Rollout Strategies

Created: 2024-02-20 13:49
#quicknote

After we decided the Model Deployment, we have to consider how we will release new versions of the model or in general of the service.

Recreate strategy

existing versions in production are stopped and uninstalled
new versions are installed and tested, and then production resumes
simple and easy to implement, but service downtime expected
it could be used for batch applications where clients can be stopped from accessing the services for some time

Rolling deployments

used when we have n+1 deployments -> service is running on multiple nodes
updates is don eon one node at the time -> service is always available

Shadow deployments

new instance/cluster with new version
production traffic is duplicated and moved to the shadow deployment
shadow is validated for stability and performance
if validation is successful, we can move the service

Canary deployments

a separate cluster set up with the new version (canary)
part of the traffic is moved to the canary
canary is validated
roll back if there are issues
on successful validation, new version is deployed cross the board

Blue/Green deployments

new cluster with its own resources (nodes) set up for the new version
the gateway/load balancer is switched from the existing to the new version
if issues found, gateway can be switched back, otherwise, older version can be retired
it is easy to move between versions, but additional resources are needed for the new cluster

Tags

#mlops #ml #linkedin