supervisor process memory consumption

[gsar]

It looks like the supervisor process consumes substantial amount of memory, mainly due to it being forked after the intial falcon host loads preload.rb. I'm running multiple falcon instances in 1GB dynos on heroku, and this leaves only sufficient room for a single worker.

Is is possible to reduce the memory consumed by the supervisor, or perhaps have an option to run the supervisor fibers within the master process itself?

Thanks for falcon!

[ioquatix]

I'm assuming that you are loading a lot of things in preload.rb which is totally fine.

In that case, the supervisor will share those pages with all the other instances, and it should not really count towards the usage. However, IIRC, Heroku may count this towards your usage, even if it's shared.

One option is to fork off the child workers using a separate process, e.g. fork a process that does preload independently of the other services. I think this could be useful in general, so it given your scenario and others, it would be an interesting idea to explore (this feature does not exist yet).

[gsar]

Since the supervisors don't do much I was thinking they could be a small executable that gets exec-ed.

The only thing in preload.rb is require_relative 'config/environment' and the rails app itself is not that large.

Just to add some extra detail. With 3 workers in a 2.5GB dyno, soon after a restart, heroku shows RSS of 1644MB and the ProcessMonitor stats look like the screenshot below. Of course, there is some memory growth over time as well (around 1.5x at steady state).