Flow Control - Streams and Pumps

There are several objects in Vert.x that allow data to be read from and written to in the form of Buffers.

In Vert.x, calls to write data return immediately and writes are internally queued.

It's not hard to see that if you write to an object faster than it can actually write the data to its underlying resource then the write queue could grow without bound - eventually resulting in exhausting available memory.

To solve this problem a simple flow control capability is provided by some objects in the Vert.x API.

Any flow control aware object that can be written-to provides a ReadStream, and any flow control object that can be read-from is said to provides a WriteStream.

Let's take an example where we want to read from a ReadStream and write the data to a WriteStream.

A very simple example would be reading from a NetSocket on a server and writing back to the same NetSocket - since NetSocket provides both ReadStream and WriteStream, but you can do this between any ReadStream and any WriteStream, including HTTP requests and response, async files, WebSockets, etc.

A naive way to do this would be to directly take the data that's been read and immediately write it to the NetSocket, for example:

value server = vertx.createNetServer();

server.connectHandler {
  void onConnect(NetSocket sock) {
    sock.dataHandler {
      void onData(Buffer buffer) {
        // Write the data straight back
        sock.write(buffer);
      }
    }
  }
}.listen(1234, "localhost");

There's a problem with the above example: If data is read from the socket faster than it can be written back to the socket, it will build up in the write queue of the NetSocket, eventually running out of RAM. This might happen, for example if the client at the other end of the socket wasn't reading very fast, effectively putting back-pressure on the connection.

Since NetSocket provides WriteStream, we can check if the WriteStream is full before writing to it:

value server = vertx.createNetServer();

server.connectHandler {
  void onConnect(NetSocket sock) {
    sock.dataHandler {
      void onData(Buffer buffer) {
        if (!sock.writeStream) {
          sock.write(buffer);
        }
      }
    }
  }
}.listen(1234, "localhost");

This example won't run out of RAM but we'll end up losing data if the write queue gets full. What we really want to do is pause the NetSocket when the write queue is full. Let's do that:

value server = vertx.createNetServer();

server.connectHandler {
  void onConnect(NetSocket sock) {
    sock.dataHandler {
      void onData(Buffer buffer) {
        if (!sock.writeStream) {
          sock.write(buffer);
        } else {
          sock.writeStream.pause();
      }
    }
  }
}.listen(1234, "localhost");

We're almost there, but not quite. The NetSocket now gets paused when the file is full, but we also need to unpause it when the write queue has processed its backlog:

value server = vertx.createNetServer();

server.connectHandler {
  void onConnect(NetSocket sock) {
    sock.dataHandler {
      void onData(Buffer buffer) {
        if (!sock.writeStream) {
          sock.write(buffer);
        } else {
          sock.readStream.pause();
          sock.writeStream.drainHandler {
            void onDrain() {
              sock.readStream.resume();
            }
          };
        }
      }
    }
  }
}.listen(1234, "localhost");

And there we have it. The WriteStream.drainHandler() event handler will get called when the write queue is ready to accept more data, this resumes the NetSocket which allows it to read more data.

It's very common to want to do this when writing Vert.x applications, so we provide a helper class called Pump which does all this hard work for you. You just feed it the ReadStream and the WriteStream and it tell it to start:

value server = vertx.createNetServer();

server.connectHandler {
  void onConnect(NetSocket sock) {
    Pump(sock, sock).start();
  }
}.listen(1234, "localhost");

Which does exactly the same thing as the more verbose example.

Let's look at the methods on ReadStream and WriteStream in more detail:

ReadStream

ReadStream is provided by HttpClientResponse, HttpServerRequest, WebSocket, NetSocket, SockJSSocket and AsyncFile.

• ReadStream.dataHandler() set a handler which will receive data from the ReadStream. As data arrives the handler will be passed a Buffer.
• ReadStream.pause() pause the handler. When paused no data will be received in the dataHandler.
• ReadStream.resume() resume the handler. The handler will be called if any data arrives.
• ReadStream.exceptionHandler() Will be called if an exception occurs on the ReadStream.
• ReadStream.endHandler() Will be called when end of stream is reached. This might be when EOF is reached if the ReadStream represents a file, or when end of request is reached if it's an HTTP request, or when the connection is closed if it's a TCP socket.

WriteStream

WriteStream is provided by , HttpClientRequest, HttpServerResponse, WebSocket, NetSocket, SockJSSocket and AsyncFile.

• WriteStream.write() write a Buffer to the WriteStream. This method will never block. Writes are queued internally and asynchronously written to the underlying resource.
• WriteStream.setWriteQueueMaxSize() set the number of bytes at which the write queue is considered full, and the method writeQueueFull() returns true. Note that, even if the write queue is considered full, if `write is called the data will still be accepted and queued.
• WriteStream.writeQueueFull() returns true if the write queue is considered full.
• WriteStream.exceptionHandler() Will be called if an exception occurs on the WriteStream.
• WriteStream.drainHandler() The handler will be called if the WriteStream is considered no longer full.

Pump

Instances of Pump have the following methods:

• Pump.start() Start the pump.
• Pump.stop() Stops the pump. When the pump starts it is in stopped mode.
• Pump.setWriteQueueMaxSize() This has the same meaning as WriteStream.setWriteQueueMaxSize().
• Pump.bytesPumped Returns total number of bytes pumped.

A pump can be started and stopped multiple times.

When a pump is first created it is not started. You need to call the Pump.start() method to start it.