feat: add /health endpoint for microservice monitoring
Checks database, background workers (ingestion, collector, attacker, sniffer), and Docker daemon. Reports healthy/degraded/unhealthy status with per-component details. Returns 503 when required services fail, 200 for healthy or degraded (only optional services down).
This commit is contained in:
@@ -14,6 +14,11 @@ from .stream.api_stream_events import router as stream_router
|
||||
from .attackers.api_get_attackers import router as attackers_router
|
||||
from .attackers.api_get_attacker_detail import router as attacker_detail_router
|
||||
from .attackers.api_get_attacker_commands import router as attacker_commands_router
|
||||
from .config.api_get_config import router as config_get_router
|
||||
from .config.api_update_config import router as config_update_router
|
||||
from .config.api_manage_users import router as config_users_router
|
||||
from .config.api_reinit import router as config_reinit_router
|
||||
from .health.api_get_health import router as health_router
|
||||
|
||||
api_router = APIRouter()
|
||||
|
||||
@@ -42,3 +47,10 @@ api_router.include_router(attacker_commands_router)
|
||||
# Observability
|
||||
api_router.include_router(stats_router)
|
||||
api_router.include_router(stream_router)
|
||||
api_router.include_router(health_router)
|
||||
|
||||
# Configuration
|
||||
api_router.include_router(config_get_router)
|
||||
api_router.include_router(config_update_router)
|
||||
api_router.include_router(config_users_router)
|
||||
api_router.include_router(config_reinit_router)
|
||||
|
||||
0
decnet/web/router/health/__init__.py
Normal file
0
decnet/web/router/health/__init__.py
Normal file
80
decnet/web/router/health/api_get_health.py
Normal file
80
decnet/web/router/health/api_get_health.py
Normal file
@@ -0,0 +1,80 @@
|
||||
from typing import Any
|
||||
|
||||
from fastapi import APIRouter, Depends
|
||||
from fastapi.responses import JSONResponse
|
||||
|
||||
from decnet.web.dependencies import require_viewer, repo
|
||||
from decnet.web.db.models import HealthResponse, ComponentHealth
|
||||
|
||||
router = APIRouter()
|
||||
|
||||
_OPTIONAL_SERVICES = {"sniffer_worker"}
|
||||
|
||||
|
||||
@router.get(
|
||||
"/health",
|
||||
response_model=HealthResponse,
|
||||
tags=["Observability"],
|
||||
responses={
|
||||
401: {"description": "Could not validate credentials"},
|
||||
503: {"model": HealthResponse, "description": "System unhealthy"},
|
||||
},
|
||||
)
|
||||
async def get_health(user: dict = Depends(require_viewer)) -> Any:
|
||||
components: dict[str, ComponentHealth] = {}
|
||||
|
||||
# 1. Database
|
||||
try:
|
||||
await repo.get_total_logs()
|
||||
components["database"] = ComponentHealth(status="ok")
|
||||
except Exception as exc:
|
||||
components["database"] = ComponentHealth(status="failing", detail=str(exc))
|
||||
|
||||
# 2. Background workers
|
||||
from decnet.web.api import get_background_tasks
|
||||
for name, task in get_background_tasks().items():
|
||||
if task is None:
|
||||
components[name] = ComponentHealth(status="failing", detail="not started")
|
||||
elif task.done():
|
||||
if task.cancelled():
|
||||
detail = "cancelled"
|
||||
else:
|
||||
exc = task.exception()
|
||||
detail = f"exited: {exc}" if exc else "exited unexpectedly"
|
||||
components[name] = ComponentHealth(status="failing", detail=detail)
|
||||
else:
|
||||
components[name] = ComponentHealth(status="ok")
|
||||
|
||||
# 3. Docker daemon
|
||||
try:
|
||||
import docker
|
||||
|
||||
client = docker.from_env()
|
||||
client.ping()
|
||||
client.close()
|
||||
components["docker"] = ComponentHealth(status="ok")
|
||||
except Exception as exc:
|
||||
components["docker"] = ComponentHealth(status="failing", detail=str(exc))
|
||||
|
||||
# Compute overall status
|
||||
required_failing = any(
|
||||
c.status == "failing"
|
||||
for name, c in components.items()
|
||||
if name not in _OPTIONAL_SERVICES
|
||||
)
|
||||
optional_failing = any(
|
||||
c.status == "failing"
|
||||
for name, c in components.items()
|
||||
if name in _OPTIONAL_SERVICES
|
||||
)
|
||||
|
||||
if required_failing:
|
||||
overall = "unhealthy"
|
||||
elif optional_failing:
|
||||
overall = "degraded"
|
||||
else:
|
||||
overall = "healthy"
|
||||
|
||||
result = HealthResponse(status=overall, components=components)
|
||||
status_code = 503 if overall == "unhealthy" else 200
|
||||
return JSONResponse(content=result.model_dump(), status_code=status_code)
|
||||
Reference in New Issue
Block a user