Falcon no longer respects count, new likely increases memory usage on shared hosts

[trevorturk]

I think there's an issue with Falcon 0.45.0 where it uses significantly more memory as opposed to 0.43.0.

I've isolated a diff of the minimal Gemfile(.lock) changes for my app here:

diff --git a/Gemfile b/Gemfile
index a72feb297..126e77f35 100644
--- a/Gemfile
+++ b/Gemfile
@@ -12,7 +12,7 @@ gem "chartkick"
 gem "coffee-rails"
 gem "connection_pool"
 gem "dogapi"
-gem "falcon", "0.43.0"
+gem "falcon"
 gem "geokit"
 gem "graphql"
 gem "hashie"
diff --git a/Gemfile.lock b/Gemfile.lock
index 7be649138..8ccc9c4b5 100644
--- a/Gemfile.lock
+++ b/Gemfile.lock
@@ -85,7 +85,7 @@ GEM
       fiber-annotation
       io-event (~> 1.5, >= 1.5.1)
       timers (~> 4.1)
-    async-container (0.16.13)
+    async-container (0.18.0)
       async
       async-io
     async-http (0.64.0)
@@ -102,6 +102,9 @@ GEM
       async
     async-pool (0.4.0)
       async (>= 1.25)
+    async-service (0.10.0)
+      async
+      async-container (~> 0.16)
     barnes (0.0.9)
       multi_json (~> 1)
       statsd-ruby (~> 1.1)
@@ -109,7 +112,6 @@ GEM
     bcrypt (3.1.20)
     bigdecimal (3.1.7)
     bindex (0.8.1)
-    build-environment (1.13.0)
     builder (3.2.4)
     bundler-audit (0.9.1)
       bundler (>= 1.2.0, < 3)
@@ -148,19 +150,19 @@ GEM
       railties (>= 5.0.0)
     faker (3.3.1)
       i18n (>= 1.8.11, < 2)
-    falcon (0.43.0)
+    falcon (0.45.0)
       async
-      async-container (~> 0.16.0)
+      async-container (~> 0.18)
       async-http (~> 0.57)
       async-http-cache (~> 0.4.0)
       async-io (~> 1.22)
-      build-environment (~> 1.13)
+      async-service (~> 0.10.0)
       bundler
       localhost (~> 1.1)
       openssl (~> 3.0)
       process-metrics (~> 0.2.0)
-      protocol-rack (~> 0.1)
-      samovar (~> 2.1)
+      protocol-rack (~> 0.5)
+      samovar (~> 2.3)
     ffi (1.16.3)
     fiber-annotation (0.2.0)
     fiber-local (1.0.0)
@@ -444,7 +446,7 @@ DEPENDENCIES
   dogapi
   factory_bot_rails
   faker
-  falcon (= 0.43.0)
+  falcon
   gem_updater
   geokit
   graphql

Screenshot of my memory usage metrics on Heroku:

Please let me know if there's anything I can do to help debug! Thank you!

[evkorotkov]

I've noticed the same, but looks like the problem is not in the memory consumption but in wrong instances count. Falcon 0.44.0+ started ignoring count option in configuration and runs forks count = processors count (as if option is empty). As a workaround you can explicitly override instances count using ASYNC_CONTAINER_PROCESSOR_COUNT environment variable.

cc: @ioquatix

[trevorturk]

Ah, perhaps in the move to async-service in https://github.com/socketry/falcon/pull/226 that was (inadvertently?) dropped, but I'm not sure if the plan is to migrate into a new config style? I think we should probably still support count, or raise an error if it's being used and shouldn't be?

I dug a bit and see Etc.nprocessors is used, but this reminded me of the discussion around Puma in Rails here: https://github.com/rails/rails/issues/50450#issuecomment-1881585649 so I checked and imo we should consider not using that system and instead defaulting to 1 or something. See for example on my Heroku Standard-1x Dyno which I believe should only run 1 Falcon process where I'm likely being assigned 1 virtual CPU on a machine that has 8?

$ heroku run rails console
Running rails console (Standard-1X)
irb(main):001> Etc.nprocessors
=> 8

Somewhat of a separate issue, but if we need to move over to the new configuration style, we'll need to figure out what changes are needed from the old style. For example my config for Heroku is currently:

#!/usr/bin/env -S falcon host

load :rack

hostname = File.basename(__dir__)
port = ENV["PORT"] || 3000

rack hostname do
  append preload "preload.rb"
  cache false
  count ENV.fetch("FALCON_COUNT", 1).to_i
  endpoint Async::HTTP::Endpoint.parse("http://0.0.0.0:#{port}").with(protocol: Async::HTTP::Protocol::HTTP11)
end

...and I'd want to make sure to disable caching etc.

[ioquatix]

This looks like a bug I will investigate it today.

[ioquatix]

The count was not used when creating the container:

https://github.com/socketry/falcon/blob/e17d9190fffc872bb414c3f8b880b68d091803fd/lib/falcon/service/server.rb#L53-L56

It's using container_options[:count]. We can add count to the default container options which should fix this.

[ioquatix]

This should be fixed in v0.45.1 - please let me know if you have any issues.

[trevorturk]

Thank you! ☺️

[trevorturk]

Confirmed fixed for me. Thank you for the quick fix!

[evkorotkov]

Confirmed fixed for me too. Many thanks!