My little Backpressure: Flow Control is magic

By: on September 30, 2013

When we’re designing systems that are designed to be robust against failure, it’s important to know how behaviour at your Integration points (a term borrowed from Michael Nygard’s book Relase It!) impacts the rest of the system. For example, if your database or a remote API is running slowly, then in Synchronous systems, because you (usually) have a limited number of threads on which to process the requests, any slow down will naturally be pushed back onto the clients of that service.

Assuming your application server behaves sensibly, then in the worst case, then the listen queue on the server’s listening socket will fill up, and any further connections will be refused. Any load balancer that you have in front of it should then know to stop sending requests to that instance.

However, in pervasively Asynchronous systems, it can be more difficult to manage control flow. For example, in its default configuration, the Node.js web server module will allow you to have as many outstanding requests as it can hold in memory. Naturally, for some applications that can be helpful. However, if as above, your authentication service is running slowly, then all of those requests can potentially pile up in your Node.js process until you run out of memory.

Again, as with the synchronous case it is possible to put a cap on the number of outstanding requests, but managing the flow control can get quite complicated quite quickly, unless you are really careful.

This is where core.async comes in. By allowing you to write synchronous-looking code that runs asynchronously, I find that it’s a lot easier to build reliability features, and in our example, back-pressure.

Let’s take a trivial example of doing this on a Node.JS with the Streams2 API. The new api makes backpressure far easier to manage when reading (as the consuming code is notified when data is ready, and can then pull data when it is ready), there are still difficulties around writing. For example, the state of node’s internal buffers are reflected via two mechanisms:

  • The write method on a stream will return true until a high watermark is hit, after which it returns false until these are flushed
  • The write method takes an optional callback, which gets invoked when requested write has been flushed to the operating system (or has failed).

Now, it is entirely possible to use these to build a state machine which allow you to propagate back-pressure, but such things can often be quite ad-hoc, and rather more complex than the synchronous version. Also, both need to implemented when implementing your own Stream. So unfortunately, this complexity means that back-pressure often seems to get ignored when writing data.

So, I’ve put together a full example of how one could implement a buffering version of cat using core.async.

The most complex part of the code is in the function writeable-from-chan, as below:

(defn writeable-from-chan [ch strm]
  (let [drains (chan)]
      (loop [] ;; #1
        (let [buf (<! ch) ;; #2
          drain-cb (fn drain-cb [] (put! drains :token))] ;; #4
          (if buf
            (.write strm buf drain-cb) ;; #3
            (<! drains) ;; #5
        (.end strm))))))) ;; #6

Essentially, this starts a loop (#1) which takes input from the input channel (#2) and writes (#3) it to the output stream. At this point, we pass in a callback (#4) which signals when the write is complete, and we can then effectively block on the completion of the write (#5). When the stream is closed (and the take at #2 returns nil) we close the output stream. (#6).

Now, the nice thing here, is that back-pressure is managed for you; a default channel only offers one cell of space, and so any subsequent puts will block until another process has taken that value from the channel. So, data can be pushed through a channel with no need to invoke #pause() or resume() methods on the input, nor do you have to handle back-pressure specially, or do any extra buffering.

In summary, the control flow can become clear in the structure of your code, rather having to construct an explicit state machine mechanism, which greatly enhance clarity, whilst preserving a degree of safety.



  1. Michael says:

    Also, both need to implemented when implementing your own Stream.

    Only #_write must be implemented; given that, buffering etc., is done by the Writable internals.

    So unfortunately, this complexity means that back-pressure often
    seems to get ignored when writing data.

    It’s not too bad: write until you get false, then wait for ‘drain’.

    That said I don’t think the streams(2) API is all that useful, excepting the built-in streams; and putting a coroutine facade on it seems pretty sensible.

  2. Ceri Storey says:

    Thanks; I’d only really skimmed the documentation for the new streams API, so I’m happy for the correction.

    Also, your point about using the return value from #write would help to avoid the large number of needless context switches (between the Javascript and I/O worker thread) that my code would likely cause.


Leave a Reply

Your email address will not be published.

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>