ML inference on spot instances. 70-85% cost reduction through orchestration.
This project is actively being restructured. The codebase is transitioning to a cleaner workspace layout. Updates coming soon.
What works now:
- Simulation engine for validating scheduling policies
- Core types and traits for spot orchestration
- Agent scaffolding for spot instance monitoring
In progress:
- CLI for fleet management
- Full agent implementation
- Documentation
Synkti orchestrates ML inference workloads on spot instances. Spot instances are 70-90% cheaper than on-demand, but can be terminated with only 2 minutes notice.
The key insight: you don't need complex checkpoint/restore mechanisms. GPU state can't be checkpointed anyway (CUDA contexts, VRAM). Instead, Synkti uses stateless failover - when a spot instance gets preempted, we drain in-flight requests and spawn a fresh container on a replacement instance.
Same cost savings, simpler architecture, actually works with GPUs.
crates/
├── synkti-core/ # Shared types and traits
├── synkti-agent/ # Runs on spot instances (monitoring, container lifecycle)
├── synkti-cli/ # Command-line interface
└── synkti-simulation/ # Policy testing and cost modeling
git clone https://github.com/Bobby-Math/synkti.git
cd synkti/crates
# Run the simulation engine
cargo run -p synkti-simulation -- --duration 48 --tasks 100
# Build everything
cargo build --workspaceAGPL-3.0
Bobby - github.com/Bobby-Math