Hi, i'm gonna hijack this thread a bit. My community uses ClearML and is looking at various model deployment strategies. We are looking at a seamless integration with Triton but noted they Triton does not support deployment strategies. ClearML-Serving seems to but the strategies are rather limited. Is there a roadmap to expand Clearml-serving?

Hi AgitatedDove14 , thanks.
In this case i am running k8s glue (machine glue), which will then spawn off pods in kubernetes worker (machine worker). So when you say direct access, are you refering to the Glue machine or K8S Worker machine?

Hi CostlyOstrich36 , thanks. I will check with the Enterprise team then.

This one can be solved with shared cache + pipeline step, refreshing the cache in the shared cache machine.

Would you have an example of this in your code blogs to demonstrate this utilisation?

Hi, so you meant i need to installl virtualenv in my base image?

Hi SuccessfulKoala55 , can i confirm the following comments in the docker-compose.yml ?
And after that to run docker-compose commands without loss of data?

docker-compose down docker-compose up
docker-compose.yml
`
version: "3.6"
services:

apiserver:
command:
- apiserver
container_name: clearml-apiserver
image: allegroai/clearml:latest
restart: unless-stopped
volumes:
- /opt/clearml/logs:/var/log/clearml
- /opt/clearml/config:/opt/clearml/config
#...

Alright will give it a try.

Thanks CostlyOstrich36 , how do i know how is the parts indexed in the first place? Or rather, how is chunk and parts defined? Say in the context of images, videos, text documents...etc.

The server is running only the ClearML components. Could you advise on the ELB part, how should we optimise it?

Ok thks.

I did another test by running
kubectl exec pod-name -- echo $PIP_INDEX_URL and it returned nothing. So the env are not passed to the container at all.

Hi AgitatedDove14 , i changed everything to cuda 10.1 and tried again with the same rrror. the section as follows. I made sure torch==1.6.0+cu101 and torchvision==0.8.2+cu101 are in the pypi repo. But the same error still came up.
` # Python 3.6.9 (default, Oct 8 2020, 12:12:24) [GCC 8.4.0]
boto3 == 1.14.56
clearml == 0.17.4
numpy == 1.19.1
torch == 1.6.0
torchvision == 0.7.0

Detailed import analysis

**************************

IMPORT PACKAGE boto3

clearml.storage: 0

IMPORT PACKAG...

Yeah.. issue is ClearML unable to talk to the nodes cos pytorch distributed needs to know their IP. There is some sort of integration missing that would enable this.

Hi thanks. How about Agent, does its docker mode or k8s mode require docker.sock to be exposed?

Hi CostlyOstrich36 , That's correct.

Share data across R&D teams with searchable data catalogs available on any environment.

Ok thanks, that worked.

python k8s_glue_example.py --queue gpu --namespace default 
Traceback (most recent call last):
 File "k8s_glue_example.py", line 86, in <module>
  main()
 File "k8s_glue_example.py", line 80, in main
  namespace=args.namespace,
 File "/home/administrator/clearml-agent-k8s/venv/lib/python3.6/site-packages/clearml_agent/helper/base.py", line 239, in _ call _
  cls. instances[cls] = super(Singleton, cls). call_(*args, **kwargs)
TypeError: _ init _() got an unexpected keyword argument 'base_pod...

the hackathon is 3 days.

I'm also noticing a lot of this while the k8s glue is running.
Ex: Expecting value: line 1 column 1 (char 0) K8S Glue pods monitor: Failed parsing kubectl output:

So these (PIP_INDEX_URL) weren't used when clearml starts running pip.

Thanks SuccessfulKoala55 . I can try my hand on a patch. But the pod spinning is handled by the k8s glue, which has no link to the client side. How should the client pass the key over to k8s glue during runtime via clearml server?

Hi AgitatedDove14 . I'm trying out passing env via the code instead.
task.set_base_docker("nvcr.io/nvidia/tensorflow:19.11-tf2-py3 --env TRAINS_AGENT_GIT_USER=git_username_here --env TRAINS_AGENT_GIT_PASS=git_password_here")So the strange thing is when my k8sglue pulls a task, this happens.
` Pulling task xxxxxxxxxx launching on kubernetes cluster
Pushing task xxxxxxxxxx into temporary pending queue
Kubernetes scheduling task id=xxxxxxxxxxxx
skipping docker argument TRAINS_AGENT_GIT_USE...

Sorry, in case i misunderstood you. Are you refering to the extra_docker_shell_script .

I want to rule out the glue being the problem. Is the Glue significant in initialising clearml-agent after the pod is spawned?

Hi, scenario as follows.

client.py runs task.execute_remotely(queue='myqueue', exit_process=True) The API section of clearml.conf at client side is read in. client side calls clearml server and insert task into queue. K8S glue retrieves task from queue. Spawn a K8S pod. K8S pod performs git clone Error. ssh keys not found.
Each individual has their own key in the gitlab profile and gitlab is configured to only work via ssh.
We can't place the key in the image as this is as good as ...

Ok. That brings me back to the spawned pod. At this point, clearml-agent and its config would be a controbuting factor. Is the absence of /tmp/.clearml_agent.xxxxxx.cfg an issue?

We are using k8s glue to spawn the job. Would you be able to advise in detail of steps on what goes on when the above code executes?

This is a env var?
CLEARML_CONFIG_FILE

Ok thanks, we'll try it out on next availability.