I got a real Linux VM to cold boot in about 4ms.
In production this gives me less than 150ms response times for real Node.js applications. Not edge functions, but full Node.js or Bun apps. The same approach works with PHP and Python, and I can add more languages easily, because these are full virtual machines.
This is based on unikernels, with a layered distro model that lets me run real Node, Bun, Python and PHP applications with full package support.
Once I had that running reliably, I realized I could build a serverless server platform around it instead of just another function runtime.
I am building a serverless platform for running full servers, not functions.
The idea is simple. Instead of containers or pods, each workload runs as a tiny VM that can scale to zero and wake in a few milliseconds.
By splitting the system into distro layers, framework layers and application layers, I keep disk usage low and boot time extremely small.
This gives stronger isolation than containers, while still feeling like normal Linux from the inside.
Right now I support Node.js, Bun, Python and PHP.
You can install any dependency you want.
I collect observability data using OpenTelemetry and eBPF, and the load balancer can scale based on HTTP activity or message queue events.
It already runs across multiple physical clusters, including air gapped environments.
There is a CLI, a load balancer and the serverless server component.
The CLI can run and build any layer locally on Windows, macOS or Linux, using the same execution model as the servers, so nothing is being faked through Docker or Podman.
The system can run VMs without administrator or root permissions, and it can even run on hardware or virtual machines where virtualization is disabled. Performance takes a heavy hit in that case, but it still works.
There is still a lot missing.
I have persistent storage, so applications that need state, such as databases, can run inside the platform.
But I am still working on TCP and UDP support so I can run full stacks like Postgres or Redis.
When testing Postgres inside a tiny VM I can cold boot it in less than 250ms, but I still need a clean way to expose and manage these services across regions.
If anyone wants to dive deeper into the tech or discuss the implementation details, feel free to email me at allan at openiap.io
Comments URL: https://news.ycombinator.com/item?id=45952591
Points: 1
# Comments: 0
Source: news.ycombinator.com
