Cloudflare Workers are a genuinely different piece of engineering: a V8 isolate rather than a container, cold starts measured in single-digit milliseconds, and a network that puts your code near everyone. When the workload fits, nothing else is close on latency or price.
The reason people go looking for alternatives is almost never that Workers are slow. It's that the isolate model has a specific shape, and their workload isn't that shape. Here's how to tell which case you're in, and where each one goes. I build PandaStack, which is one destination for exactly one of these cases.
The four constraints that send people away
1. CPU time per request
Workers cap CPU time per invocation. Wall-clock time waiting on a subrequest is generally fine; actual computation is what's metered. Image processing, PDF generation, cryptographic work, large parsing jobs, or anything running a model will hit this ceiling, and the failure is a terminated request rather than a slow one.
If you're compute-bound, you don't want a different edge platform — most have similar limits for the same physical reasons. You want a machine.
2. Node APIs and native modules
An isolate is not Node. Compatibility flags cover a growing subset of the Node API, but native modules — anything compiled, from image codecs to certain crypto libraries — cannot run at all, because there's no place to load a shared object.
This is a hard boundary rather than a performance trade-off. If your dependency tree contains a native module you can't replace with a pure-JS or WASM equivalent, the isolate model is closed to you.
3. Database connections
Traditional database drivers open a long-lived TCP connection and hold it. That's a poor match for a runtime with thousands of short-lived isolates, and the ecosystem's answer has been HTTP-based drivers and connection poolers. That works well — until you need a feature only the wire protocol exposes, like LISTEN/NOTIFY, session-level advisory locks, or COPY.
4. Long-lived state and connections
WebSocket servers, background workers, in-memory caches shared across requests, and long-running jobs all assume a process that outlives a request. Durable Objects address a real part of this and are excellent for coordination, but they're a programming model you adopt, not a place to run your existing server.
Two things to try before migrating anything
Some Workers limits are real walls and some are just the first thing you hit. Before assuming you need a different platform, check whether the work can move off the request path entirely — a queue plus a consumer turns a request that exceeds the CPU budget into an acknowledgement and a job, and that pattern fits the isolate model perfectly. A surprising share of planned migrations get cancelled at this step.
The second is caching. Workers sit in front of a cache that most teams under-use, and an expensive computation that runs once per unique input rather than once per request often stops being expensive at all. Neither of these helps if a native module is your blocker — that is a genuine wall — but both are worth an afternoon before committing to a platform change measured in weeks.
Where each case goes
- Still edge-shaped, want a different vendor: Deno Deploy, Fastly Compute, Vercel or Netlify edge functions, AWS Lambda@Edge. Similar constraints, different ergonomics and pricing — a lateral move, so be sure it fixes your actual problem.
- Need Node proper but still request-scoped: AWS Lambda, Google Cloud Run, Vercel's Node runtime. Full Node, native modules, longer durations, higher cold starts. Cloud Run in particular is a sensible middle ground — a real container, scale-to-zero, request-scoped billing.
- Compute-bound or holding connections: a container PaaS (Render, Railway, Fly.io, Koyeb) or a VM. One long-lived process, no per-request ceiling, no compatibility questions.
- Executing code you didn't write: microVM platforms — Fly Machines, PandaStack (mine). Each workload gets its own kernel rather than an isolate sharing a runtime, and a machine restores from a snapshot in roughly 179ms at p50, so per-request isolation stays affordable. This is the case where I'd argue for my own category and not otherwise.
The answer that's usually right: don't move everything
Most apps that outgrow Workers outgrow them in one specific place. The routing, auth checks, redirects, caching, and A/B logic are perfectly edge-shaped and benefit enormously from running close to users. It's the image resize, the report generation, or the WebSocket that doesn't fit.
Keep the Worker as the front door and give the heavy part somewhere to live. You keep the latency benefit for the 95% of requests that never need the origin, and the awkward 5% stops being a fight with the runtime.
// Worker stays as the edge router; the compute-bound path goes to an origin
export default {
async fetch(request, env) {
const url = new URL(request.url);
// Edge-shaped: cached, close to the user, sub-millisecond
if (url.pathname.startsWith("/api/config")) {
return new Response(env.CONFIG, { headers: { "cache-control": "max-age=60" } });
}
// Compute-bound: hand off to a process that has no CPU-time ceiling
if (url.pathname.startsWith("/api/render")) {
return fetch(env.ORIGIN + url.pathname, request);
}
return env.ASSETS.fetch(request);
},
};The short version
Hitting the CPU ceiling: you need a machine, not a different edge. Blocked by a native module: same answer, and no compatibility flag will save you. Fighting the database driver: try an HTTP driver or a pooler before migrating. Needing a long-lived process: a container PaaS. Running untrusted code: microVMs. And in most cases, keep the Worker at the front and move only the part that never belonged there.