agentleFS
saintgo7saintgo7/claude-skills/gem-llm-gateway-debug/SKILL.md

gem-llm-gateway-debug

Skill0 starsChanged 6 months agoMITView on GitHub

GEM-LLM Gateway (FastAPI port 8080) 디버깅. 사용 시점 — "Gateway 500", "QueuePool", "unknown_model", "401 Invalid API key", "스트리밍 끊김", "/v1/chat 안됨", "/metrics 추가", "rate limit". 인증/quota/스트리밍 프록시 구조 + 흔한 함정.

  • Reads credentials

The file

Files
SKILL.md

gem-llm-gateway-debug

구조

[Request: Bearer token]
  → middleware/auth.py        # API key SHA256 검증, AuthContext 주입
  → middleware/rate_limit.py  # slowapi (60 RPM)
  → middleware/logging.py     # 요청/사용량 로깅 (key prefix 8자만)
  → routes/chat.py            # /v1/chat/completions
  → services/quota.py         # daily/concurrent 검사
  → services/proxy.py         # httpx → vLLM upstream
  → services/usage.py         # async UsageLog INSERT

정상 endpoint

GET /healthz200 — 프로세스 살아있음
GET /readyz200 — 두 vLLM 모두 health OK
GET /metricsPrometheus format
GET /v1/models인증 필요. qwen2.5-coder-32b, qwen3-coder-30b 반환
POST /v1/chat/completions인증 + quota + 프록시
POST /admin/usersX-Admin-Key 헤더 필요
POST /admin/keys키 발급 (raw_key 1회 노출)
GET /admin/usage사용량 조회

시작 명령

cd /home/jovyan/gem-llm/src/gateway
set -a; source /home/jovyan/gem-llm/.env; set +a
nohup /home/jovyan/vllm-env/bin/python -m uvicorn gateway.main:app \
  --host 0.0.0.0 --port 8080 \
  > /home/jovyan/gem-llm/_logs/gateway.log 2>&1 &

.env는 src/gateway/.env에도 같은 사본 필요 (pydantic-settings가 cwd .env 찾음).

.env 필수 변수

GATEWAY_PORT=8080
GATEWAY_ADMIN_KEY=admin_<32hex>     # 자동생성, 한 번만 표시
GATEWAY_DB_URL=sqlite+aiosqlite:////home/jovyan/gem-llm/_data/gateway.db
GATEWAY_API_KEY_SALT=salt_<32hex>   # 자동생성, 변경 시 기존 모든 키 무효
GATEWAY_VLLM_31B_URL=http://localhost:8001
GATEWAY_VLLM_26B_URL=http://localhost:8002
GATEWAY_DEFAULT_DAILY_TOKEN_LIMIT=50000
GATEWAY_DEFAULT_RPM=60

모델 라우팅 매핑

src/gateway/gateway/config.py:

@property
def upstream_map(self) -> dict[str, str]:
    return {
        "qwen2.5-coder-32b": self.vllm_31b_url,  # → 8001
        "qwen3-coder-30b": self.vllm_26b_url,    # → 8002
    }

새 모델 추가 시 여기 + routes/models.py의 served_model_names 둘 다 업데이트.

흔한 에러 → 원인

응답원인
401 invalid_api_key키 prefix 또는 hash 불일치, salt 변경, 회수됨
400 unknown_model: gem-31bupstream_map 옛 이름 — case 11 마이그레이션 후 수정 필요
400 admin_key requiredX-Admin-Key 헤더 누락 또는 GATEWAY_ADMIN_KEY 불일치
429 rate_limited 60 per 1 minuteslowapi 정상 작동
429 daily_token_limitquota.py — 사용자별 50K 한도
500 internal_error내부 — 로그 traceback 확인
502 bad_gatewayupstream vLLM 응답 안 함 — gem-llm-vllm-debug

500 traceback 패턴

grep -B2 -A5 "ERROR\|Exception\|Traceback" /home/jovyan/gem-llm/_logs/gateway.log | tail -30

QueuePool limit of size 5 overflow 10 reached

케이스 12. db.py에서 pool_size=50, max_overflow=150, pool_timeout=10 설정.

sqlite3.OperationalError: disk I/O error

케이스 13. SQLite WAL 파일 손상. _data/를 _trash/로 격리 + alembic upgrade head + 사용자 재발급.

unknown_model

upstream_map 또는 routes/models.py의 served_model_names 옛 이름. grep + sed로 일괄 치환.

Invalid args for response field

chat.py의 @router.post("/chat/completions")에 response_model=None 추가 (이미 적용됨).

인증 흐름

  1. Bearer 헤더 → gem_live_<32hex> 형식 검증
  2. prefix 8자 추출 → DB의 api_keys.key_prefix lookup
  3. 후보 hash 비교 (SHA256 + salt)
  4. 매치 → AuthContext 주입 (user_id, plan)
  5. revoked=True 또는 미매치 → 401

quota 검사 (services/quota.py)

1. RPM (slowapi 미들웨어, in-memory) — 60/min/key
2. concurrent — asyncio.Semaphore(quotas.concurrent_limit), default 5
3. daily — UsageLog 테이블에서 오늘자 토큰 SUM ≥ daily_token_limit

증액 (개별 사용자):

UPDATE quotas SET daily_token_limit=200000, rpm_limit=120, concurrent_limit=10
WHERE user_id='01K...';

스트리밍 동작

1. client → POST /v1/chat/completions {"stream": true}
2. proxy.py → httpx.AsyncClient.stream()으로 vLLM 호출
3. SSE chunks (data: {...}\n\n) 그대로 패스스루
4. data: [DONE] 도착 시 종료
5. 스트림 끝나면 services/usage.py.record() async (블로킹 X)

/metrics (Prometheus)

# routes/health.py
from prometheus_client import generate_latest, CONTENT_TYPE_LATEST
@router.get("/metrics")
async def metrics():
    return Response(generate_latest(), media_type=CONTENT_TYPE_LATEST)

기본 metrics: process_cpu, process_memory, python_gc, http_requests_total. 커스텀 metric 추가 시 Counter/Histogram import.

부하 테스트 시 Gateway 동작

50동접에서:

  • p50 = 144ms (DB lookup ~10ms + vLLM ~120ms)
  • p99 = 4.1s (vLLM 큐 적체로 인한 outliers)
  • 처리량 = 45 req/s, 1441 tok/s

병목이 의심되면:

  1. SQLAlchemy pool size (case 12)
  2. asyncio.Semaphore concurrent (default 5)
  3. vLLM 자체 큐
  4. nvidia-smi GPU util

관련

  • 책 Part III Ch.9 — Gateway 구현 상세
  • SPEC-03 — Gateway API 설계
  • src/gateway/ 코드 + tests/integration/test_e2e_gateway_vllm.py

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.