Netlify processes roughly a billion Edge Functions daily — personalization, routing, auth — and has rebuilt the entire execution layer from scratch. The old architecture sent requests out over the public internet to a hosted V8 isolate service. The new one runs Firecracker MicroVMs inside Netlify's own edge network, partnering with Unikraft for the VM lifecycle management. Median latency dropped from 25-40ms to 5-6ms. The p99 improved 47.4%. Availability sits at 99.998%. The technical architecture is methodical. Each request that matches an edge function route gets a machine specification written at the edge node — naming a runtime image, platform image, and function image with CPU/memory/connection limits. The spec's hash becomes a service ID, meaning different deploys or environment variables never share a MicroVM. Rendezvous hashing pins services to specific compute nodes for cache warmth, with a spillover mechanism that relaxes stickiness above a traffic threshold to prevent hotspots. Firecracker MicroVMs boot in under a millisecond and reach p99 start in about 2ms because they launch a stripped-down Linux kernel, not a full OS. Function bundles mount as uncompressed EROFS images and are memory-mapped — the VM reads only what it needs. After the JavaScript server initializes, a snapshot is taken. Idle MicroVMs scale to zero; the next invocation restores from snapshot without waiting for full memory reload. Cold starts affect about 1.2% of invocations and add roughly 9ms. The isolation story is the real architectural argument. Netlify states plainly that V8 isolates, despite the name, do not provide hardware-level isolation. A compromised deploy in a MicroVM cannot escape to poison other tenants or the compute layer itself. Each deploy gets its own VM boundary. This is a structural security upgrade, not a performance-only play. The operational design reflects hard-won lessons. Compute nodes run local DNS resolvers after discovering that DNS latency was a real production problem. Circuit breakers handle rerouting and decommissioning. Deployments use blue-green fleet replacement — a new fleet scales up alongside the running one and takes traffic only when healthy. Metrics now cover boot time, time to first port open, and time to user code execution. The customer-facing contract is unchanged: same URL imports, npm packages, Node built-ins, netlify.toml declarations, same pricing, no migration step. The entire migration happened transparently under production traffic. This is notable because invisible infrastructure migrations at this scale — a billion daily invocations — tend to be where things break. The strategic implications extend beyond Netlify. This is a public datapoint in the V8 isolates vs. MicroVM debate that has shaped edge compute for years. Cloudflare Workers built on V8 isolates; Netlify just moved off them. The 5x latency improvement and the isolation argument together make a case that the MicroVM approach, once considered too heavy for edge workloads, is now viable at scale — provided the snapshot/restore lifecycle is fast enough.
