Oh that's awesome. So medium memory is basically shm?

Yea it’s stored in /dev/shm on the host

Excellent

@devth I found this code useful to manage development jobs https://gist.github.com/jeroenvandijk/95a7d09b3c8cb7904547c95c3b4a3360#file-user-clj-L32-L47. Maybe a (simple) variation would be useful for managing remote jobs

Would it make sense to use Onyx as a microservices architecture?

and hi everyone 🙂

jeroenvandijk: thanks! that helps.

@emil0r: possibly, depends on exactly what your trying to do

So https://github.com/onyx-platform/lib-onyx/blob/master/doc/server-api.md doesn't include an endpoint to kill or restart a job?

my company just asked them to figure out how to chain Map reduce jobs together. 😢

@drewverlee: Time to go job hunting. 🙂

@devth: Not yet it doesn't! 🙂

Ok. Eagerly awaiting endpoints 🙂

Or are you hoping for PRs?

PR's always welcome, this is definitely on our radar. My approach was going to be allowing lookups by job metadata

Btw, Onyx dashboard is so cool! Watching it react in realtime to submitting and killing jobs 💯

Yea it's pretty slick

However, I still don't understand why a newly submitted job isn't getting any host allocation

I set NPEERS to 6 and created 3 replicas (Onyx peers)

Hm, it looks like 1 job is consuming all my peers

Can I tell it to share? These are streaming / indefinite jobs that never complete.

Nevermind let me just read up on this some more 🙂

Your cluster needs at least 1 peer for every task in each of its jobs

Yeah I think I need to be using the balanced scheduler

So if you have 2 jobs that each have 5 tasks, you will need your host machines to run JVM's that spawn a total of 10 peers

More than a 1 for 1 peer to task ratio means more concurrency

Got it. Thanks

And NPEERS should be roughly set to number of CPUs?

From a performance standpoint it does not make sense to be running more peers than cores on a box. However it's fine to have many more peers than cores.

I recommend getting your jobs setup and tuning for that stuff last, while developing give yourself enough peers to run your jobs

Also if your jobs might be idle for some time, like when handling a streaming workload, it's reasonable to over-allocate peers on a box.

Ah, cool. That makes sense.

Yea the peers terminology is a bit tricky, most immediately think of physical nodes or hosts. It's more akin to a thread pool

What do you call a machine that runs Onyx? A node?

Are all peers "virtual peers"?

Yup

Yea nodes are the machines/hosts that are running a JVM

Ok. I was confusing nodes with Peers before.

Yea a diagram would be helpful

I also have Datomic peers, so it's extra confusing 🙂

Lol yup

@devth: A PR would be great if you add any new endpoints. I don't actively work on that project.

Ok. I'd like contribute. I'll see if I can find the time

Cool, yeah if anything comes up as needed, just send it in. Thanks!

This stuff generally turns out better if people contribute directly from the project's they work on

Cool

@devth: I was able to get going with the official java alpine image. Thanks for your PR.

gardnervickers: great! sorry I didn't get around to figuring out the issue. glad you found it.

it'd be nice if s6 could surface the actual error

It was in a sense, there was no /bin/bash so file not found was appropriate

But yea I get what you mean

@gardnervickers: i wish to build a modular financial system with a set of core modules and the ability to extend the system by adding more modules. a module would be an area of interest for the system such as accounting, accounts, etc

@emil0r: If your system is flowing information unidirectionally and asynchronously, it would be a good fit.

ok. is it possible to change flow on the fly?

As far I understand model, the whole cluster needs to have the same code available. If that's true, the above idea is feasible if designed around shared primitives (running on Onyx). Right Mike?

So you wouldn’t be able to have a job called foo.bar running in the cluster and then add another called foo.baz which would have a completely different code base but still use Onyx?

You can technically have different code running on each node so long as the interfaces match up, but the only time you'd want to do that is during a rolling deployment.

@emil0r: Think of it like this. You have a jar running on each node in your cluster with some Clojure functions on it. You can submit as many jobs, which are all data, as needed to the cluster for concurrent execution that uses those functions in different configurations.

By interface you mean?

Fn signature

ah right

OK thanks. I’ll play around a bit with some code 🙂

@andrewhr: Yeah, definitely the right idea with shared primitives.

@emil0r: Np.