Flink - RocksDBStateBackend

如果要考虑易用性和效率，使用rocksDB来替代普通内存的kv是有必要的

有了rocksdb，可以range查询，可以支持columnfamily，可以各种压缩

但是rocksdb本身是一个库，是跑在RocksDBStateBackend中的

所以taskmanager挂掉后，数据还是没了，

所以RocksDBStateBackend仍然需要类似HDFS这样的分布式存储来存储snapshot

kv state需要由rockdb来管理，这是和内存或file backend最大的不同

AbstractRocksDBState

/**
 * Base class for {@link State} implementations that store state in a RocksDB database.
 *
 * <p>State is not stored in this class but in the {@link org.rocksdb.RocksDB} instance that
 * the {@link RocksDBStateBackend} manages and checkpoints.
 *
 * @param <K> The type of the key.
 * @param <N> The type of the namespace.
 * @param <S> The type of {@link State}.
 * @param <SD> The type of {@link StateDescriptor}.
 */
public abstract class AbstractRocksDBState<K, N, S extends State, SD extends StateDescriptor<S, ?>>
        implements KvState<K, N, S, SD, RocksDBStateBackend>, State {

    /** Serializer for the namespace */
    private final TypeSerializer<N> namespaceSerializer;

    /** The current namespace, which the next value methods will refer to */
    private N currentNamespace;

    /** Backend that holds the actual RocksDB instance where we store state */
    protected RocksDBStateBackend backend;

    /** The column family of this particular instance of state */
    protected ColumnFamilyHandle columnFamily;

    /**
     * We disable writes to the write-ahead-log here.
     */
    private final WriteOptions writeOptions;

    /**
     * Creates a new RocksDB backed state.
     *
     * @param namespaceSerializer The serializer for the namespace.
     */
    protected AbstractRocksDBState(ColumnFamilyHandle columnFamily,
            TypeSerializer<N> namespaceSerializer,
            RocksDBStateBackend backend) {

        this.namespaceSerializer = namespaceSerializer;
        this.backend = backend;

        this.columnFamily = columnFamily;

        writeOptions = new WriteOptions();
        writeOptions.setDisableWAL(true);
    }

    @Override
    public KvStateSnapshot<K, N, S, SD, RocksDBStateBackend> snapshot(long checkpointId,
            long timestamp) throws Exception {
        throw new RuntimeException("Should not be called. Backups happen in RocksDBStateBackend.");
    }
}

RocksDBValueState

/**
 * {@link ValueState} implementation that stores state in RocksDB.
 *
 * @param <K> The type of the key.
 * @param <N> The type of the namespace.
 * @param <V> The type of value that the state state stores.
 */
public class RocksDBValueState<K, N, V>
    extends AbstractRocksDBState<K, N, ValueState<V>, ValueStateDescriptor<V>>
    implements ValueState<V> {

    @Override
    public V value() {
        ByteArrayOutputStream baos = new ByteArrayOutputStream();
        DataOutputViewStreamWrapper out = new DataOutputViewStreamWrapper(baos);
        try {
            writeKeyAndNamespace(out);
            byte[] key = baos.toByteArray();
            byte[] valueBytes = backend.db.get(columnFamily, key); //从db读出value
            if (valueBytes == null) {
                return stateDesc.getDefaultValue();
            }
            return valueSerializer.deserialize(new DataInputViewStreamWrapper(new ByteArrayInputStream(valueBytes)));
        } catch (IOException|RocksDBException e) {
            throw new RuntimeException("Error while retrieving data from RocksDB.", e);
        }
    }

    @Override
    public void update(V value) throws IOException {
        if (value == null) {
            clear();
            return;
        }
        ByteArrayOutputStream baos = new ByteArrayOutputStream();
        DataOutputViewStreamWrapper out = new DataOutputViewStreamWrapper(baos);
        try {
            writeKeyAndNamespace(out);
            byte[] key = baos.toByteArray();
            baos.reset();
            valueSerializer.serialize(value, out);
            backend.db.put(columnFamily, writeOptions, key, baos.toByteArray()); //将kv写入db
        } catch (Exception e) {
            throw new RuntimeException("Error while adding data to RocksDB", e);
        }
    }
}

因为对于kv state，key就是当前收到数据的key，所以key是直接从backend.currentKey()中读到；参考，Flink - Working with State

RocksDBStateBackend

初始化过程，

/**
 * A {@link StateBackend} that stores its state in {@code RocksDB}. This state backend can
 * store very large state that exceeds memory and spills to disk.
 *
 * <p>All key/value state (including windows) is stored in the key/value index of RocksDB.
 * For persistence against loss of machines, checkpoints take a snapshot of the
 * RocksDB database, and persist that snapshot in a file system (by default) or
 * another configurable state backend.
 *
 * <p>The behavior of the RocksDB instances can be parametrized by setting RocksDB Options
 * using the methods {@link #setPredefinedOptions(PredefinedOptions)} and
 * {@link #setOptions(OptionsFactory)}.
 */
public class RocksDBStateBackend extends AbstractStateBackend {

    // ------------------------------------------------------------------------
    //  Static configuration values
    // ------------------------------------------------------------------------

    /** The checkpoint directory that we copy the RocksDB backups to. */
    private final Path checkpointDirectory;

    /** The state backend that stores the non-partitioned state */
    private final AbstractStateBackend nonPartitionedStateBackend;

    /**
     * Our RocksDB data base, this is used by the actual subclasses of {@link AbstractRocksDBState}
     * to store state. The different k/v states that we have don't each have their own RocksDB
     * instance. They all write to this instance but to their own column family.
     */
    protected volatile transient RocksDB db; //RocksDB实例

    /**
     * Creates a new {@code RocksDBStateBackend} that stores its checkpoint data in the
     * file system and location defined by the given URI.
     *
     * <p>A state backend that stores checkpoints in HDFS or S3 must specify the file system
     * host and port in the URI, or have the Hadoop configuration that describes the file system
     * (host / high-availability group / possibly credentials) either referenced from the Flink
     * config, or included in the classpath.
     *
     * @param checkpointDataUri The URI describing the filesystem and path to the checkpoint data directory.
     * @throws IOException Thrown, if no file system can be found for the scheme in the URI.
     */
    public RocksDBStateBackend(String checkpointDataUri) throws IOException {
        this(new Path(checkpointDataUri).toUri());
    }

    /**
     * Creates a new {@code RocksDBStateBackend} that stores its checkpoint data in the
     * file system and location defined by the given URI.
     *
     * <p>A state backend that stores checkpoints in HDFS or S3 must specify the file system
     * host and port in the URI, or have the Hadoop configuration that describes the file system
     * (host / high-availability group / possibly credentials) either referenced from the Flink
     * config, or included in the classpath.
     *
     * @param checkpointDataUri The URI describing the filesystem and path to the checkpoint data directory.
     * @throws IOException Thrown, if no file system can be found for the scheme in the URI.
     */
    public RocksDBStateBackend(URI checkpointDataUri) throws IOException {
        // creating the FsStateBackend automatically sanity checks the URI
        FsStateBackend fsStateBackend = new FsStateBackend(checkpointDataUri); //仍然使用FsStateBackend来存snapshot

        this.nonPartitionedStateBackend = fsStateBackend;
        this.checkpointDirectory = fsStateBackend.getBasePath();
    }

    // ------------------------------------------------------------------------
    //  State backend methods
    // ------------------------------------------------------------------------

    @Override
    public void initializeForJob(
            Environment env,
            String operatorIdentifier,
            TypeSerializer<?> keySerializer) throws Exception {

        super.initializeForJob(env, operatorIdentifier, keySerializer);

        this.nonPartitionedStateBackend.initializeForJob(env, operatorIdentifier, keySerializer);

        RocksDB.loadLibrary(); //初始化rockdb

        List<ColumnFamilyDescriptor> columnFamilyDescriptors = new ArrayList<>(1); //columnFamily的概念和HBase相同，放在独立的文件
        // RocksDB seems to need this...
        columnFamilyDescriptors.add(new ColumnFamilyDescriptor("default".getBytes()));
        List<ColumnFamilyHandle> columnFamilyHandles = new ArrayList<>(1);
        try {
            db = RocksDB.open(getDbOptions(), instanceRocksDBPath.getAbsolutePath(), columnFamilyDescriptors, columnFamilyHandles); //真正的open rocksDB
        } catch (RocksDBException e) {
            throw new RuntimeException("Error while opening RocksDB instance.", e);
        }
    }

snapshotPartitionedState

@Override
public HashMap<String, KvStateSnapshot<?, ?, ?, ?, ?>> snapshotPartitionedState(long checkpointId, long timestamp) throws Exception {
    if (keyValueStatesByName == null || keyValueStatesByName.size() == 0) {
        return new HashMap<>();
    }

    if (fullyAsyncBackup) {
        return performFullyAsyncSnapshot(checkpointId, timestamp);
    } else {
        return performSemiAsyncSnapshot(checkpointId, timestamp);
    }
}

snapshot分为全异步和半异步两种，

半异步，

/**
 * Performs a checkpoint by using the RocksDB backup feature to backup to a directory.
 * This backup is the asynchronously copied to the final checkpoint location.
 */
private HashMap<String, KvStateSnapshot<?, ?, ?, ?, ?>> performSemiAsyncSnapshot(long checkpointId, long timestamp) throws Exception {
    // We don't snapshot individual k/v states since everything is stored in a central
    // RocksDB data base. Create a dummy KvStateSnapshot that holds the information about
    // that checkpoint. We use the in injectKeyValueStateSnapshots to restore.

    final File localBackupPath = new File(instanceBasePath, "local-chk-" + checkpointId);
    final URI backupUri = new URI(instanceCheckpointPath + "/chk-" + checkpointId);

    long startTime = System.currentTimeMillis();

    BackupableDBOptions backupOptions = new BackupableDBOptions(localBackupPath.getAbsolutePath());
    // we disabled the WAL
    backupOptions.setBackupLogFiles(false);
    // no need to sync since we use the backup only as intermediate data before writing to FileSystem snapshot
    backupOptions.setSync(false); //设为异步

    try (BackupEngine backupEngine = BackupEngine.open(Env.getDefault(), backupOptions)) {
        // wait before flush with "true"
        backupEngine.createNewBackup(db, true); //利用rocksDB自己的backupEngine生成新的backup，存在本地磁盘
    }

    long endTime = System.currentTimeMillis(); //这部分是同步做的，需要计时看延时
    LOG.info("RocksDB (" + instanceRocksDBPath + ") backup (synchronous part) took " + (endTime - startTime) + " ms.");

    // draw a copy in case it get's changed while performing the async snapshot
    List<StateDescriptor> kvStateInformationCopy = new ArrayList<>();
    for (Tuple2<ColumnFamilyHandle, StateDescriptor> state: kvStateInformation.values()) {
        kvStateInformationCopy.add(state.f1);
    }
    SemiAsyncSnapshot dummySnapshot = new SemiAsyncSnapshot(localBackupPath, //
            backupUri,
            kvStateInformationCopy,
            checkpointId);

    HashMap<String, KvStateSnapshot<?, ?, ?, ?, ?>> result = new HashMap<>();
    result.put("dummy_state", dummySnapshot);
    return result;
}

SemiAsyncSnapshot.materialize

@Override
public KvStateSnapshot<Object, Object, ValueState<Object>, ValueStateDescriptor<Object>, RocksDBStateBackend> materialize() throws Exception {
    try {
        long startTime = System.currentTimeMillis();
        HDFSCopyFromLocal.copyFromLocal(localBackupPath, backupUri);  //从本地磁盘copy到hdfs
        long endTime = System.currentTimeMillis();
        LOG.info("RocksDB materialization from " + localBackupPath + " to " + backupUri + " (asynchronous part) took " + (endTime - startTime) + " ms.");
        return new FinalSemiAsyncSnapshot(backupUri, checkpointId, stateDescriptors);
    } catch (Exception e) {
        FileSystem fs = FileSystem.get(backupUri, HadoopFileSystem.getHadoopConfiguration());
        fs.delete(new org.apache.hadoop.fs.Path(backupUri), true);
        throw e;
    } finally {
        FileUtils.deleteQuietly(localBackupPath);
    }
}

全异步

/**
 * Performs a checkpoint by drawing a {@link org.rocksdb.Snapshot} from RocksDB and then
 * iterating over all key/value pairs in RocksDB to store them in the final checkpoint
 * location. The only synchronous part is the drawing of the {@code Snapshot} which
 * is essentially free.
 */
private HashMap<String, KvStateSnapshot<?, ?, ?, ?, ?>> performFullyAsyncSnapshot(long checkpointId, long timestamp) throws Exception {
    // we draw a snapshot from RocksDB then iterate over all keys at that point
    // and store them in the backup location

    final URI backupUri = new URI(instanceCheckpointPath + "/chk-" + checkpointId);

    long startTime = System.currentTimeMillis();

    org.rocksdb.Snapshot snapshot = db.getSnapshot(); //生成snapshot，但不用落盘

    long endTime = System.currentTimeMillis();
    LOG.info("Fully asynchronous RocksDB (" + instanceRocksDBPath + ") backup (synchronous part) took " + (endTime - startTime) + " ms.");

    // draw a copy in case it get's changed while performing the async snapshot
    Map<String, Tuple2<ColumnFamilyHandle, StateDescriptor>> columnFamiliesCopy = new HashMap<>();
    columnFamiliesCopy.putAll(kvStateInformation);
    FullyAsyncSnapshot dummySnapshot = new FullyAsyncSnapshot(snapshot, //直接把snapshot传入
            this,
            backupUri,
            columnFamiliesCopy,
            checkpointId);

    HashMap<String, KvStateSnapshot<?, ?, ?, ?, ?>> result = new HashMap<>();
    result.put("dummy_state", dummySnapshot);
    return result;
}

FullyAsyncSnapshot.materialize

可以看到需要自己去做db内容的序列化到文件的过程

@Override
public KvStateSnapshot<Object, Object, ValueState<Object>, ValueStateDescriptor<Object>, RocksDBStateBackend> materialize() throws Exception {
    try {
        long startTime = System.currentTimeMillis();

        CheckpointStateOutputView outputView = backend.createCheckpointStateOutputView(checkpointId, startTime);

        outputView.writeInt(columnFamilies.size());

        // we don't know how many key/value pairs there are in each column family.
        // We prefix every written element with a byte that signifies to which
        // column family it belongs, this way we can restore the column families
        byte count = 0;
        Map<String, Byte> columnFamilyMapping = new HashMap<>();
        for (Map.Entry<String, Tuple2<ColumnFamilyHandle, StateDescriptor>> column: columnFamilies.entrySet()) {
            columnFamilyMapping.put(column.getKey(), count);

            outputView.writeByte(count);

            ObjectOutputStream ooOut = new ObjectOutputStream(outputView);
            ooOut.writeObject(column.getValue().f1);
            ooOut.flush();

            count++;
        }

        ReadOptions readOptions = new ReadOptions();
        readOptions.setSnapshot(snapshot);

        for (Map.Entry<String, Tuple2<ColumnFamilyHandle, StateDescriptor>> column: columnFamilies.entrySet()) {
            byte columnByte = columnFamilyMapping.get(column.getKey());

            synchronized (dbCleanupLock) {
                if (db == null) {
                    throw new RuntimeException("RocksDB instance was disposed. This happens " +
                            "when we are in the middle of a checkpoint and the job fails.");
                }
                RocksIterator iterator = db.newIterator(column.getValue().f0, readOptions);
                iterator.seekToFirst();
                while (iterator.isValid()) {
                    outputView.writeByte(columnByte);
                    BytePrimitiveArraySerializer.INSTANCE.serialize(iterator.key(),
                            outputView);
                    BytePrimitiveArraySerializer.INSTANCE.serialize(iterator.value(),
                            outputView);
                    iterator.next();
                }
            }
        }

        StateHandle<DataInputView> stateHandle = outputView.closeAndGetHandle();

        long endTime = System.currentTimeMillis();
        LOG.info("Fully asynchronous RocksDB materialization to " + backupUri + " (asynchronous part) took " + (endTime - startTime) + " ms.");
        return new FinalFullyAsyncSnapshot(stateHandle, checkpointId);
    } finally {
        synchronized (dbCleanupLock) {
            if (db != null) {
                db.releaseSnapshot(snapshot);
            }
        }
        snapshot = null;
    }
}

CheckpointStateOutputView

backend.createCheckpointStateOutputView

public CheckpointStateOutputView createCheckpointStateOutputView(
        long checkpointID, long timestamp) throws Exception {
    return new CheckpointStateOutputView(createCheckpointStateOutputStream(checkpointID, timestamp));
}

关键createCheckpointStateOutputStream

RocksDBStateBackend

@Override
public CheckpointStateOutputStream createCheckpointStateOutputStream(
        long checkpointID, long timestamp) throws Exception {

    return nonPartitionedStateBackend.createCheckpointStateOutputStream(checkpointID, timestamp);
}

看看nonPartitionedStateBackend是什么？

public RocksDBStateBackend(URI checkpointDataUri) throws IOException {
    // creating the FsStateBackend automatically sanity checks the URI
    FsStateBackend fsStateBackend = new FsStateBackend(checkpointDataUri);

    this.nonPartitionedStateBackend = fsStateBackend;
    this.checkpointDirectory = fsStateBackend.getBasePath();
}

其实就是FsStateBackend，最终rocksDB还是要用FsStateBackend来存储snapshot

restoreState

@Override
public final void injectKeyValueStateSnapshots(HashMap<String, KvStateSnapshot> keyValueStateSnapshots) throws Exception {
    if (keyValueStateSnapshots.size() == 0) {
        return;
    }

    KvStateSnapshot dummyState = keyValueStateSnapshots.get("dummy_state");
    if (dummyState instanceof FinalSemiAsyncSnapshot) {
        restoreFromSemiAsyncSnapshot((FinalSemiAsyncSnapshot) dummyState);
    } else if (dummyState instanceof FinalFullyAsyncSnapshot) {
        restoreFromFullyAsyncSnapshot((FinalFullyAsyncSnapshot) dummyState);
    } else {
        throw new RuntimeException("Unknown RocksDB snapshot: " + dummyState);
    }
}

同样也分为两种，半异步和全异步，过程基本就是snapshot的逆过程