfig().getRemoteWALDir();
        Path remoteSnapshotDirForPeer = ReplicationUtils.getPeerSnapshotWALDir(remoteWALDir, peerId);
        FileSystem remoteFs = ReplicationUtils.getRemoteWALFileSystem(conf, remoteWALDir);
        if (remoteFs.exists(new Path(remoteSnapshotDirForPeer, filename))) {
            // the file has been replayed when the remote cluster was transited from S to DA, the
            // content will be replicated back to us so give up split it.
            LOG.warn("Giveup splitting {} since it has been replayed in the remote cluster and " + "the content will be replicated back", filename);
            return false;
        }
    }
    return true;
}

Example 2 with ReplicationPeerImpl

use of org.apache.hadoop.hbase.replication.ReplicationPeerImpl in project hbase by apache.

the class PeerProcedureHandlerImpl method transitSyncReplicationPeerState.

@Override
public void transitSyncReplicationPeerState(String peerId, int stage, HRegionServer rs) throws ReplicationException, IOException {
    ReplicationPeers replicationPeers = replicationSourceManager.getReplicationPeers();
    Lock peerLock = peersLock.acquireLock(peerId);
    try {
        ReplicationPeerImpl peer = replicationPeers.getPeer(peerId);
        if (peer == null) {
            throw new ReplicationException("Peer with id=" + peerId + " is not cached.");
        }
        if (!peer.getPeerConfig().isSyncReplication()) {
            throw new ReplicationException("Peer with id=" + peerId + " is not synchronous.");
        }
        SyncReplicationState newSyncReplicationState = peer.getNewSyncReplicationState();
        if (stage == 0) {
            if (newSyncReplicationState != SyncReplicationState.NONE) {
                LOG.warn("The new sync replication state for peer {} has already been set to {}, " + "this should be a retry, give up", peerId, newSyncReplicationState);
                return;
            }
            // refresh the peer state first, as when we transit to STANDBY, we may need to disable the
            // peer before processing the sync replication state.
            PeerState oldState = peer.getPeerState();
            boolean success = false;
            try {
                PeerState newState = replicationPeers.refreshPeerState(peerId);
                if (oldState.equals(PeerState.ENABLED) && newState.equals(PeerState.DISABLED)) {
                    replicationSourceManager.refreshSources(peerId);
                }
                success = true;
            } finally {
                if (!success) {
                    peer.setPeerState(oldState.equals(PeerState.ENABLED));
                }
            }
            newSyncReplicationState = replicationPeers.refreshPeerNewSyncReplicationState(peerId);
            SyncReplicationState oldSyncReplicationState = peer.getSyncReplicationState();
            peerActionListener.peerSyncReplicationStateChange(peerId, oldSyncReplicationState, newSyncReplicationState, stage);
        } else {
            if (newSyncReplicationState == SyncReplicationState.NONE) {
                LOG.warn("The new sync replication state for peer {} has already been clear, and the " + "current state is {}, this should be a retry, give up", peerId, newSyncReplicationState);
                return;
            }
            if (newSyncReplicationState == SyncReplicationState.STANDBY) {
                replicationSourceManager.drainSources(peerId);
                // Need to roll the wals and make the ReplicationSource for this peer track the new file.
                // If we do not do this, there will be two problems that can not be addressed at the same
                // time. First, if we just throw away the current wal file, and later when we transit the
                // peer to DA, and the wal has not been rolled yet, then the new data written to the wal
                // file will not be replicated and cause data inconsistency. But if we just track the
                // current wal file without rolling, it may contains some data before we transit the peer
                // to S, later if we transit the peer to DA, the data will also be replicated and cause
                // data inconsistency. So here we need to roll the wal, and let the ReplicationSource
                // track the new wal file, and throw the old wal files away.
                LogRoller roller = rs.getWalRoller();
                roller.requestRollAll();
                try {
                    roller.waitUntilWalRollFinished();
                } catch (InterruptedException e) {
                    // reset the interrupted flag
                    Thread.currentThread().interrupt();
                    throw (IOException) new InterruptedIOException("Interrupted while waiting for wal roll finish").initCause(e);
                }
            }
            SyncReplicationState oldState = peer.getSyncReplicationState();
            peerActionListener.peerSyncReplicationStateChange(peerId, oldState, newSyncReplicationState, stage);
            peer.transitSyncReplicationState();
        }
    } finally {
        peerLock.unlock();
    }
}

Example 3 with ReplicationPeerImpl

use of org.apache.hadoop.hbase.replication.ReplicationPeerImpl in project hbase by apache.

the class ReplicationSourceManager method claimQueue.

void claimQueue(ServerName deadRS, String queue) {
    // This sleep may not be enough in some cases.
    try {
        Thread.sleep(sleepBeforeFailover + (long) (ThreadLocalRandom.current().nextFloat() * sleepBeforeFailover));
    } catch (InterruptedException e) {
        LOG.warn("Interrupted while waiting before transferring a queue.");
        Thread.currentThread().interrupt();
    }
    // We try to lock that rs' queue directory
    if (server.isStopped()) {
        LOG.info("Not transferring queue since we are shutting down");
        return;
    }
    // After claim the queues from dead region server, wewill skip to start the
    // RecoveredReplicationSource if the peer has been removed. but there's possible that remove a
    // peer with peerId = 2 and add a peer with peerId = 2 again during failover. So we need to get
    // a copy of the replication peer first to decide whether we should start the
    // RecoveredReplicationSource. If the latest peer is not the old peer, we should also skip to
    // start the RecoveredReplicationSource, Otherwise the rs will abort (See HBASE-20475).
    String peerId = new ReplicationQueueInfo(queue).getPeerId();
    ReplicationPeerImpl oldPeer = replicationPeers.getPeer(peerId);
    if (oldPeer == null) {
        LOG.info("Not transferring queue since the replication peer {} for queue {} does not exist", peerId, queue);
        return;
    }
    Pair<String, SortedSet<String>> claimedQueue;
    try {
        claimedQueue = queueStorage.claimQueue(deadRS, queue, server.getServerName());
    } catch (ReplicationException e) {
        LOG.error("ReplicationException: cannot claim dead region ({})'s " + "replication queue. Znode : ({})" + " Possible solution: check if znode size exceeds jute.maxBuffer value. " + " If so, increase it for both client and server side.", deadRS, queueStorage.getRsNode(deadRS), e);
        server.abort("Failed to claim queue from dead regionserver.", e);
        return;
    }
    if (claimedQueue.getSecond().isEmpty()) {
        return;
    }
    String queueId = claimedQueue.getFirst();
    Set<String> walsSet = claimedQueue.getSecond();
    ReplicationPeerImpl peer = replicationPeers.getPeer(peerId);
    if (peer == null || peer != oldPeer) {
        LOG.warn("Skipping failover for peer {} of node {}, peer is null", peerId, deadRS);
        abortWhenFail(() -> queueStorage.removeQueue(server.getServerName(), queueId));
        return;
    }
    if (server instanceof ReplicationSyncUp.DummyServer && peer.getPeerState().equals(PeerState.DISABLED)) {
        LOG.warn("Peer {} is disabled. ReplicationSyncUp tool will skip " + "replicating data to this peer.", peerId);
        return;
    }
    ReplicationSourceInterface src;
    try {
        src = createSource(queueId, peer);
    } catch (IOException e) {
        LOG.error("Can not create replication source for peer {} and queue {}", peerId, queueId, e);
        server.abort("Failed to create replication source after claiming queue.", e);
        return;
    }
    // synchronized on oldsources to avoid adding recovered source for the to-be-removed peer
    synchronized (oldsources) {
        peer = replicationPeers.getPeer(src.getPeerId());
        if (peer == null || peer != oldPeer) {
            src.terminate("Recovered queue doesn't belong to any current peer");
            deleteQueue(queueId);
            return;
        }
        // replicated back.
        if (peer.getPeerConfig().isSyncReplication()) {
            Pair<SyncReplicationState, SyncReplicationState> stateAndNewState = peer.getSyncReplicationStateAndNewState();
            if ((stateAndNewState.getFirst().equals(SyncReplicationState.STANDBY) && stateAndNewState.getSecond().equals(SyncReplicationState.NONE)) || stateAndNewState.getSecond().equals(SyncReplicationState.STANDBY)) {
                src.terminate("Sync replication peer is in STANDBY state");
                deleteQueue(queueId);
                return;
            }
        }
        // track sources in walsByIdRecoveredQueues
        Map<String, NavigableSet<String>> walsByGroup = new HashMap<>();
        walsByIdRecoveredQueues.put(queueId, walsByGroup);
        for (String wal : walsSet) {
            String walPrefix = AbstractFSWALProvider.getWALPrefixFromWALName(wal);
            NavigableSet<String> wals = walsByGroup.get(walPrefix);
            if (wals == null) {
                wals = new TreeSet<>();
                walsByGroup.put(walPrefix, wals);
            }
            wals.add(wal);
        }
        oldsources.add(src);
        LOG.info("Added source for recovered queue {}", src.getQueueId());
        for (String wal : walsSet) {
            LOG.trace("Enqueueing log from recovered queue for source: " + src.getQueueId());
            src.enqueueLog(new Path(oldLogDir, wal));
        }
        src.startup();
    }
}

Example 4 with ReplicationPeerImpl

use of org.apache.hadoop.hbase.replication.ReplicationPeerImpl in project hbase by apache.

the class PeerProcedureHandlerImpl method updatePeerConfig.

@Override
public void updatePeerConfig(String peerId) throws ReplicationException, IOException {
    Lock peerLock = peersLock.acquireLock(peerId);
    ReplicationPeers peers = replicationSourceManager.getReplicationPeers();
    ReplicationPeerImpl peer = null;
    ReplicationPeerConfig oldConfig = null;
    PeerState oldState = null;
    boolean success = false;
    try {
        peer = peers.getPeer(peerId);
        if (peer == null) {
            throw new ReplicationException("Peer with id=" + peerId + " is not cached.");
        }
        oldConfig = peer.getPeerConfig();
        oldState = peer.getPeerState();
        ReplicationPeerConfig newConfig = peers.refreshPeerConfig(peerId);
        // also need to refresh peer state here. When updating a serial replication peer we may
        // disable it first and then enable it.
        PeerState newState = peers.refreshPeerState(peerId);
        // RS need to start work with the new replication config change
        if (!ReplicationUtils.isNamespacesAndTableCFsEqual(oldConfig, newConfig) || oldConfig.isSerial() != newConfig.isSerial() || (oldState.equals(PeerState.ENABLED) && newState.equals(PeerState.DISABLED))) {
            replicationSourceManager.refreshSources(peerId);
        }
        success = true;
    } finally {
        if (!success && peer != null) {
            // Reset peer config if refresh source failed
            peer.setPeerConfig(oldConfig);
            peer.setPeerState(oldState.equals(PeerState.ENABLED));
        }
        peerLock.unlock();
    }
}

Example 5 with ReplicationPeerImpl

use of org.apache.hadoop.hbase.replication.ReplicationPeerImpl in project hbase by apache.

the class PeerProcedureHandlerImpl method refreshPeerState.

private void refreshPeerState(String peerId) throws ReplicationException, IOException {
    Lock peerLock = peersLock.acquireLock(peerId);
    ReplicationPeerImpl peer = null;
    PeerState oldState = null;
    boolean success = false;
    try {
        peer = replicationSourceManager.getReplicationPeers().getPeer(peerId);
        if (peer == null) {
            throw new ReplicationException("Peer with id=" + peerId + " is not cached.");
        }
        oldState = peer.getPeerState();
        PeerState newState = replicationSourceManager.getReplicationPeers().refreshPeerState(peerId);
        // RS need to start work with the new replication state change
        if (oldState.equals(PeerState.ENABLED) && newState.equals(PeerState.DISABLED)) {
            replicationSourceManager.refreshSources(peerId);
        }
        success = true;
    } finally {
        if (!success && peer != null) {
            // Reset peer state if refresh source failed
            peer.setPeerState(oldState.equals(PeerState.ENABLED));
        }
        peerLock.unlock();
    }
}