3 papers
Memrun is a Python SDK and platform for deploying data-intensive services that need warm memory, sticky routing, and bounded concurrency. Deploy a handler with a decorator. The platform handles NVMe caching, locality-aware routing via Kafka, and worker lifecycle on commodity hardware.
Serverless platforms optimized for stateless, elastic compute. But memory-bound workloads need the opposite: warm RAM, fast local disk, and sticky routing. This isn't better serverless. It's a different abstraction entirely. We describe the programming model, the platform mechanics, and how memrun implements it.
Serverless removed servers from the developer's mental model, but it also removed state from first-class support. For data-intensive workloads, that tradeoff is increasingly costly. We built memrun to prove that warm, locality-aware workers can remain serverless in feel while eliminating the state tax that stateless platforms impose.