This page lists every page in the Snapshot documentation, grouped by section.
- What Is Snapshot? - What Snapshot is, why it exists, and where it fits.
- Core Concepts - Checkpoint and restore, resources, the workload contract, the snapshot store, and multi-GPU.
- Architecture - The components, the Kubernetes resources that drive them, and how a checkpoint and a restore proceed.
- Performance - Cold start and restore times for vLLM, SGLang, and TensorRT-LLM on one to eight B200 GPUs.
- Limitations and Known Issues - Current limitations and what is on the roadmap.
- Prerequisites - Minimum GPU driver and other dependency versions, per feature.
- Installation - Helm install and uninstall.
- Quickstart: Checkpoint and Restore One Pod - Checkpoint and restore one pod end to end.
- Using Snapshot - The checkpoint and restore flow, with a recipe for each inference framework.
- Prepare a vLLM Workload - Make a vLLM workload snapshot-ready.
- Prepare an SGLang Workload - Make an SGLang workload snapshot-ready.
- Prepare a TensorRT-LLM Workload - Make a TensorRT-LLM workload snapshot-ready.
- Use Your Own Engine or Image - Use an engine other than vLLM, SGLang, or TensorRT-LLM.
- Checkpoint a Pod - Save a running replica's state as a
PodSnapshot. - Restore a Pod - Start a new replica from a snapshot.
- Configure Storage - The shared snapshot volume and how to configure it.
- Tune Storage - Storage option tuning and examples.
- Upgrade Snapshot and the CRDs - Upgrade with
helm upgrade; the operator updates the CRDs. - Liqo metadata plane — Consumer metadata, provider-local runtime and Snapshot-owned provenance on stock Liqo.
- API Reference -
PodSnapshot,PodSnapshotContent,SnapshotJob, and therestore-fromannotation. - Helm Chart Values - The Helm values that configure Snapshot, with their defaults.
snapshotctlCLI - Lower-level checkpoint/restore from a pod manifest.- Workload Contract - What a workload must do, and how its pod must be shaped, to be checkpointed and restored; a custom image is one packaging method.
- Restore Pod Contract - The pod interface for programmatic restore: annotations, control volume, startup gate, and seccomp.
- Reading a Failure Message - Find the step a checkpoint or restore failed at.
- Troubleshoot Checkpoints - A checkpoint does not finish.
- Troubleshoot Restores - A restore is slow or fails.
- Troubleshoot Slow Restores - Check timings, NFS clients, mount options, and fio results.
- Troubleshoot the Agent - The node agent or a component it runs fails.
- Building from Source - Build the images and install locally.
- Reproducing the Benchmarks - Run the cold-start, checkpoint, and restore measurements.
- Writing Documentation - Where a change goes and how to write it.
- Naming Conventions - Terms and naming patterns for the API, chart, and docs.
- Snapshot Enhancement Proposal (SNEP) - The design process for significant Snapshot changes.