I don't know about this, but could you turn your whole project into a pip-installable package using a setup.py and/or pyproject.toml ?

I've never tried this, but maybe then you could do pip install -e . locally before executing the task. Then execute. And then maybe the pip freeze that ClearML does would contain the symlink to your directory.

(so that from my_package import ... statements would work)

I could imagine other useful automations for reacting to failed tasks that have certain tags, including alerting.

I realize we could move a lot of this logic into ClearML itself: make handler functions that run within the services queue. That would work for logic that is implemented in Python. But I believe it would be harder for our team to detect and respond to failures in the event handler functions if they were placed there because it seems unclear how we could use our existing systems a...

Thanks for the response @<1523701205467926528:profile|AgitatedDove14> !

What would you consider an event?

I was thinking of the TriggerScheduler 's definition of an event. Pretty much, any thing the TriggerSchedule allows you to react to, it'd be great to be able to publish those events to a queue external to ClearML, e.g. a tag added to a model (or removed), a state in a task changing, etc. We'd want as much metadata about that event as possible. So if the event is due to a task...

Exactly

For now, I've written a headless selenium script to generate credentials for the fresh ClearML instance in CI.

But the extension will need credentials to connect to it.

I could potentially write a selenium script to make a set of keys, but I'd prefer to avoid that 😅

I ultimately resorted to creating a selenium script combined with docker-compose. Not a beautiful solution but I can confirm that it works 😕 None

If this works, we might be able to fully replace Metaflow with ClearML!

(Refering to the feature where Metaflow creates Step Functions state machines for you, and then you can use those to trigger event-driven batch jobs in the same way described here)

On it

Man, I owe you lunch sometime @<1523701205467926528:profile|AgitatedDove14> . Thanks for being so detailed in your answers.

Okay! So the pipeline ID is really just a task ID. So cool!

Not sure I fully understand what you mean here...

Sorry, I'll try again. Here's an illustrated example with AWS Step Functions (pretend this is a ClearML pipeline). If the pipeline fails, I'd want to have a chance to do some logic to react to that. Maybe in a step called "on_pipeline_failed" or someth...

The agent commands are nothing special.

clearml-agent daemon --queue sessions --cpu-only --create-queue true --docker

Haha, that was a total gotcha for me. Yeah, a lot just wasn't even getting run due to the #!/bin/bash part.

Anyway, wow! I finally got the precious console logs you thought to find, here they are:

2023-05-06 00:19:21
User aborted: stopping task (3)
2023-05-06 00:19:21
Successfully installed PyYAML-6.0 attrs-22.2.0 certifi-2022.12.7 charset-normalizer-3.1.0 clearml-agent-1.5.2 distlib-0.3.6 filelock-3.12.0 furl-2.1.3 idna-3.4 jsonschema-4.17.3 orderedmultidict-1.0.1 pathlib2-2.3.7....

One idea: is it possible to store usable credentials in advance and place them in a volume that the ClearML containers can access and then use?

When you run the docker-compose.yml on an EC2 instance, you can configure user login for the ClearML webserver. But the files API is still open to the world, right? (and same with the backend?)

We could solve this by placing the EC2 instance into a VPN.

One disadvantage to that approach is it becomes annoying to reach the model registry from outside the VPN, like if you have a deployment pipeline based in GitHub Actions. Or if you wanted to trigger a ClearML pipeline from a VPC that isn...

Yes, it's pretty lame that a clearml-agent can only process one task at a time if it's not listening to a services queue 🤔

@<1523701070390366208:profile|CostlyOstrich36> Oh that’s smart. Is that to make sure no transactions happen during the backup? Would there be a risk of ongoing or pending tasks somehow getting corrupted if you shut the server down?

Oh, that is cool. I captured all this. Maybe I'll make a user-data.sh script and docker-compose.yml file that brings all these things together. Probably won't have time for a few weeks.

To do this, I think I need to know:

• Can you trigger a pre-existing Pipeline via the ClearML REST API? I'd want to have a Lambda function trigger the Pipeline for a batch without needing to have all the Pipeline code in the lambda function. Something like curl -u '<clearml credetials>' None ,...
• [probably a big ask] If the pipeline succeeds/fails, can ClearML emit an event that I can react to? Like mayb...

Earlier in the thread they mentioned that the agents are all resilient. So no ongoing tasks should be lost. I imagine even in a large organization, you could afford 5-10 minutes of downtime at 2AM or something.

That said, you'd only have 1 backup per day which could be a big deal depending on the experiments your running. You might want more than that.

Thanks for this!! I may try it and if I do and it works I’ll look into writing a plugin for ZenML and Metaflow that auto initializes the parent task and registers the steps as child tasks. Super helpful thank you!

You know, you could probably add some immortal containers to the docker-compose.yml that use images with mongodump and the ES equivalent installed.

The container(s) could have a bash script with a while loop in it that sleeps for 30 minutes and then does a backup. If you installed the AWS CLI inside, it could even take care of uploading to S3.

I like this idea, because docker-compose.yml could make sure that if the backup container ever dies, it would be restarted.

So I get output with this one, but the console only shows me the output from my machine. For example, the SSH key is present, and whoami results in ericriddoch

I’d really prefer it was modular enough to use serving with any model registry

Oh that's interesting. To serve a model from MLflow, would you have to copy it over to ClearML first?

As an infrastructure engineer, I feel that this is a fairly significant shortcoming of ClearML.

Having the ability to pack jobs/tasks onto the same "resource" (underlying server/EC2 instance) would

• simplify the experience for data scientists
• open up a streaming use case, wherein batch (offline) inference could be done directly inside of a ClearML pipeline in reaction to an event/trigger (like new data landing in your data lake). As it is, you can make this work, but if you start to get ...

But from your other answer, I think I'm understanding that you can have multiple agents on a single instance listening to the same queue.

So we could maybe initialize 4 instances of the agent on a single EC2 instance which would allow us to handle a higher volume of small batches concurrently without tying up the entire instance.

That's fabulous. This is definitely how my team prefers to structure projects. I hadn't gotten around to trying that out in our POC of ClearML yet, but I'm certain this is how our group will solve this problem

Oh hooray! So docker-compose manages the restarting of crashed containers? I didn't know that, and that is great 😄

possibly cheaper on the cloud (Lambda vs EC2 instance)

Whoa, are you saying there's an autoscaler that doesn't use EC2 instances? I may be misunderstanding, but that would be very cool.

Maybe I should have said: my plan is to use AWS StepFunctions where a single task in the DAG is an entire ClearML pipeline . The non-ClearML steps would orchestrate putting messages into a queue, doing retry logic, and triggering said pipeline.

I think at some point, there has to be some amount of...

I'm imagining:

• The EC2 instance would be in a private subnet, accessible only on the VPN (read: VPC)
• The API Gateway and Load Balancer would also be on the VPC and therefore have access to the private subnet BUT the API Gateway or Load Balancer themselves would be exposed to the public internet.
  That way, to do the JWT authentication, the load balancer or API Gateway could reach out to the EC2 instance on the private network to authenticate any incoming ClearML SDK requests.