Dylan Di Filippo
Dylan Di FilippoAI journal

Modal evaluates candidate endpoints on mirrored traffic

ModalAgents and applications

Endpoint Candidates creates alternative deployments alongside a Dedicated Endpoint. They can receive sampled copies of traffic to compare quality, latency, throughput and cost while production continues to supply the answer users actually see.

Feature release / rollout

Endpoint Candidates creates alternative deployments alongside a Dedicated Endpoint. They can receive sampled copies of traffic to compare quality, latency, throughput and cost while production continues to supply the answer users actually see.

The feature is generally available to all Modal workspaces. Each candidate has its own evaluation URL, and promotion retains the production endpoint URL. Mirrored compute is billed at standard rates.

Mirroring is best effort. Under memory pressure or when candidates are slow, copies may be canceled to protect production. Received samples may therefore omit difficult requests. Real traffic can also contain sensitive data.

Testing models on realistic traffic should not require exposing users directly to them.

← Back to this weekly edition

A journal by Dylan Di Filippo. Facts, vendor claims and editorial analysis are kept separate. Coverage is not exhaustive.