Round robin pool selection #45
Conversation
lib/finch/pool/strategy.ex
Outdated
| @@ -0,0 +1,27 @@ | |||
| defmodule Finch.Pool.Strategy do | |||
There was a problem hiding this comment.
This module should probably be private/@moduledoc false, but i felt the documentation could be useful when implementing additional strategies
lib/finch/pool.ex
Outdated
| {:ok, _} = Registry.register(registry, shp, []) | ||
| {:ok, {shp, opts}} | ||
| def init_pool({registry, state}) do | ||
| {:ok, _} = Registry.register(registry, state.shp, state.registry_value) |
There was a problem hiding this comment.
I realize that we will have this registry value duplicated for every single pool, but I think this is the best way to quickly look it up at the same time as the pool pid.
lib/finch/pool/round_robin.ex
Outdated
|
|
||
| @impl true | ||
| def choose_pool([{_, %{counter: counter, count: count}} | _] = pids) do | ||
| index = :counters.get(counter, 1) |
There was a problem hiding this comment.
Unfortunately I think this does not really guarantee round robin, because you can have handle_enqueue being called multiple times before you get to choose the pool. Or you choose the pool multiple times before handle_enqueue is called and bumped. You think you would need to use ets:update_counter if you want an actual round robin.
There was a problem hiding this comment.
Ahh right of course :( Ok, I'll switch to the ets counter. Thanks for pointing that out!
There was a problem hiding this comment.
You can probably use the atomics module that has a add_get function
There was a problem hiding this comment.
That's right! I try that out
sneako
force-pushed
the
round-robin
branch
2 times, most recently
from
46ccef8 to
e806473
Compare
|
|
||
| @impl true | ||
| def registry_value(%{count: count}) do | ||
| atomics = :atomics.new(1, []) |
There was a problem hiding this comment.
I think you might want this to be unsigned. I think you're relying on overflow behavior and this is going into undefined behavior as it's eventually a signed-int overflow. I could be wrong.
There was a problem hiding this comment.
Great catch! I am relying on the overflow behaviour, but you're right that if the value becomes negative, things will break. Thank you!
There was a problem hiding this comment.
AFAIK atomics use intrinsics that guarantee proper wraparound for signed integers as well, but yeah you need to be prepared for negative numbers. Unsigned is probably simpler 👍
Change `node.pid` into `node.vm_pid` in MetadataFormatters.ELK.
|
I did not see any noticeable improvement from these changes in my benchmarks so I left this PR sitting around for that reason. I am definitely still interested in looking in to alternative pool selection strategies, but I will close this PR because it is so outdated. |
I used the
:atomicsmodule, so that means we would no longer support OTP versions < 21.2. We could fallback to an :ets counter if we feel that supporting older OTP versions is necessary. Or we could possibly just not allow this strategy in older versions