This project is deprecated. Please check out https://github.com/ucbrise/clipper for more recent work in model serving.
Velox is a system for serving machine learning predictions.
- Supports real-time personalized predictions
- Integration with Spark and KeystoneML
- Automatic model training in batch and online
Installing Velox using the provided scripts requires fabric >= 1.10.0 and dependencies (Paramiko >= 1.10.0)
Install and start Velox:
git clone https://github.com/amplab/velox-modelserver.git
cd velox-modelserver/bin/cluster
pip install fabric
fab install_velox_local:~/
fab start_velox:start_local=y
This start a local Velox instance listening on localhost at port 8080.
You can interact with the Velox REST API using cURL.
curl -H "Content-Type: application/json" -d '{"context": 4, "uid":1000}' http://localhost:8080/predict/matrixfact
curl -H "Content-Type: application/json" -d '{"context": 4, "uid":4, "score":1.3}' http://localhost:8080/observe/matrixfact
curl http://localhost:8080/retrain/matrixfact
For more details and a guide to deploying Velox on a cluster, check out our deployment guide.
- Mailing list: velox-modelserver@googlegroups.com
- crankshaw@cs.berkeley.edu
- CIDR paper
- Tech talk
- Video and slides from presentation at AMPCamp 5
Velox is under the Apache 2.0 License.
More documentation coming soon