Skip to content

Latest commit

 

History

History

04-deployment

README.md

Model Deployment

Ask yourself:

  • Need to have predictions inmediately or can wait for them (day or weeks)?

Batch Offline Deployment

  • Run regularly against batches of data (e.g. daily, weekly, etc)
  • Have some sort of DB with the data.
  • Have a (scoring) job running that host the model inside of it (e.g. AWS Batch, ECS)
  • Apply the model to all the data from a previous interval/batch.
  • Write the predictions to another DB.
  • Other applications use those predictions.
  • For applications where is not necessarily to know a result inmediately.

Online Deployment

Model is always available, up and running all the time.

Web Service

  • Web service that contains the model, gets the data in request and send prediction back on a response.
  • Backend talks to the web service that host the model.
  • Needs to be up and running all the time.
  • 1 to 1 relationship: there is a connection kept alive while processing the request.
  • For applications where the use needs the result inmediately.

Streaming

  • There are producers and consumers.
  • Producers pushes some events to an event stream/broker.
  • Consumers fetch data from the event stream and then do something the data (multiple models) consume the data.
  • Then, consumers can also send their predictions to another event stream/broker.
  • Producers send events but doesn't actually wait/expect for a result.
  • 1 to many or many to many relationship.

E.g. Content moderation (e.g. Youtube videos, Twitch Streaming). Copyright violation, NSFW, Violence, etc.

Homework

Terraform

Initialize terraform (installs providers plugin from terraform registry)

terraform init

Format configuration files for readability and consistency.

terraform fmt

Validates terraform configuration.

terraform validate

Creates and updates infrastructure according to Terraform configuration. First, prints out the execution plan which describes actions Terraform will take.

terraform apply

Inspect state:

terraform show

Docker

Build the image:

docker build -t mlopz-zoomcamp:hw4 .

Execute the script as depicted below:

docker run --rm -v ~/.aws:/root/.aws \
    -e DATE="2021-03-01" \
    -e AWS_PROFILE="nicolas-devops" \
    -e BUCKET="mlops-zoomcamp-nicolas" \
    mlops-zoomcamp:hw4