[ZephyrP]

BEAM doesn't do everything containers or a container orchestration system can, but in view of my relatively short time as a regular Erlang programmer (definitely less than 10 years), the spirit of tidepod's point accords with my experience (with a weaker interpretation of his term "node clustering"). BEAM and OTP don't do automatic load distribution between nodes, but can be fashioned to do so easily with application support.

We never used pool. The nodes were mapped onto heterogenous machines sharing the host with a 3rd-party daemon. It's configuration changes even took place through a module update hook written in Erlang itself. We both deployed new code and distributed work "manually" across them entirely on OTP.

[NOTE] It it surprising, or was to me, that there are problems with having a fairly small number of nodes fully connected. I'm lucky enough to have avoided learning this the hard way, but imagine this could serve as a painful backbone to an "Erlang deployment war story".

[kungfooguru]

Sure, you can write it yourself, and probably it is even easier to write in Erlang -- up to a limit due to the issues with distributed Erlang discussed above.

I worry about, and have seen this both in the first hype phase of Erlang a number of years ago, the misconception about what Erlang offers and the resulting frustration, blaming it on the tool and quitting.

Hoping in the chapters I'm working on for https://adoptingerlang.org/docs/production/ I can better explain the benefits of running on k8s (or similar), while also making clear it certainly isn't the right choice in all cases.

- Tristan

[elcritch]

Nice synopsis in the docs. I’d suggest that instead of saying k8s and OTP don’t overlap that it’s more that "they don’t completely overlap" Fenn diagram style. That’d be a bit more nuanced answer and help people judge which category they fall into.

There are many cases where having an Erlang cluster and doing ‘naive pooling’ at the application level would get you 80% of what k8s + routing layer would get you. I’m assuming more of an an all Erlang/Elixir environment which many small startups could get by with. Even then Docker + Fargate or whatnot would still simplify deployment. Personally I harbor a secret desire to see if I could replicate part of the k8s interface using Nerves images + some otp tooling. Probably better things to do with my life though. ;)

Also by ‘heart’ are you referring to how ‘epmd’ works or Erlang distribution?

[kungfooguru]

For 'heart' here I'm referring to the heart program that'll start if you run erl with `-heart` http://erlang.org/doc/man/heart.html -- it will restart a down node.

I think part of the confusion is you in general don't need k8s, but when your size and requirements get to a point that k8s makes sense (which arguably lowers as hosted k8s becomes better) it is not in conflict with your also use of Erlang/Elixir.

I find distributed Erlang much nicer in k8s env (I don't have to maintain it obviously) where I get an IP per node, can add a k8s service making it possible to use DNS SRV queries to find all other nodes and letting k8s worry about where pods run and keeping them up.

Plus there is configuration management and consistent storage (resources in etcd).