Shadow mode deployment vs "dark launch"

Course Q&A Machine Learning in Production
1

Would you consider “shadow mode deployment” to be the same thing as “dark launch”? According to Deploying Machine Learning Models in Shadow Mode (which also points to https://cloud.google.com/blog/products/gcp/cre-life-lessons-what-is-a-dark-launch-and-what-does-it-do-for-me) they seem to be used interchangeably but I’m curious from this community what your thoughts on it are.

Fwiw, I’ve seen “dark launch” used as another way to say “silent launch” at certain companies – that is, a product (not necessarily a model) is deployed to production and is user-facing, but just not announced. That is not how Andrew Ng describes shadow mode deployments in Deployment patterns, so I’d like to see what y’all think.

2

According to shadow mode deployment, the candidate model (i.e. your new neural network model) and the existing api receive calls from the caller. The existing api returns the response to the caller.

The logging / evaluation infrastructure compares both outputs from both candidate and existing endpoint to to finetune your new model and determine when it’s ready to replace the old system (this could be a manual task / a rule based system). In the case of an online learning implementation, this setting can be used for training your new candidate model as well.

The goal is to shadow an existing system to replace with a new model. This is what this link says as well.