Back pressure in Process API

[bergie]

There are many situations where you want to throttle component's activity level. Examples could be rate limited API calls, or amount of files you can open simultaneously.

In the legacy AsyncComponent implementation we had support for a load outport, where the component would send its current load. Then you could use a component like flow/Throttle in front of the component to prevent new jobs being triggered before the component has finished previous ones, like:

# Only allow max 100 directory reads to happen at same time
'100' -> MAX ThrottleDir(flow/Throttle)
Directory OUT -> IN ThrottleDir OUT -> SOURCE Dir(filesystem/ReadDir)
Dir LOAD -> LOAD ThrottleDir

Another concept that has been used in various FBP systems is "back pressure" where an downstream component can tell its upstream "don't send me more data".

• http://www.elixirfbp.org/2015/07/reactive-streams.html
• Back pressure jpaulm/jsfbp#25
• http://www.jpaulmorrison.com/fbp/concat.html

With backpressure, one question is how the max numbers of jobs for a component would be configured on graph level. We have several mechanisms that could be used, like control inports or node/edge metadata.

[bergie]

Based on a chat with @trustmaster today, the way forward would be to allow graph designers to add a load port to component instances. If such port is present, component activation/deactivation would send current load out to that port. Throttling logic can then be implemented as another component on the graph level.

Depends on #594

[bergie]

Actually, much nicer to properly implement back pressure in NoFlo than doing “load” ports. Buffer size can be handled via edge metadata (see “connection capabilities in noflo/noflo-ui#370).

Quick ideas:

• we can make a Promisified sending method that checks if connection is ready to receive and waits otherwise
• for components that don’t use this, we can probably control their activation when their downstream is blocked, and only activate when downstream is ready to receive again

[linonetwo]

Use backpressure to prevent new input from being processed until the current conversation is done. #1042

Please review this guide, I'm not sure if ChatGPT knows your implementation well.