This message was deleted.

Hi @Dawid Smoleń To deploy to AWS sagemaker, you would have to configure your AWS cred on the Azure machine. BentoML uses

boto3

under the hood for AWS operations. The order of how

boto3

gets credentials, is first look at the environment var

AWS_ACCESS_KEY_ID

and

AWS_SECRET_ACCESS_KEY

, if no envvars are set, it will look into

./aws/credentials

and then

./aws/config

. You can find relevant documentations at https://boto3.amazonaws.com/v1/documentation/api/latest/guide/configuration.html#guide-configuration just reply here or

@

me, if you have any questions

Thanks! I see, I had problem with ~/.aws/ dir because everything is happening on CI machine on Azure DevOps. Then I figured out I can set the variables you mentioned, although I don't have it set fully automatically I hope I'll be able to do that. Thanks for your help!

Let me know how it goes

actually I am wondering how can I update or delete sagemaker deployment, when it was created on Azure CI machine, that I'm not sure I have access to. Updating deployment requires

--bento-service-bundle

but as far as I understand after new commit we are creating a new fresh instance and I have no information about

name:version

created by the previous pipeline, am I right?

@Dawid Smoleń Typically, when deploy a model from local machine, you just need

bentoml sagemaker update same_deployment_name --options

and

bentoml sagemaker delete same_deployment_name

for update and delete operations. BentoML's backend service, Yatai, uses a local

sqlite

db as default for managing models and deployments. Yatai service could also take remote sqlite file (from s3) or postgres db as its storage backend. There is a CLI command to start a local Yatai services,

bentoml yatai-service-start

that can takes

db-url

and

repo-base-url

(where the model will be stored, default is local file system) as options. For your use case with Azure CI, I think a good option for now is have a remote postgres or sqlite db. I would first start an Yatai service with

bentoml yatai-service-start --db-url=MY_REMOTE_DB_URL

, and then either run

bentoml config set yatai_service.url=YATAI_SERVICE_ADDRESS

(for local, it is 127.0.0.1:50051) . When you set yatai_service_url, all BentoML operations afterward, will use this Yatai service. Since you have the remote db setup, all of the model and deployment operations will be stored for future use. You can also assign

repo-base-url

to a s3 bucket, so all of the model version will be also stored as well. We are working on have good documentation for this, for now, you can checkout our test for reference. For using postgres db(https://github.com/bentoml/BentoML/blob/3d4f7b9db101af1feaac7051127ce8cce4931b1d/e2e_tests/yatai_server/test_postgres_local_fs.py#L16) We are actively working on make this process much smoother and easier. We will soon dockerlize the yatai service and you can deploy it as a service for your CI operations. When you run your CI, you just need to pass yatai service url as an envvar, without doing the work I mentioned above. Let me know if this help you move forward or not.