advanced Estimated learning time: 5 h

10.10 Scaling inference

Your service survives a hundred times the traffic.

Before:05. Classical Machine Learning

Scaling inference is standard web engineering applied to models — horizontal replicas, autoscaling, caching, request deduplication — plus model-specific tricks like quantization for latency. Load testing reveals the real ceiling. It sits near the module's end because it stresses everything built. The habit is testing to failure on purpose in daylight, because the alternative is discovering the capacity limit from an outage, at night, with users watching.

Work through these

  • Horizontal scaling and autoscaling policies

    Adding more instances behind a load balancer, with rules deciding when. For prediction services the rules need to account for slow model loading.

  • Caching layers and request deduplication

    Storing previous answers, and recognising when the same request arrives twice. On many workloads caching removes most of the traffic.

  • Quantization and distillation for latency

    Making the model itself smaller and faster, using the compression techniques from earlier modules. This is frequently cheaper than adding hardware.

  • Load testing and capacity planning

    Generating realistic load and measuring what breaks first tells you what capacity you actually have. Estimating it instead is how outages happen at launch.

Sign in to keep your progress.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.