Komputer makes most ordinary Linux OCI workloads transparently durable without requiring an application framework. It runs each computer under gVisor, records nondeterministic inputs and externally visible effects in local S2 streams, and restores from the latest fenced epoch after failure. CPU work that has not produced an observed effect may be lost and re-executed.
gVisor is the sole execution substrate. Its Sentry application kernel gives Komputer one semantic boundary for clocks, entropy, signals, readiness, filesystem observations, sockets, wakeups, and process I/O while preserving the familiar OCI userspace contract. The compatibility target is intentionally “most ordinary Linux OCI workloads,” not every Linux syscall, device, or privileged container behavior.
Each computer owns separate S2 streams for control, incremental checkpoint data, stdin, stdout, and stderr. Local in-memory or filesystem-backed s2-lite is the development durability tier. S3-compatible object storage, exercised with RustFS, holds asynchronously compacted checkpoint bases and is never on the small-effect acknowledgement path.
The active engineering spike starts from upstream gVisor full-state checkpoint/restore and asks whether a narrow, maintainable extension can:
- Track dirty application pages, including both application writes and Sentry writes into application memory.
- Represent changed Sentry objects without instrumenting mutation sites throughout the application kernel.
- Seal a short incremental epoch to local S2 before releasing effects.
- Reconstruct an upstream-compatible full checkpoint from a compacted base plus ordered deltas.
See the architecture, the gVisor boundary, the durability protocol, and the roadmap. The previous WASM implementation in ../core remains research input, not a runtime dependency.