SpotWarp: Local failover daemon for Vast.ai spot GPUs finds 5 real bugs

The core problem SpotWarp addresses is the hidden cost of spot GPU eviction: not the minutes of downtime, but the hours of training progress lost when an instance terminates without notice. Standard spot workflows require manual recovery—logging into a console, finding a new GPU, re-uploading code, and restarting training from epoch zero—making spot pricing a risky choice for long training runs despite the 70% savings. SpotWarp converts that gamble into a safe operation by inverting the protection strategy: instead of hoping the instance survives, it assumes eviction will happen and continuously protects the workspace before it does.

The daemon's architecture centers on local execution and zero trust. All orchestration runs on the user's PC—the rented container is treated as ephemeral and untrusted, receiving only workspace data and resume commands over SSH. This design eliminates the security risk of placing API keys on rented hosts, a vulnerability that normally forces users to choose between spot savings and credential safety. The parallel candidate racing feature (renting up to 4 potential replacements concurrently and picking the cheapest reachable one) shrinks failover time to under a minute, while automatic failback watches the original cloud in the background and migrates workloads back the instant cheaper capacity returns—verified end-to-end in a real scenario: Vast.ai eviction → RunPod bridge → automatic return to Vast.ai. This removes the penalty of accidentally leaving an expensive bridge instance running indefinitely.