Coursera

Week 2

Deployment strategies

Static approaches:

• Blue/Green
  • Swap prediction request traffic
  • Easy Rollback
• Shadow/Challenger
  • Parallel prediction request traffic
  • Validate new version without impact
• Canary
  • Split traffic
  • Target smaller specific users/groups
  • Shorter validation cycles
  • Minimize risk of low performing model
• A/B
  • Split traffic
  • Target larger users/groups OR Distribute % of traffic
  • Longer validation cycles
  • Minimize risk of low performing model

Dynamic approaches:

• Multi-Armed bandits
  • Dynamic testing method for model version using Reinforcement Learning
  • Exploit and Explore:
    • Exploit: Reward the winning model with more traffic
    • Explore: Continue to send traffic to the non-winning model(s) in case behaviour changes.