Edge Computing

From Cache to Compute

 classic edge: serves cached copies; misses travel to origin.
 edge compute: YOUR CODE executes AT the edge location.

 user ──15ms──► [edge: runs your function]
                    │ can now:
                    │ - transform requests/responses
                    │ - auth/route/personalize LOCALLY
                    │ - call origin only when truly needed

 the shift: from "cache decisions made centrally" to
 "logic distributed globally."

What Edge Compute Actually Is

 deployment model: upload code once → replicated to ~300 POPs
 execution model: serverless-style — no servers to manage,
                  scale is automatic (per-request isolation)
 latency model:   compute happens ~10-30ms from users
                  instead of ~100-300ms at a central region

 platforms: Cloudflare Workers, Fastly Compute@Edge,
            AWS Lambda@Edge/CloudFront Functions, Akamai EdgeWorkers
 
 all share a shape: V8-isolate-class runtimes, sub-millisecond
 cold starts, aggressive per-request resource caps.

The Workloads That Belong There

Workload	Why edge wins
Auth token verification	Reject bad requests before they cross oceans
A/B routing / personalization	Decision latency IS UX
Request/response transforms	Header rewrites, image negotiation
API aggregation for mobile	Parallel fan-out from near the user
Bot filtering / WAF logic	Attack mitigation at ingress
Real-time fan-out	Chat/presence relayed regionally

 pattern uniting them: SHORT, STATELESS, LATENCY-SENSITIVE.
 anything needing >50ms CPU or heavy state doesn't fit —
 see constraints below before falling in love.

The Constraints That Define It

 ✗ LONG CPU WORK:      CPU-time caps (ms to low seconds)
 ✗ HEAVY DEPENDENCIES: small bundles; limited native libs
 ✗ LOCAL DISK:         mostly none; storage is networked/KV
 ✗ TRADITIONAL DATABASES: connection pooling across 300 POPs
                        against central DBs = anti-pattern;
                        use HTTP APIs/caches/regional replicas
 ✗ STRONG CONSISTENCY: writes still need an authoritative region

 edge compute EXTENDS the architecture (fast paths + gateways);
 it does not REPLACE regions of record.

The Architecture Pattern

 [user] ──► [EDGE: auth, routing, personalization, cache mgmt]
                 │ allowed & cacheable → serve locally
                 │ needs truth
                 ▼
          [REGIONAL CORE: databases, transactions, workers]

 edge as intelligent filter + accelerator;
 core as system of record. most successful edge deployments
 are exactly this two-tier story.

Interview Framing

“Use edge computing for checkout?” tests constraint-awareness. Scored answer distinguishes layers: token validation and cart pricing DISPLAY at edge (fast reads), but payment AUTHORIZATION stays regional (consistency + compliance). Candidates who move everything to the edge fail; those who identify the fast-path subset pass. Name CPU/stateless constraints explicitly — that’s the depth signal.