app.py

from fastapi import FastAPI, HTTPException, Header
from pydantic import BaseModel
import pandas as pd, os, joblib, time, requests
from huggingface_hub import hf_hub_download
from prometheus_client import Counter, Histogram, Gauge, generate_latest

app = FastAPI()

API_KEY = os.getenv("API_KEY")
HF_MODEL_REPO = os.getenv("HF_MODEL_REPO")
PROM_PUSHGATEWAY = os.getenv("PROM_PUSHGATEWAY")

# Prometheus metrics
REQS = Counter("pred_requests_total", "Total prediction requests")
LAT = Histogram("pred_request_latency_seconds", "Request latency")
LATEST = Gauge("latest_prediction", "Last predicted value")

# Load model and preprocessing artifacts
try:
    m = hf_hub_download(repo_id=HF_MODEL_REPO, filename="best_model.joblib")
    e = hf_hub_download(repo_id=HF_MODEL_REPO, filename="models/encoders.joblib")
    s = hf_hub_download(repo_id=HF_MODEL_REPO, filename="models/scaler.joblib")
    f = hf_hub_download(repo_id=HF_MODEL_REPO, filename="models/feature_columns.joblib")

    model = joblib.load(m)
    encoders = joblib.load(e)
    scaler = joblib.load(s)
    feature_columns = joblib.load(f)
    loaded = True
except Exception as ex:
    print("Model load error:", ex)
    loaded = False


@app.get("/")
def health():
    return {"status": "ok", "model_loaded": loaded, "features": feature_columns}


@app.post("/predict")
def predict(payload: dict, x_api_key: str = Header(None)):
    if x_api_key != API_KEY:
        raise HTTPException(status_code=401, detail="Invalid API key")

    if not loaded:
        raise HTTPException(status_code=503, detail="Model not loaded")

    # Convert input dict to DataFrame and reindex to match training columns
    df = pd.DataFrame([payload]).reindex(columns=feature_columns, fill_value=0)

    # Scale numeric values
    df_scaled = scaler.transform(df)
    start = time.time()
    pred = model.predict(df_scaled)[0]
    LAT.observe(time.time() - start)
    REQS.inc()
    LATEST.set(pred)

    if PROM_PUSHGATEWAY:
        try:
            requests.post(f"{PROM_PUSHGATEWAY}/metrics/job/loan_model", data=generate_latest())
        except Exception:
            pass

    return {"prediction": int(pred), "used_features": feature_columns}


@app.get("/metrics")
def metrics():
    return generate_latest()