15. Accelerating model deployment velocity, Emmanuel Ameisen, Stripe

https://www.youtube.com/watch?v=tClDQk7DqlY&ab_channel=Tecton (opens in a new tab)

• Speeding up the art of ML model deployment

1. The value of redeployed models

   • Modelling and eng part to ship it ⇒ what happens when we have new features, or drift, etc
   • Any model that you train today will be obsolete tomorrow. By how much will it be obsolete?

   

   • Domain shift

   

   • Bottleneck when it comes to production

   

2. Skill set for regular ML deployments

   • Need to be 10x DS?

     

     • very operational work to bring value

3. Improving model release processes

   • automate the majority of the pipeline

     

   • another trick to de-risk the pipeline is to leverage shadow mode

     • deploy our prod model and also shadow, to observe how it behaves in production

   • Schedule it

     

   ⇒ Stripe radar technical guide (opens in a new tab)