Documentation
Membrane documentation
Install, run, secure, and operate Membrane, the shared KV-cache for LLM serving. Start with the quickstart, or jump to the guide for what you are doing.
Quickstart
From a fresh clone to a running node and your first cache hit in about five minutes.
Deploy to production
Docker, Kubernetes, TLS, persistence, and the settings a production node needs.
Security model
Authentication, tenant isolation, encryption at rest, and audit logging.
Get started
Guides
Run Membrane in production as a container, on Docker Compose, on Kubernetes, or under systemd. This guide covers installation, configuration,…
How Membrane authenticates callers, authorizes every request, isolates tenants, protects data in transit and at rest, and records administrative…
Choose how much durability each write gets. A single node always writes locally; in a cluster, the consistency level decides whether a write also…
Ask a node what an inference engine needs to know before it runs a prompt: what is already cached, where it lives, and whether to compute the rest…
Run prefill and decode on different machines. Prefill is compute-bound and decode is memory-bound, so they run best on different hardware. A…
Extend Membrane without forking it. Compute backends, authenticators, content stores, persistence backends, eviction policies, placement policies,…
Operations
Recommended service-level objectives for a Membrane deployment, the Prometheus queries that measure them, and the alerts that protect them. Tune…
Size memory, storage, Redis, and node count for a Membrane deployment. A node's memory budget holds the KV cache for the prompts you want to keep…
Protect warm-cache performance through node, volume, and regional failures. Membrane is a cache: every fragment can be recomputed by running…
Upgrade a Membrane deployment without downtime. Patch and minor releases roll out one node at a time while the cluster keeps serving; major…
The on-call runbook for Membrane: how to triage an alert, diagnose the common failure modes, and communicate during an incident. Pair it with SLOs…
Reference
How Membrane is built: the request path, the components of a node, the cluster layer, and the design principles behind them. Membrane separates the…
The specification of Membrane's v5 formats: fragment identity, the wire message, canonical byte framing, cluster snapshots, and tombstones. It is…
What you can depend on across Membrane releases. The public API follows Semantic Versioning, with the guarantees below, so integrations built on…
The runtimes, serving engines, transports, and GPUs supported by the Membrane 3.x series. This page is the single source of truth for…
How maintainers cut, publish, and roll back a Membrane release. A release is the version in pyproject.toml plus its vX.Y.Z git tag. On every tag,…