Postgres 9.4 (released in December 2014) introduced a new feature called logical replication. Logical replication allows changes from a database to be streamed in real-time to an external system. The difference between physical replication and logical replication is that logical replication sends data over in a logical format whereas physical replication sends data over in a binary format. Additionally logical replication can send over a single table, or database. Binary replication replicates the entire cluster in an all or nothing fashion; which is to say there is no way to get a specific table or database using binary replication

Prior to logical replication keeping an external system synchronized in real time was problematic. The application would have to update/invalidate the appropriate cache entries, reindex the data in your search engine, send it to your analytics system, and so on. This suffers from race conditions and reliability problems. For example if slightly different data gets written to two different datastores (perhaps due to a bug or a race condition),the contents of the datastores will gradually drift apart — they will become more and more inconsistent over time. Recovering from such gradual data corruption is difficult.

Logical decoding takes the database's write-ahead log (WAL), and gives us access to row-level change events: every time a row in a table is inserted, updated or deleted, that's an event. Those events are grouped by transaction, and appear in the order in which they were committed to the database. Aborted/rolled-back transactions do not appear in the stream. Thus, if you apply the change events in the same order, you end up with an exact, transactionally consistent copy of the database. It's looks like the Event Sourcing pattern that you previously implemented in your application, but now it's available out of the box from the LightDB database.

For access to real-time changes LightDB provides the streaming replication protocol. Replication protocol can be physical or logical. Physical replication protocol is used for Master/Secondary replication. Logical replication protocol can be used to stream changes to an external system.

Since the JDBC API does not include replication PGConnection implements the LightDB API

Configure database

lightdb.conf

lt_hba.conf

Configuration for examples

max_wal_senders = 4             # max number of walsender processes
wal_keep_segments = 4           # in logfile segments, 16MB each; 0 disables
wal_level = logical             # minimal, replica, or logical
max_replication_slots = 4       # max number of replication slots

Logical replication uses a replication slot to reserve WAL logs on the server and also defines which decoding plugin to use to decode the WAL logs to the required format, for example you can decode changes as json, protobuf, etc. To demonstrate how to use the ltjdbc replication API we will use the test_decoding plugin that is include in the lightdb-contrib package, but you can use your own decoding plugin. There are a few on github which can be used as examples.

In order to use the replication API, the Connection has to be created in replication mode, in this mode the connection is not available to execute SQL commands, and can only be used with replication API. This is a restriction imposed by LightDB.

    String url = "jdbc:lightdb://localhost:5432/postgres";
    Properties props = new Properties();
    PGProperty.USER.set(props, "postgres");
    PGProperty.PASSWORD.set(props, "postgres");
    PGProperty.ASSUME_MIN_SERVER_VERSION.set(props, "9.4");
    PGProperty.REPLICATION.set(props, "database");
    PGProperty.PREFER_QUERY_MODE.set(props, "simple");

    Connection con = DriverManager.getConnection(url, props);
    PGConnection replConnection = con.unwrap(PGConnection.class);

The entire replication API is grouped in org.postgresql.replication.PGReplicationConnection and is available via org.postgresql.PGConnection#getReplicationAPI.

Before you can start replication protocol, you need to have replication slot, which can be also created via ltjdbc API.

    replConnection.getReplicationAPI()
        .createReplicationSlot()
        .logical()
        .withSlotName("demo_logical_slot")
        .withOutputPlugin("test_decoding")
        .make();

    PGReplicationStream stream =
        replConnection.getReplicationAPI()
            .replicationStream()
            .logical()
            .withSlotName("demo_logical_slot")
            .withSlotOption("include-xids", false)
            .withSlotOption("skip-empty-xacts", true)
            .start();

The replication stream will send all changes since the creation of the replication slot or from replication slot restart LSN if the slot was already used for replication. You can also start streaming changes from a particular LSN position,in that case LSN position should be specified when you create the replication stream.

    LogSequenceNumber waitLSN = LogSequenceNumber.valueOf("6F/E3C53568");

    PGReplicationStream stream =
        replConnection.getReplicationAPI()
            .replicationStream()
            .logical()
            .withSlotName("demo_logical_slot")
            .withSlotOption("include-xids", false)
            .withSlotOption("skip-empty-xacts", true)
            .withStartPosition(waitLSN)
            .start();

Via withSlotOption we also can specify options that will be sent to our output plugin, this allows customize decoding. For example I have my own output plugin that has a property sensitive=true which will include changes by sensitive columns to change event.

During replication the database and consumer periodically exchange ping messages. When the database or client do not receive ping message within the configured timeout, replication has been deemed to have stopped and an exception will be thrown and the database will free resources. In LightDB the ping timeout is configured by the property wal_sender_timeout (default = 60 seconds). Replication stream in pgjdc can be configured to send feedback(ping) when required or by time interval. It is recommended to send feedback(ping) to the database more often than configured wal_sender_timeout. In production I use value equal to wal_sender_timeout / 3. It's avoids a potential problems with networks and changes to be streamed without disconnects by timeout. To specify the feedback interval use withStatusInterval method.

Example 9.10. Replication stream with configured feedback interval equal to 20 sec

    PGReplicationStream stream =
        replConnection.getReplicationAPI()
            .replicationStream()
            .logical()
            .withSlotName("demo_logical_slot")
            .withSlotOption("include-xids", false)
            .withSlotOption("skip-empty-xacts", true)
            .withStatusInterval(20, TimeUnit.SECONDS)
            .start();

After create PGReplicationStream, it's time to start receive changes in real-time. Changes can be received from stream as blocking(org.postgresql.replication.PGReplicationStream#read) or as non-blocking(org.postgresql.replication.PGReplicationStream#readPending). Both methods receive changes as a java.nio.ByteBuffer with the payload from the send output plugin. We can't receive part of message, only the full message that was sent by the output plugin. ByteBuffer contains message in format that is defined by the decoding output plugin, it can be simple String, json, or whatever the plugin determines. That why ltjdbc returns the raw ByteBuffer instead of making assumptions.

    while (true) {
      //non blocking receive message
      ByteBuffer msg = stream.readPending();

      if (msg == null) {
        TimeUnit.MILLISECONDS.sleep(10L);
        continue;
      }

      int offset = msg.arrayOffset();
      byte[] source = msg.array();
      int length = source.length - offset;
      System.out.println(new String(source, offset, length));
    }

As mentioned previously, replication stream should periodically send feedback to the database to prevent disconnect via timeout. Feedback is automatically sent when read or readPending are called if it's time to send feedback. Feedback can also be sent via org.postgresql.replication.PGReplicationStream#forceUpdateStatus() regardless of the timeout. Another important duty of feedback is to provide the server with the Logial Sequence Number (LSN) that has been successfully received and applied to consumer, it is necessary for monitoring and to truncate/archive WAL's that that are no longer needed. In the event that replication has been restarted, it's will start from last successfully processed LSN that was sent via feedback to database.

The API provides the following feedback mechanism to indicate the successfully applied LSN by the current consumer. LSN's before this can be truncated or archived. org.postgresql.replication.PGReplicationStream#setFlushedLSN and org.postgresql.replication.PGReplicationStream#setAppliedLSN. You always can get last receive LSN via org.postgresql.replication.PGReplicationStream#getLastReceiveLSN.

    while (true) {
      //Receive last successfully send to queue message. LSN ordered.
      LogSequenceNumber successfullySendToQueue = getQueueFeedback();
      if (successfullySendToQueue != null) {
        stream.setAppliedLSN(successfullySendToQueue);
        stream.setFlushedLSN(successfullySendToQueue);
      }

      //non blocking receive message
      ByteBuffer msg = stream.readPending();

      if (msg == null) {
        TimeUnit.MILLISECONDS.sleep(10L);
        continue;
      }

      asyncSendToQueue(msg, stream.getLastReceiveLSN());
    }

    String url = "jdbc:lightdb://localhost:5432/test";
    Properties props = new Properties();
    PGProperty.USER.set(props, "postgres");
    PGProperty.PASSWORD.set(props, "postgres");
    PGProperty.ASSUME_MIN_SERVER_VERSION.set(props, "9.4");
    PGProperty.REPLICATION.set(props, "database");
    PGProperty.PREFER_QUERY_MODE.set(props, "simple");

    Connection con = DriverManager.getConnection(url, props);
    PGConnection replConnection = con.unwrap(PGConnection.class);

    replConnection.getReplicationAPI()
        .createReplicationSlot()
        .logical()
        .withSlotName("demo_logical_slot")
        .withOutputPlugin("test_decoding")
        .make();

    //some changes after create replication slot to demonstrate receive it
    sqlConnection.setAutoCommit(true);
    Statement st = sqlConnection.createStatement();
    st.execute("insert into test_logic_table(name) values('first tx changes')");
    st.close();

    st = sqlConnection.createStatement();
    st.execute("update test_logic_table set name = 'second tx change' where pk = 1");
    st.close();

    st = sqlConnection.createStatement();
    st.execute("delete from test_logic_table where pk = 1");
    st.close();

    PGReplicationStream stream =
        replConnection.getReplicationAPI()
            .replicationStream()
            .logical()
            .withSlotName("demo_logical_slot")
            .withSlotOption("include-xids", false)
            .withSlotOption("skip-empty-xacts", true)
            .withStatusInterval(20, TimeUnit.SECONDS)
            .start();

    while (true) {
      //non blocking receive message
      ByteBuffer msg = stream.readPending();

      if (msg == null) {
        TimeUnit.MILLISECONDS.sleep(10L);
        continue;
      }

      int offset = msg.arrayOffset();
      byte[] source = msg.array();
      int length = source.length - offset;
      System.out.println(new String(source, offset, length));

      //feedback
      stream.setAppliedLSN(stream.getLastReceiveLSN());
      stream.setFlushedLSN(stream.getLastReceiveLSN());
    }

API for physical replication looks like the API for logical replication. Physical replication does not require a replication slot. And ByteBuffer will contain the binary form of WAL logs. The binary WAL format is a very low level API, and can change from version to version. That is why replication between different major LightDB versions is not possible. But physical replication can contain many important data, that is not available via logical replication. That is why pgjdc contains an implementation for both.

    LogSequenceNumber lsn = getCurrentLSN();

    Statement st = sqlConnection.createStatement();
    st.execute("insert into test_physic_table(name) values('previous value')");
    st.close();

    PGReplicationStream stream =
        pgConnection
            .getReplicationAPI()
            .replicationStream()
            .physical()
            .withStartPosition(lsn)
            .start();

    ByteBuffer read = stream.read();