Model Serving

Overview

Serving turns a saved model into a reliable, low-latency API. The concerns shift from accuracy to latency, throughput, robustness, and observability. The model artifact and its preprocessing must travel together (use the pipeline from sklearn-pipelines).

When to use

• A validated model needs to be callable by other systems.
• Moving from notebook to production.

Minimal FastAPI service

from fastapi import FastAPI
from pydantic import BaseModel
import joblib, numpy as np

app = FastAPI()
model = joblib.load("model.joblib")  # full pipeline: preprocessing + estimator

class Features(BaseModel):
    age: float
    income: float
    country: str
    plan: str

@app.get("/health")
def health():
    return {"status": "ok"}

@app.post("/predict")
def predict(f: Features):
    import pandas as pd
    X = pd.DataFrame([f.model_dump()])
    proba = float(model.predict_proba(X)[0, 1])
    return {"probability": proba, "label": int(proba >= 0.5)}

Run: uvicorn app:app --host 0.0.0.0 --port 8000 --workers 4.

Production checklist

• Load the model once at startup, not per request.
• Validate inputs with Pydantic; return 422 on bad payloads.
• Health/readiness endpoints for orchestrators (k8s).
• Batch requests where possible to raise throughput.
• Timeouts + graceful degradation for downstream calls.
• Version the model in the response (model_version) for traceability.
• Pin the artifact's training env — preprocessing must match training exactly.

Speed: ONNX + quantization

# Export sklearn/torch model to ONNX, then serve with onnxruntime
import onnxruntime as ort
sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
out = sess.run(None, {"input": X.astype(np.float32)})

ONNX runtime + dynamic quantization often gives 2-4x CPU speedups. For LLMs, use vLLM/TGI rather than rolling your own.

Monitoring (don't deploy blind)

• Operational: latency p50/p95/p99, error rate, throughput.
• ML-specific: input feature drift, prediction distribution shift, and (when labels arrive) live metric decay.
• Alert on drift — a silently degrading model is worse than a down one.

Pitfalls

• Training/serving skew — different preprocessing in prod than training. Ship the pipeline, not just the estimator.
• Loading the model per request — kills latency.
• No input validation — one bad payload takes down the worker.
• `pickle`/`joblib` from untrusted sources — only load artifacts you produced.
• No monitoring — you won't notice drift until users do.

Hand-off

A containerized, monitored endpoint serving the exact pipeline that model-evaluation validated and experiment-tracking registered.