Platform for AI (PAI) and Model Lifecycle
PAI is Alibaba Cloud's machine learning platform covering the model lifecycle: data preparation, training, tuning, deployment, and monitoring. Know the main components at a conceptual level. PAI-DSW provides interactive development notebooks for experimentation. PAI-DLC handles distributed training jobs. PAI-EAS serves models as online inference endpoints with scaling. PAI-Designer offers a visual, drag-and-drop pipeline for no-code or low-code workflows. Understand that these components are complementary: experiment in DSW, train at scale with DLC, deploy with EAS.
Lifecycle thinking matters: data labeling and feature preparation precede training; evaluation gates promotion; deployment includes autoscaling and canary or blue-green rollout; monitoring covers latency, throughput, and drift, which triggers retraining. Know that fine-tuning (including parameter-efficient methods) adapts a pretrained model with less data than training from scratch.
Traps: questions may suggest EAS is for training (it is for serving) or that DSW is a production serving environment (it is for development). Another trap is treating drift monitoring as optional; production models degrade as data shifts. Watch for confusion between batch inference and real-time endpoints.
Rule of thumb: pick the PAI component by lifecycle stage: develop in DSW, train in DLC, serve in EAS, and orchestrate visually in Designer.
Worked example: a team prototypes a classifier in a notebook, then needs to train on a large dataset across multiple GPUs, then serve low-latency predictions. The path is DSW for prototyping, DLC for distributed training, EAS for online serving with autoscaling.