ML Infer gRPC Service

Runtime ML gating calls a Python gRPC inference service.

Public interface:

• @tradejs/node
• gRPC Predict contract (PredictRequest / PredictResponse)

Start / Stop

Start and stop the inference service with your deployment tooling (Docker/Kubernetes/systemd/etc).

Default endpoint:

• 127.0.0.1:50051

Set runtime address:

ML_GRPC_ADDRESS=127.0.0.1:50051

RPC Contract

PredictRequest:

• strategy (string)
• features (map<string,double>)
• threshold (double)

PredictResponse:

• probability
• threshold
• passed

Model Loading

Service resolves model files from MODEL_DIR:

• ensemble aliases: <Strategy>.modelN.joblib
• single alias: <Strategy>.joblib

If ensemble files exist, prediction is mean probability across members.

Typical local setup:

• trained models are in ./data/ml/models
• container uses MODEL_DIR=/app/data/ml/models
• host ./data is mounted to /app/data in infer container

So aliases written by training (./data/ml/models/<Strategy>.joblib) are immediately visible to inference.

Runtime Integration

On runtime side, ML gating:

1. Builds ML payload from signal.
2. Applies same window trim policy as train (trimMlTrainingRowWindows(..., 5)).
3. Removes non-feature columns.
4. Calls gRPC Predict.

If service is unavailable, runtime logs error and returns null for ML decision.

Health Check

npx @tradejs/cli doctor --require-ml