Kubernetes Pod Commandline Code Executor custom image

[mattwg]

Has anyone managed to use a custom image when using the K8s code executor?

I do not want to waste tokens and round trips between agents having them install python packages - and instead want to use an image with the python packages I want the agents to use pre-installed.

I built an image locally and can run it in Docker but am not sure how I reference it in the code via the image parameter. I tried referring to it using the tag name etc. but nothing seems to work - everything just hangs indefinitely. I am not sure where to go next. Do I need to push the image somehow to the K8s cluster first? I am using minikube locally at the moment. For reference here is the documentation I am looking at : https://microsoft.github.io/autogen/0.2/docs/topics/code-execution/kubernetes-pod-commandline-code-executor/

At the moment my dockerfile is super simple:

FROM python:3.11-slim

RUN pip3 install pandas seaborn sqlalchemy

CMD ["tail", "-f", "/dev/null"]

I can build this and it spins up a container just fine with no errors using docker build -t my-code-executor .

In my agent setup I tried to attach the code executor like this:

executor = PodCommandLineCodeExecutor(namespace="default", image="my-code-executor:latest");
agent_x = ConversableAgent(
    "CodeExecutor",
    llm_config=False,  
    code_execution_config={"executor": executor},  
    human_input_mode="NEVER",  
)               

As soon as I add the image parameter the code just hangs when I initiate a conversation. No errors show - it just hangs. If I do not specify an image and it uses the default image all is good - the code executes spinning up K8s pods on the cluster.

What am I missing? Thanks in advance for ideas!

[mattwg]

Quick update - I now get this error after 60seconds:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[1], [line 37](vscode-notebook-cell:?execution_count=1&line=37)
     [26](vscode-notebook-cell:?execution_count=1&line=26) agent_q = ConversableAgent(
     [27](vscode-notebook-cell:?execution_count=1&line=27)     "QueryExpert",
     [28](vscode-notebook-cell:?execution_count=1&line=28)     llm_config={"config_list": [{"model": "gpt-4o", "temperature": 0.7, "api_key": os.environ.get("OPENAI_API_KEY")}]},
   (...)
     [32](vscode-notebook-cell:?execution_count=1&line=32)     human_input_mode="NEVER",  
     [33](vscode-notebook-cell:?execution_count=1&line=33) )
     [35](vscode-notebook-cell:?execution_count=1&line=35) # need to build Dockerfile first
     [36](vscode-notebook-cell:?execution_count=1&line=36) # docker build -t cadie-ai-code-executor 
---> [37](vscode-notebook-cell:?execution_count=1&line=37) executor = PodCommandLineCodeExecutor(namespace="default", image="cadie-ai-code-executor");
     [38](vscode-notebook-cell:?execution_count=1&line=38) agent_x = ConversableAgent(
     [39](vscode-notebook-cell:?execution_count=1&line=39)     "CodeExecutor",
     [40](vscode-notebook-cell:?execution_count=1&line=40)     llm_config=False,  
     [41](vscode-notebook-cell:?execution_count=1&line=41)     code_execution_config={"executor": executor},  
     [42](vscode-notebook-cell:?execution_count=1&line=42)     human_input_mode="NEVER",  
     [43](vscode-notebook-cell:?execution_count=1&line=43) )
     [45](vscode-notebook-cell:?execution_count=1&line=45) analyst_system_message = open("system_prompt_code_writer.txt").read()

File ~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:158, in PodCommandLineCodeExecutor.__init__(self, image, pod_name, namespace, pod_spec, container_name, timeout, work_dir, kube_config_file, stop_container, execution_policies)
    [155](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:155) except ApiException as e:
    [156](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:156)     raise ValueError(f"Creating pod failed: {e}")
--> [158](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:158) self._wait_for_ready()
    [160](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:160) def cleanup() -> None:
    [161](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:161)     try:

File ~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:184, in PodCommandLineCodeExecutor._wait_for_ready(self, stop_time)
    [182](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:182) elapsed_time += stop_time
    [183](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:183) if elapsed_time > self._timeout:
--> [184](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:184)     raise ValueError(
    [185](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:185)         f"pod name {name} on namespace {namespace} is not Ready after timeout {self._timeout} seconds"
    [186](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:186)     )
    [187](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:187) try:
    [188](https://file+.vscode-resource.vscode-cdn.net/Users/mgardner/Development/autogen-analyst/~/Development/autogen-analyst/.venv/lib/python3.10/site-packages/autogen/coding/kubernetes/pod_commandline_code_executor.py:188)     pod_status = self._api_client.read_namespaced_pod_status(name, namespace)

ValueError: pod name autogen-code-exec-6ef95481-5f79-4198-999d-fdd4f457145b on namespace default is not Ready after timeout 60 seconds

It's like the pod is not getting spun up when I provide a different image - like it is searching for it and cannot find it.

Do I need to register the image with a Kubernetes Registry?

If I run kubectl get pods I see this for the failed pod:

NAME                                                     READY   STATUS             RESTARTS      AGE
autogen-code-exec-6ef95481-5f79-4198-999d-fdd4f457145b   0/1     ImagePullBackOff   0

Matt

[mattwg]

OK - I got it working! Was less an autogen and more a Kubernete's question!

We need to ensure that the image gets stored in minikube's registry (not the local registry). After that, when creating pods, Kubernetes will find the image in minikube's local registry because we set image_pull_policy="Never" in the pod spec.

# Set up Docker environment for Minikube use the Docker daemon running inside the Minikube VM not the local machine
eval $(minikube docker-env)

# Build the Docker image
docker build -t my-code-executor .

After building the local docker image, another key insight (thank you Claude!) was to include image_pull_policy="Never" in the pod spec so that the image was not pulled from a remote registry - that was the step that was timing out.

For future reference I used the Kubernetes pod configuration that worked for me was this:

from kubernetes import client

pod = client.V1Pod(
    metadata=client.V1ObjectMeta(
        name="autogen-code-executor" + str(uuid.uuid4()),
        namespace="default",
        annotations={"sidecar.istio.io/inject": "false"}
    ),
    spec=client.V1PodSpec(
        restart_policy="Never", 
        ## THIS IS FOR MY USE CASE WHERE I NEED TO COMMUNICATE WITH LOCAL POSTGRESQL RUNNING IN DOCKER CONTAINER ON A MAC (FOR NOW)
        host_aliases=[
            client.V1HostAlias(
                ip="xx.xx.xx.xx",  # Use your actual host IP address here
                hostnames=["host.docker.internal"]
            )
        ],
        containers=[
            client.V1Container(
                args=["-c", "while true;do sleep 5; done"],
                command=["/bin/sh"],
                name="code-executor",
                image="my-code-executor:latest",
                image_pull_policy="Never",  ## THIS IS IMPORTANT SINCE I WANT TO GET THE IMAGE LOCALLY (FOR NOW)
                env=[
                    client.V1EnvVar(name="PYTHONUNBUFFERED", value="1")
                ]
            )
        ]
    )
)