Upload folder using huggingface_hub
Browse files- api/main.py +18 -3
api/main.py
CHANGED
|
@@ -13,7 +13,7 @@ import time
|
|
| 13 |
from contextlib import asynccontextmanager
|
| 14 |
from pathlib import Path
|
| 15 |
|
| 16 |
-
from fastapi import FastAPI, HTTPException, Request, status
|
| 17 |
from fastapi.responses import JSONResponse
|
| 18 |
|
| 19 |
from api import db, settings
|
|
@@ -144,7 +144,11 @@ async def model_info(request: Request) -> ModelInfoResponse:
|
|
| 144 |
|
| 145 |
|
| 146 |
@app.post("/predict", response_model=PredictionResponse, tags=["scoring"])
|
| 147 |
-
async def predict(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 148 |
raw_inputs = payload.model_dump()
|
| 149 |
sk_id = raw_inputs.pop("SK_ID_CURR")
|
| 150 |
|
|
@@ -208,7 +212,7 @@ async def predict(payload: PredictionRequest, request: Request) -> PredictionRes
|
|
| 208 |
plumbing_ms = max(0, latency_ms - feature_assembly_ms - inference_ms)
|
| 209 |
else:
|
| 210 |
plumbing_ms = None
|
| 211 |
-
|
| 212 |
sk_id_curr=sk_id,
|
| 213 |
client_known=client_known,
|
| 214 |
raw_input=raw_inputs,
|
|
@@ -225,3 +229,14 @@ async def predict(payload: PredictionRequest, request: Request) -> PredictionRes
|
|
| 225 |
status_code=status_code,
|
| 226 |
error_message=error_message,
|
| 227 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
from contextlib import asynccontextmanager
|
| 14 |
from pathlib import Path
|
| 15 |
|
| 16 |
+
from fastapi import BackgroundTasks, FastAPI, HTTPException, Request, status
|
| 17 |
from fastapi.responses import JSONResponse
|
| 18 |
|
| 19 |
from api import db, settings
|
|
|
|
| 144 |
|
| 145 |
|
| 146 |
@app.post("/predict", response_model=PredictionResponse, tags=["scoring"])
|
| 147 |
+
async def predict(
|
| 148 |
+
payload: PredictionRequest,
|
| 149 |
+
request: Request,
|
| 150 |
+
background_tasks: BackgroundTasks,
|
| 151 |
+
) -> PredictionResponse:
|
| 152 |
raw_inputs = payload.model_dump()
|
| 153 |
sk_id = raw_inputs.pop("SK_ID_CURR")
|
| 154 |
|
|
|
|
| 212 |
plumbing_ms = max(0, latency_ms - feature_assembly_ms - inference_ms)
|
| 213 |
else:
|
| 214 |
plumbing_ms = None
|
| 215 |
+
log_kwargs: dict[str, object | None] = dict(
|
| 216 |
sk_id_curr=sk_id,
|
| 217 |
client_known=client_known,
|
| 218 |
raw_input=raw_inputs,
|
|
|
|
| 229 |
status_code=status_code,
|
| 230 |
error_message=error_message,
|
| 231 |
)
|
| 232 |
+
# Success: defer the Supabase round-trip until AFTER the response
|
| 233 |
+
# is sent, so the client never waits on ~350 ms of DB log overhead.
|
| 234 |
+
# Failure: log synchronously — BackgroundTasks scheduled on a route
|
| 235 |
+
# are attached to its Response, and the exception-handler chain
|
| 236 |
+
# builds its own Response, silently dropping pending tasks. Losing
|
| 237 |
+
# observability on errors would be a worse trade-off than the
|
| 238 |
+
# one-shot latency hit paid by failing requests.
|
| 239 |
+
if status_code < 400:
|
| 240 |
+
background_tasks.add_task(log_prediction, **log_kwargs)
|
| 241 |
+
else:
|
| 242 |
+
log_prediction(**log_kwargs)
|