When my GenStage consumer(s) are too slow to handle the amount of incoming data (Tweets) the producer will stop ack the HTTP/TCP connection packages. The Twitter streaming endpoint will stop sending more data through the connection.

This is what whe want and follows the same pattern GenStage and Flow follow to consume data.

Without back-pressure, my producer would have to buffer the incoming Tweets or it’s internal message box would grow indefinitly. In both cases this would result in taking more and more memory, crashing the VM in the long term.

I wrote a Twitter library from scratch (at that time, it was my first project to get startet with Elixir). It uses HTTPoison (wrapper for :hackney).

A few weeks ago, I’ve updated the streaming part of the lib to use GenStage.

You can have a look at the implementation here: https://github.com/almightycouch/twittex/blob/master/lib/twittex/client/stream.ex

Basically, the Twittex.Client.Stream receives incoming HTTP chunks from :hackney applying back-pressure at the TCP level with :hackney.stream_next/1.

The next thing I want to try out is using Flow and Flow.Window to implement features such as peak detection.

It is a bug in ExTwitter. The stream is discarding messages from the process inbox, which is a big no-no. This means that the subscriptions for the second and third stages are never consumed:

ExTwitter should use a reference and make sure it only consumes messages it knows about. This must be reported as a bug.

Ideally ExTwitter would provide a GenStage producer, to avoid hijacking the process inbox, which won’t work with a GenStage since both GenStage and ExTwitter are trying to use the same process inbox. You can check the issues tracker so see if someone already requested this feature.

Hi @maz, no problem at all, you helped me found an annoying bug here :

@api_key Application.get_env(:twittex, :consumer_key)
@api_secret Application.get_env(:twittex, :consumer_secret)

Basically, @consumer_key and @consumer_secret where defined at compile time in the Twittex.Client.Base module. Updating those configuration keys did not force the module to recompile which I think is why you encountered you authentication problems (the first time you compiled :twittex, those keys where compiled into the application even when you updated your config).

You could try to run mix deps.clean twittex and recompile your application (or run iex -S mix directly), this should fix the problem. Otherwise, I just published version 0.3.5 which fixes the issue and loads :consumer_key and :consumer_secret at runtime.

You haven’t overridden the default demand quantities. GenStage operates in batches, and the min / max batch sizes are 500 / 1000 by default, so unless you have more than that it isn’t gonna use more than one worker.

This is a good example of https://github.com/elixir-lang/gen_stage/issues/72