Skip to content

/talks ·

Cocktail: A Multidimensional Optimization for Model Serving in Cloud

19th USENIX Symposium on Networked Systems Design and Implementation · Renton, WA

NSDI

What this talk is about

Cocktail is an ensemble-based model-serving framework for the public cloud that simultaneously optimises three things production teams actually care about:

  • Cost — by leveraging transient cloud instances and dynamically selecting the smallest viable subset of models per request.
  • Accuracy — by combining predictions across the active ensemble with a weighted majority vote.
  • Latency / SLO — by predicting per-request resource needs and provisioning ahead of demand spikes.

The talk walks through the per-component optimisations and the joint control loop that ties them together. It also covers the decoupling between model-selection (a slow loop) and request-routing (a fast loop) that makes the system responsive without flapping.

Quick references