Edge Computing Runtime Lab (Interactive)
Route a Sydney or Tokyo user through Anycast PoPs and compare V8 isolates, microVMs, containers, and VMs. Computes speed-of-light RTT from user-to-PoP distance, adds per-runtime cold-start probability, and contrasts local edge reads with origin write pass-through.
Edge Runtime Latency Physics Lab
Beat the speed of light: route a Sydney user through Anycast PoPs and compare V8 isolates against containers and VMs.
Edge p50
11.1 ms
meets 15ms SLA (local read)
Centralized cloud p50
171 ms
Virginia origin RTT 159 ms alone
Nearest PoP
805 km
8.1 ms RTT at 5ms/1,000km
Cold start cost
+0.5 ms
p99 11.6 ms
Workloads per node
50,000
RAM for 10k tenants
39 GiB
Latency saved vs origin
160 ms
Isolates share one permanently warm V8 process, so their cold slot costs microseconds; a container image pull or VM boot lands on the p99 tail. Reads served from a local edge SQLite/KV replica never cross an ocean, but toggling to a write pays the full round trip to the primary region — exactly why mutating workloads at the edge need CRDTs or Durable Object actors instead of naive pass-through.
How It Works Under the Hood
Fiber propagates at roughly 5ms per 1,000km one-way, so a Sydney user hitting a Virginia origin pays 160ms+ of pure physics before any code runs. Anycast BGP re-anchors compute to the nearest of 300+ PoPs, shrinking that distance, while V8 isolates instantiate sandboxed heaps inside an already-warm process in under 0.5ms instead of seconds for containers or minutes for VMs. The catch is state: reads from edge KV or SQLite replicas stay local, but writes route across an ocean to the primary, which is why coordinated writes need CRDTs or Durable Objects.
Core Architectural Principles
- Round-trip latency = distance / 100 ms; PoP count collapses user-to-compute kilometers non-linearly.
- Cold starts: isolate 0.5ms, microVM ~500ms, container seconds, VM minutes, landing on the p99 tail.
- Edge reads hit local replicas in ~1ms; write pass-through pays the full WAN round trip to the primary region.
Draw the topology user to Anycast to nearest PoP to local replica, and quote the 5ms-per-1000km physics so your latency numbers are defensible. Contrast V8 isolates against containers on cold start, RAM, and density. Then address state proactively: reads stay local, writes use CRDTs or Durable Object actors, and heavy compute or strict multi-table locking is an argument for keeping the workload in a region.
Edge runtimes deliver sub-15ms global latency with microsecond cold starts, but impose tight CPU/memory limits and eventual-consistency semantics that centralized clouds avoid.