Startup performance
What determines how long a sandbox takes to serve its first page, and how to make yours faster.
Startup time is dominated by your application, not by the environment around it. Understanding the split is what lets you act on it.
The two starts
First start. The project has never run: dependencies are installed and your framework builds its output from nothing. This is the slowest case, and it happens once.
Subsequent starts. Dependencies and build caches are already on persistent storage, so both steps are skipped or reduced to an incremental rebuild. This is what you experience day to day.
Where the time goes
Ordered by weight, on a typical Node project:
- Your framework's own startup — loading the runtime and the framework before any of your code is compiled. This is the largest fixed cost, and no platform setting changes it.
- The first compilation — building the route you land on. Heavily dependent on your dependency graph, and cached afterwards.
- Dependency installation — first start only.
- The environment itself — starting the isolated environment, attaching storage, and wiring the preview endpoint. A small, roughly constant share.
Once the first page has been served, navigating within your application is near-instant: subsequent routes compile incrementally.
Making your project start faster
- Keep your dependency graph lean. It is the single biggest lever, and it compounds across installation, compilation and framework startup.
- Use your framework's fastest bundler. Modern defaults are usually much faster on incremental rebuilds; check your framework's release notes rather than carrying an old flag forward.
- Do not clear build caches. They are preserved on purpose between sessions. Deleting your framework's cache directory turns every start into a first start.
- Avoid heavy work at module load. Anything at the top level of your entry point runs before your first page can be served.
- Declare the right port so the platform detects your server as soon as it listens instead of waiting for a probe.
Figures per framework
Measured figures for each template are being collected on identical hardware and will be published here. We would rather leave this table empty than fill it with estimates.