Edge Computing

Edge runtimes move logic closer to users—with strict CPU/time limits, different APIs than full Node, and tenant isolation requirements.

When to Offer This Workflow

Trigger conditions:

• Auth, redirects, or personalization at the CDN layer
• HTML rewriting, A/B assignment, or bot mitigation at the edge
• Global latency SLOs for read-heavy paths

Initial offer:

Use six stages: (1) workload fit, (2) edge vs origin split, (3) data & state, (4) security & tenancy, (5) limits & cost, (6) testing & rollout). Confirm platform (Workers, Lambda@Edge, Fastly Compute, etc.).

Stage 1: Workload Fit

Goal: Prefer short, CPU-light, request-scoped work—not long jobs or huge body buffering.

Exit condition: Explicit list of what remains on origin (heavy SSR, large uploads).

Stage 2: Edge vs Origin Split

Goal: Document what runs where: geo headers, redirects, cache key logic, A/B bucketing, partial HTML injection.

Practices

• Cache Vary and cookie behavior documented to avoid wrong personalization leakage

Stage 3: Data & State

Goal: If using edge KV/Durable Objects/regional stores, state consistency (eventual vs strong) and rate of round-trips to origin.

Stage 4: Security & Tenancy

Goal: Validate JWT/session at edge; isolate tenants; never embed secrets in deploy bundles visible to clients.

Stage 5: Limits & Cost

Goal: Wall-clock CPU limits, request size caps, egress pricing; graceful degradation or fallback to origin.

Stage 6: Testing & Rollout

Goal: Canary per region/PoP; synthetics from multiple locations; compare p95 vs origin-only path.

Final Review Checklist

• Workload fits edge constraints
• Edge vs origin responsibilities documented
• State/consistency strategy clear
• Multi-tenant security reviewed
• Limits, cost, fallback documented
• Multi-region validation performed

Tips for Effective Guidance

• Edge runtimes differ from full Node—verify available APIs (fs, streams, crypto).
• Read platform-specific cold-start and isolate model docs.

Handling Deviations

• Hybrid: edge for headers/cache only; heavy compute stays on origin.