DockerImageBuild state machine implementation

[stephensmith-aws]

Description of changes:

Implement docker image build piece of state machine

[petermuller]

Ah I see what you mean now- for these, we can use the names as they're defined in the CreateModelRequest, which we can pass as direct input into the state machine. I would suggest a few things here:

1. Let's define any aws clients at the top of the file, similar to how they're done here:

   LISA/lambda/models/state_machine/delete_model.py

   Lines 28 to 30 in 283dd8f

   	cloudformation = boto3.client("cloudformation", region_name=os.environ["AWS_REGION"], config=retry_config)
   	dynamodb = boto3.resource("dynamodb", region_name=os.environ["AWS_REGION"], config=retry_config)
   	ddb_table = dynamodb.Table(os.environ["MODEL_TABLE_NAME"])

   (these variables are already set from the CDK definition- we can add more variables as needed)
2. For this create_infra variable, let's make it be True if all of AutoScalingConfig, ContainerConfig, InferenceContainer, InstanceType, and LoadBalancerConfig are not None. For False, all of these need to be None, and then let's throw an exception if they're partially configured (some null, some not). Basically we have two use cases:
   1. We define the ECS infrastructure which is all of those fields, and ModelUrl must also be none
   2. We define a LiteLLM-only entry, which requires all of the others to be null, and we'd set the ModelUrl as specified in the request. For the initial revision here, let's just worry about the first case.
3. In this step specifically, let's make sure that we set the model state to Creating in DDB, and we can add other details, like last modified date, the entire request blob, and anything else that seems like it'll be useful to have around. We can also add all that to the output_dict so that future states can have that info without calling out to ddb

[petermuller]

I'm not sure which errors we're expecting, but I think we can just let them boil out to the state machine. From the state machine definition after it's deployed, it automatically handles some form of sdk errors, but I'm not 100% sure if those are strictly lambda or if they're any sdk call:

      "Retry": [
        {
          "ErrorEquals": [
            "Lambda.ClientExecutionTimeoutException",
            "Lambda.ServiceException",
            "Lambda.AWSLambdaException",
            "Lambda.SdkClientException"
          ],
          "IntervalSeconds": 2,
          "MaxAttempts": 6,
          "BackoffRate": 2
        }
      ],

[petermuller]

same note for clients, let's define them at the top of the file with the region and retry config

[petermuller]

I think we should probably keep W503 to keep with the pep standards https://www.flake8rules.com/rules/W503.html

where did this happen?

[stephensmith-aws]

happened in create model state machine implementation. One of our formatters wants to put each condition on its own line, and flake8 doesn't. Couldn't make them both happy at the same time.
lambda/models/state-machine/create-model.py:38

[petermuller]

Oh for that we could try putting the and's at the end of the line. i don't think there's anything wrong with needing the backslash at the end to complete the line either. otherwise we can just do an inline ignore so we ignore the one line, but not the rest. so adding a comment like # noqa: W503 on the specific line could work