Wiki
Core12 min read

Deployment and serving

Canary, blue-green and rolling releases; batch versus online serving; and the rollback that turns a bad version into a non-event.

A model that scores beautifully offline is still a hypothesis about production. Deployment is the experiment that tests it, and the whole design challenge is to run that experiment without exposing every user to it at once. What you want is a release in which a bad version is cheap, observable and reversible — a fraction of traffic, a clear signal, and a one-step rollback.

The simulator below puts a deliberately mediocre new version behind three strategies. Watch how canary limits exposure by ramping traffic, how blue-green switches everything at once, and how the rollback threshold decides between the two.

Choose a release strategy and a bad new version — watch the guardrail decide

bars = blended error · pale bars = % traffic on new version · dashed = threshold

Strategy

Rolled back at step 6 — blended error crossed 2.0%.

old-version error 1.0% · peak blended 2.01%

Canary limits exposure by ramping traffic, so a bad version trips the guardrail while only a fraction of users see it; blue-green switches everything at once and relies on a fast, clean rollback; rolling replaces instances in batches. The simulation assumes a single fixed error rate and instant rollback — real error rates are noisy and rollbacks take time.

Blast radius is a design choice

The same bad model can affect 1% of requests or 100%, for the same training cost, purely because of how it was released. Staged rollout is not a deployment convenience — it is the primary control on how much damage an inevitable mistake can do. Optimise for reversibility before you optimise for speed.

The three release patterns

  • Canary — serve the new version to a small slice of traffic, watch its metrics, and ramp up only if they hold. Best error containment; needs traffic routing and enough volume for the slice's metrics to be meaningful.
  • Blue-green — run two identical environments, switch traffic from blue to green atomically, and keep blue warm to switch back instantly. Clean rollback; the whole user base sees the switch at once.
  • Rolling — replace instances in batches. No duplicate environment needed, but old and new versions coexist during the rollout, so the system must be compatible with both.

For models specifically, two extra patterns matter. Shadow deployment sends production input to the new model without showing anyone its output, so you can compare offline against live data safely. A/B testing splits traffic and measures the business outcome, which is the only way to know the model actually helped rather than merely scored well.

Serving: batch, online, and streaming

  • Batch scoring runs on a schedule over stored data — cheapest, highest throughput, unacceptable latency for interactive use.
  • Online serving answers per-request with the model resident in memory or behind an inference server; latency budgets of tens of milliseconds rule out heavy preprocessing at request time.
  • Streaming scores events as they arrive, sitting between the two.

The recurring failure in all three is training-serving skew: the feature computed at training time differs from the one computed at serving time, often because the serving path cannot run the same expensive transformation. A feature store exists to make both paths read the same definition.

Careful

A model endpoint is not just a function. Inputs must be validated against the schema the model was trained on, because a silently renamed or unit-changed column produces confident garbage rather than an error. Version the model, the feature schema, and the code together; log every prediction with its inputs so incidents are interrogable; and make rollback a routine, one-command operation rather than an emergency procedure.

Cost, and the model you actually ship

Latency and cost are first-class constraints. A larger model that is two points more accurate may be unusable if it cannot meet the latency budget or its per-request cost dwarfs the value of the decision. Common levers are quantisation, distillation into a smaller student, caching, batching, and choosing a simpler model class when the accuracy gap does not justify the complexity. The best deployed model is the simplest one that clears the bar.

Illustrative vs real

The simulator uses one fixed error rate per version, instant rollback and a clean traffic split. Real error rates are noisy, rollback takes real time, and canary slices need enough volume to detect a difference. The traffic schedules and blended-error arithmetic are real; the environment is a simplification.

Check yourself

Eduspheria wiki · Data, MLOps & Deployment, MLOps

0 / 4 answered

  1. 1A canary sends 10% of traffic to a new version with a 4% error rate while the old version has 1%. What is the blended error rate (in percent)?
    %
    Numeric answer
  2. 2Which release pattern keeps a second full environment warm for an atomic, instant switch back?
    Multiple choice
  3. 3What pattern sends live production input to a new model without exposing its output to users?
    Short answer
  4. 4Online serving lets you run the same heavy preprocessing at request time that you used during training.
    True / false

From the assignment paper

Modeled on NITJ AI-511, Assignment/Quiz

0 / 6 answered

  1. 1A new loan model's output is logged for analysis but not used in any decision; a human officer decides. What is this deployment called?
    Multiple choice
  2. 2Which factor most directly decides whether a prediction service should be real-time or batch?
    Multiple choice
  3. 3For a deployed anti-spam filter, which change is an example of concept or data drift?
    Multiple choice
  4. 4A canary sends 20% of traffic to a new version with a 5% error rate while the old version has 2%. What is the blended error rate in percent?
    %
    Numeric answer
  5. 5Deployment should be treated as a one-off event that must work perfectly immediately after launch.
    True / false
  6. 6Which release pattern exposes a new version to a small slice of traffic and ramps up only if its metrics hold?
    Short answer

Where next: monitoring and drift — measuring, after release, whether the world has moved away from the model.