You already know: request latency, load balancing, timeouts and retries (D1 Networking); the model registry and lineage (ML lifecycle, Platform tools); online feature lookups and freshness (Feature infrastructure); release gates, canaries and A/B tests for LLM features (D10 Production eval pipeline).
Why this matters for Atlas. Deployment is where a model meets real traffic, latency budgets and failure. Two decisions dominate: how predictions are produced (precomputed in batch, computed per request, or queued) and how a new version replaces the old one (all at once, side by side, gradually). Get the first wrong and you pay for latency or staleness you didn't need; get the second wrong and every model bug becomes an outage. This lesson sizes the serving patterns with numbers, builds the rollout mechanisms (shadow, canary with automatic rollback, sticky traffic splits) and makes rollback boring.
1. Batch, online or asynchronous?
The question. Should the struggle predictor run when the student opens a page, or every night?
Start from nothing. Three patterns: