Serving modes

Deployment e ciclo di vita in MLOps

Nemanja Radojkovic

Senior Machine Learning Engineer

Serving and serving mode

• Providing prediction service == Model serving
• Implementation of a specific type of serving == Serving mode

 

Choose carefully!

Batch prediction: Keep it simple

• Batch prediction is the simplest
• If use case allows it, go for it
• Good fit: monthly generation of sales forecasts

Acceptable latency

What is acceptable?

• < 1 hour?
• < 1 minute?
• < 1 second?
• < 1 millisecond?

Near-real time prediction a.k.a. Stream processing

Acceptable latency ~= X minutes

Also known as stream processing (requests and responses form "data streams")

Real-time prediction

Acceptable latency < 1 sec

Example:

• Credit card fraud detection
• Late prediction as good as useless

When latency is a priority

• Weaker, but faster model more valuable than a stronger, but slower one
• Models deployed to end user devices to reduce latency => "edge deployment"
  • ML-infused smartphone apps:
    • navigation apps
    • unlocking via facial recognition
    • image filters

Let's practice!

Deployment e ciclo di vita in MLOps