Platform for AI (PAI): Model Development and Deployment
What to know
PAI is Alibaba Cloud's machine-learning platform covering the model lifecycle: data preparation, training, and serving. Candidates should recognize its components at a conceptual level-interactive development environments (notebooks) for experimentation, distributed training for large jobs, and online inference/serving endpoints for deployment. Know the distinction between training (batch, GPU-heavy, produces artifacts) and inference (low-latency serving of a saved model). Understand artifact/model registry concepts, resource specs (GPU type, memory), autoscaling for endpoints, and the role of quantization to reduce serving cost.
Traps to watch
- Treating a notebook environment as production serving; notebooks are for development.
- Assuming training and inference use identical hardware; serving optimizes latency and cost, training optimizes throughput.
- Forgetting that a deployed endpoint needs the exact preprocessing used at training time-mismatches cause silent quality drops.
- Confusing autoscaling (replicas) with model scaling (parameters); they solve different problems.
- Overlooking that quantization trades small accuracy loss for large memory/latency gains.
- Believing a trained model is served automatically; it must be registered and deployed.
Rule of thumb
Separate concerns: develop in notebooks, train reproducibly, register the artifact, then deploy behind an autoscaling endpoint. Always ship preprocessing with the model.
Worked example
An endpoint times out under load. Check, in order: replica count and autoscaling policy, per-request latency (is the model too large for the GPU?), batch size, and whether quantization is feasible. Increasing replicas fixes concurrency; quantization fixes per-request cost.