Import prometheus client to collect metrics (#64)

Signed-off-by: Kaiyuan Hu <[email protected]>

main.py

@@ -1,24 +1,35 @@
-import argparse
 import os
+import argparse
+import uuid
 
 import uvicorn
 from fastapi import FastAPI, UploadFile
 from fastapi.encoders import jsonable_encoder
+from fastapi.responses import PlainTextResponse
 
 from config import TEMP_DIR
 
-
 os.makedirs(TEMP_DIR, exist_ok=True)
+os.environ['PROMETHEUS_DISABLE_CREATED_SERIES'] = 'true'
 
 # Specify mode
 parser = argparse.ArgumentParser(description='Start service with different modes.')
 parser.add_argument('--langchain', action='store_true')
 parser.add_argument('--towhee', action='store_true')
-
+parser.add_argument('--moniter', action='store_true')
+parser.add_argument('--max_observation', default=1000)
+parser.add_argument('--name', default=str(uuid.uuid4()))
 args = parser.parse_args()
 
+app = FastAPI()
+origins = ['*']
+
+# Apply args
 USE_LANGCHAIN = args.langchain
 USE_TOWHEE = args.towhee
+MAX_OBSERVATION = args.max_observation
+ENABLE_MONITER = args.moniter
+NAME = args.name
 
 assert (USE_LANGCHAIN and not USE_TOWHEE ) or (USE_TOWHEE and not USE_LANGCHAIN), \
     'The service should start with either "--langchain" or "--towhee".'
@@ -27,19 +38,29 @@
     from src_langchain.operations import chat, insert, drop, check, get_history, clear_history, count  # pylint: disable=C0413
 if USE_TOWHEE:
     from src_towhee.operations import chat, insert, drop, check, get_history, clear_history, count  # pylint: disable=C0413
+if ENABLE_MONITER:
+    from moniter import enable_moniter  # pylint: disable=C0413
+    from prometheus_client import generate_latest, REGISTRY  # pylint: disable=C0413
+
+    enable_moniter(app, MAX_OBSERVATION, NAME)
+
+    @app.get('/metrics')
+    async def metrics():
+        registry = REGISTRY
+        data = generate_latest(registry)
+        return PlainTextResponse(content=data, media_type='text/plain')
 
-app = FastAPI()
-origins = ['*']
 
 @app.get('/')
 def check_api():
-    return jsonable_encoder({'status': True, 'msg': 'ok'}), 200
+    res = jsonable_encoder({'status': True, 'msg': 'ok'}), 200
+    return res
+
 
 @app.get('/answer')
 def do_answer_api(session_id: str, project: str, question: str):
     try:
-        new_question, final_answer = chat(session_id=session_id,
-                            project=project, question=question)
+        new_question, final_answer = chat(session_id=session_id, project=project, question=question)
         assert isinstance(final_answer, str)
         return jsonable_encoder({
             'status': True,

moniter.py

@@ -0,0 +1,77 @@
+import os
+import time
+
+from starlette.middleware.base import BaseHTTPMiddleware
+from prometheus_client import Counter, Histogram, Summary
+
+# Define metrics
+latencies = {}
+latencies['all'] = []
+
+requests_total = Counter('requests_total', 'Cumulative requests', ['name'])
+requests_success_total = Counter('requests_success_total', 'Cumulative successful requests', ['name'])
+requests_failed_total = Counter('requests_failed_total', 'Cumulative failed requests', ['name'])
+
+endpoint_requests_total = Counter('endpoint_requests_total', 'Cumulative requests of each endpoint', ['name', 'endpoint'])
+endpoint_requests_success_total = Counter('endpoint_requests_success_total', 'Cumulative successful requests of each endpoint', ['name', 'endpoint'])
+endpoint_requests_failed_total = Counter('endpoint_requests_failed_total', 'Cumulative failed requests of each endpoint', ['name', 'endpoint'])
+
+latency_seconds_histogram = Histogram('latency_seconds_histogram', 'Request process latency histogram', ['name'])
+endpoint_latency_seconds_histogram = Histogram(
+    'endpoint_latency_seconds_histogram', 'Request process latency histogram of each endpoint', ['name', 'endpoint']
+)
+
+latency_seconds_summary = Summary('latency_seconds_summary', 'Request process latency summary', ['name'])
+endpoint_latency_seconds_summary = Summary(
+    'endpoint_latency_seconds_summary', 'Request process latency summary of each endpoint', ['name', 'endpoint']
+)
+
+
+def enable_moniter(app, max_observation, name):
+
+    # Define middleware to collect metrics
+    class RequestMetricsMiddleware(BaseHTTPMiddleware):
+        """
+        Middleware to process requests.
+        """
+        async def dispatch(self, request, call_next):
+            path = request.scope.get('path')
+            is_req = path != '/metrics'
+
+            if not is_req:
+                try:
+                    response = await call_next(request)
+                    return response
+                except Exception as e:
+                    raise e
+
+            begin = time.time()
+            requests_total.labels(name).inc()
+            endpoint_requests_total.labels(name, path).inc()
+            try:
+                response = await call_next(request)
+                if response.status_code / 100 < 4:
+                    requests_success_total.labels(name).inc()
+                    endpoint_requests_success_total.labels(name, path).inc()
+                    end = time.time()
+                    if path in latencies:
+                        latencies[path].append(end - begin)
+                        latencies[path] = latencies[path][-max_observation:]
+                    else:
+                        latencies[path] = [end - begin]
+                    latencies['all'].append(end - begin)
+                    latencies['all'] = latencies['all'][-max_observation:]
+                    latency_seconds_histogram.labels(name).observe(end - begin)
+                    endpoint_latency_seconds_histogram.labels(name, path).observe(end - begin)
+                    latency_seconds_summary.labels(name).observe(end - begin)
+                    endpoint_latency_seconds_summary.labels(name, path).observe(end - begin)
+                    return response
+                else:
+                    requests_failed_total.labels(name).inc()
+                    endpoint_requests_failed_total.labels(name, path).inc()
+            except Exception as e:
+                requests_failed_total.labels(name).inc()
+                endpoint_requests_failed_total.labels(name, path).inc()
+                raise e
+
+    app.add_middleware(RequestMetricsMiddleware)

requirements.txt

@@ -11,3 +11,4 @@ uvicorn
 towhee>=1.1.0
 pymilvus
 elasticsearch>=8.0.0
+prometheus-client