Fly.io Machines

• CategoryServerless function
• LanguageAny (Docker)
• PricingPaid
• LicenseProprietary
• Created2017
• GitHub starsclosed
• Cold startfast
• PoPs30+
• Node-compatYes

Fly.io Machines run Docker containers in 30+ regions globally. Not a true edge runtime, closer to a globally-distributed PaaS. Cold starts are seconds (not milliseconds), but full Node / Python / Go / any-binary support. Strong on stateful and regional workloads.

• Apps requiring full Node / Python / Go runtimes globally
• Stateful workloads (Postgres, game servers, real-time)
• Multi-region deployments where regional latency beats true edge

Skip Fly Machines for true edge needs (sub-50ms cold start globally), V8 isolates win there. Skip for stateless API workloads where Workers / Edge Functions are simpler.

Fly Machines is "edge-flavoured" PaaS, not true edge runtime. Right answer when you need full Node / Python globally; wrong answer when you need V8-isolate performance.

A Machine is a Firecracker microVM running your Docker image, placed in one of Fly's thirty-plus regions. That is a fundamentally different unit from a V8 isolate and it changes what is possible. You get a whole operating system: any language, native modules, background processes, a real filesystem, and long-lived TCP and WebSocket connections that stay open as long as you need them. You also get real start times, seconds rather than milliseconds, though Machines can be kept warm or configured to auto-stop and auto-start on incoming requests, which recovers most of the idle-cost benefit. The accurate mental model is a globally distributed platform-as-a-service with fast placement, not an edge runtime. Against Cloudflare Workers the geography is coarser, tens of regions rather than hundreds of points of presence, but the capability available inside each location is effectively unlimited by comparison.

Most edge platforms are stateless by design and push you toward their own key-value or object stores. Fly goes the other way. Volumes attach to Machines, Postgres can run as an ordinary app, and you can pin compute to a region so data locality is something you decide rather than something the platform decides for you. That makes it a sensible home for game servers, collaborative back ends holding thousands of WebSockets open, queue workers, and anything that wants a database sitting in the same region as the code reading it. The cost is operational: you own the image, the health checks, the migrations, the backups, and the scaling policy, which is meaningfully more work than deploying a Worker. If your workload is stateless request handling, Cloudflare Workers or Vercel Edge is less effort and less money. Portability is a genuine consolation, since a Docker image runs anywhere.

• Official homepage →

The 30-min call is where your edge runtime choice becomes a real architecture, an integration plan with your data layer + auth + build pipeline, and a price range you can take to your stakeholders.