feat: implement full-platform STT API (v2.3 consensus plan)

Batch+realtime transcription API: faster-whisper engine w/ EngineOwner
(single GPU owner, OOM downgrade chain, persisted attempted_profiles),
hardware-adaptive device manager (T0-T3 VRAM tiers), Redis/in-proc job
queue w/ leases + crash recovery, postprocess (glossary/rules/LLM w/
egress guard), privacy-first result store (UUID keys, source deleted
after transcribe), retention sweeper, API-key auth (HMAC digests,
scopes, job ownership), WebSocket realtime lane (LocalAgreement),
CLI (detect/transcribe/bench/serve/key), Docker, benchmark runner.

127 mock-based tests pass; ruff clean. Includes verification checklist
and autoplan review notes.
This commit is contained in:
2026-08-12 16:01:21 +09:00
parent b7c30f8b71
commit 7327145d7a
82 changed files with 7351 additions and 0 deletions
View File
+210
View File
@@ -0,0 +1,210 @@
"""API 통합 테스트 — TestClient 기반 인증/업로드/소유권/큐 흐름.
실제 모델·Redis 없이 in-proc 브로커 + mock 결과로 동작한다.
"""
from __future__ import annotations
import json
import pytest
from fastapi.testclient import TestClient
from luke_scribe.api.app import create_app
from luke_scribe.config import Settings
@pytest.fixture
def settings(tmp_path) -> Settings:
return Settings(
_env_file=None,
results_dir=str(tmp_path / "results"),
api_key_file=str(tmp_path / "api_keys.json"),
api_keys="key-transcribe,key-admin:admin,transcribe",
queue_backend="inproc",
model_cache_dir=None,
tunnel="none",
max_queue=10,
)
@pytest.fixture
def client(settings: Settings) -> TestClient:
app = create_app(settings)
with TestClient(app) as c:
yield c
class TestHealth:
def test_health_public(self, client: TestClient):
r = client.get("/health")
assert r.status_code == 200
body = r.json()
assert body["status"] == "ok"
assert body["model_ready"] is False # 모델 캐시 없음 → 프로비저닝 보류
class TestAuth:
def test_missing_key_rejected(self, client: TestClient):
r = client.get("/v1/jobs")
assert r.status_code == 401
def test_bad_key_rejected(self, client: TestClient):
r = client.get("/v1/jobs", headers={"X-API-Key": "wrong-key"})
assert r.status_code == 401
def test_admin_scope_enforced(self, client: TestClient):
r = client.get("/v1/system", headers={"X-API-Key": "key-transcribe"})
assert r.status_code == 403
r2 = client.get("/v1/system", headers={"X-API-Key": "key-admin"})
assert r2.status_code == 200
body = r2.json()
assert body["capability_tier"] in ("T0", "T1", "T2", "T3")
assert body["queue_depth"] == 0
class TestJobs:
def test_create_get_cancel_flow(self, client: TestClient, settings: Settings):
headers = {"X-API-Key": "key-transcribe"}
r = client.post(
"/v1/jobs",
files={"file": ("meeting.mp3", b"fake-audio-bytes", "audio/mpeg")},
data={"options": json.dumps({"language": "ko", "model": "large-v3-turbo"})},
headers=headers,
)
assert r.status_code == 202, r.text
body = r.json()
job_id = body["job_id"]
assert body["status"] == "queued"
assert body["queue_position"] == 0
# 조회
r = client.get(f"/v1/jobs/{job_id}", headers=headers)
assert r.status_code == 200
assert r.json()["status"] == "queued"
# 결과는 아직 없음
r = client.get(f"/v1/jobs/{job_id}/result", headers=headers)
assert r.status_code == 409
# 취소
r = client.delete(f"/v1/jobs/{job_id}", headers=headers)
assert r.status_code == 200
assert r.json()["status"] == "cancelled"
def test_ownership_enforced(self, client: TestClient):
"""Eng P14: 다른 키가 만든 job 조회/취소 불가."""
a = client.post(
"/v1/jobs",
files={"file": ("a.mp3", b"x", "audio/mpeg")},
data={"options": "{}"},
headers={"X-API-Key": "key-transcribe"},
)
job_id = a.json()["job_id"]
r = client.get(f"/v1/jobs/{job_id}", headers={"X-API-Key": "key-admin"})
assert r.status_code == 403
r = client.delete(f"/v1/jobs/{job_id}", headers={"X-API-Key": "key-admin"})
assert r.status_code == 403
def test_unknown_job_404(self, client: TestClient):
r = client.get(
"/v1/jobs/00000000-0000-0000-0000-000000000000",
headers={"X-API-Key": "key-transcribe"},
)
assert r.status_code == 404
def test_bad_options_422(self, client: TestClient):
r = client.post(
"/v1/jobs",
files={"file": ("a.mp3", b"x", "audio/mpeg")},
data={"options": "not json {"},
headers={"X-API-Key": "key-transcribe"},
)
assert r.status_code == 422
def test_bad_format_422(self, client: TestClient, settings: Settings):
headers = {"X-API-Key": "key-transcribe"}
r = client.post(
"/v1/jobs",
files={"file": ("a.mp3", b"x", "audio/mpeg")},
data={"options": "{}"},
headers=headers,
)
job_id = r.json()["job_id"]
r2 = client.get(f"/v1/jobs/{job_id}/result?format=docx", headers=headers)
assert r2.status_code == 422
def test_oversize_413(self, client: TestClient, settings: Settings):
settings.max_upload_bytes = 10
r = client.post(
"/v1/jobs",
files={"file": ("big.mp3", b"x" * 100, "audio/mpeg")},
data={"options": "{}"},
headers={"X-API-Key": "key-transcribe"},
)
assert r.status_code == 413
class TestQueueFull:
def test_queue_full_429(self, client: TestClient, settings: Settings):
settings.max_queue = 1
headers = {"X-API-Key": "key-transcribe"}
r1 = client.post(
"/v1/jobs",
files={"file": ("a.mp3", b"x", "audio/mpeg")},
data={"options": "{}"},
headers=headers,
)
assert r1.status_code == 202
r2 = client.post(
"/v1/jobs",
files={"file": ("b.mp3", b"x", "audio/mpeg")},
data={"options": "{}"},
headers=headers,
)
assert r2.status_code == 429
assert "Retry-After" in r2.headers
class TestResultEndpoint:
def test_result_after_processing(self, client: TestClient, settings: Settings, tmp_path):
"""워커가 결과를 저장한 뒤 result 엔드포인트가 JSON/SRT를 반환."""
from luke_scribe.results.models import TranscriptResult
from luke_scribe.results.store import ResultStore
headers = {"X-API-Key": "key-transcribe"}
r = client.post(
"/v1/jobs",
files={"file": ("a.mp3", b"x", "audio/mpeg")},
data={"options": "{}"},
headers=headers,
)
job_id = r.json()["job_id"]
# 워커 대신 결과를 직접 기록 (모델 없이 mock)
result = TranscriptResult(
status="completed",
source={"name": "a.mp3", "codec": "mp3", "size_bytes": 1},
normalized_audio={"duration_sec": 2.0},
execution={"model": "large-v3-turbo", "device": "cpu", "compute_type": "int8"},
text="테스트 전사 결과",
segments=[{"index": 0, "start": 0.0, "end": 1.0, "text": "테스트 전사 결과"}],
)
from luke_scribe.jobqueue.jobs import JobStatus
store = ResultStore(settings.results_dir)
store.write_result(job_id, result)
# 상태를 completed로 (워커가 했을 일) — queued→processing→completed
job = client.app.state.broker.get(job_id)
job.transition(JobStatus.PROCESSING)
job.transition(JobStatus.COMPLETED)
client.app.state.broker.save_meta(job)
r_json = client.get(f"/v1/jobs/{job_id}/result?format=json", headers=headers)
assert r_json.status_code == 200
assert json.loads(r_json.json()["content"])["text"] == "테스트 전사 결과"
r_srt = client.get(f"/v1/jobs/{job_id}/result?format=srt", headers=headers)
assert r_srt.status_code == 200
assert "WEBVTT" not in r_srt.json()["content"] # srt 포맷
assert "--> " in r_srt.json()["content"]
+267
View File
@@ -0,0 +1,267 @@
"""Worker 통합 테스트 — 큐 → 클레임 → 전사 → 결과 저장 (전부 mock)."""
from __future__ import annotations
import time
from luke_scribe.jobqueue.broker import InProcBroker
from luke_scribe.jobqueue.jobs import Job, JobStatus
from luke_scribe.jobqueue.worker import Worker, WorkerCallbacks
from luke_scribe.results.store import ResultStore
from ..conftest import FakeEngineOwner, FakeIngestor
def _job(**kw) -> Job:
defaults = dict(
type="file",
lane="batch",
options={
"model": "large-v3-turbo",
"language": "ko",
"device": "cpu",
"compute_type": "int8",
},
source_path="/tmp/fake-source.mp3",
source_name="fake-source.mp3",
)
defaults.update(kw)
return Job(**defaults)
class TestWorkerLifecycle:
def test_complete_flow(self, settings, tmp_path):
"""enqueue → worker 처리 → completed + 결과 저장 + 콜백."""
broker = InProcBroker(settings)
store = ResultStore(str(tmp_path / "results"))
owner = FakeEngineOwner()
done: list[str] = []
def on_done(job, result):
done.append(job.id)
worker = Worker(
settings=settings,
broker=broker,
store=store,
owner=owner,
ingestor=FakeIngestor(duration_sec=4.0),
callbacks=WorkerCallbacks(on_job_done=on_done),
)
job = _job()
broker.enqueue(job)
worker.drain()
assert broker.get(job.id).status == JobStatus.COMPLETED
assert done == [job.id]
result = store.read_result(job.id)
assert result is not None
assert result.status == "completed"
assert result.text # 세그먼트 텍스트 연결
assert result.execution is not None
assert result.execution.device == "cpu"
# 강등 없음
assert result.execution.downgrade_attempts == 0
def test_meta_written(self, settings, tmp_path):
"""완료 시 job 메타(완료 시각 포함)가 결과 디렉터리에 저장."""
broker = InProcBroker(settings)
store = ResultStore(str(tmp_path / "results"))
worker = Worker(
settings=settings,
broker=broker,
store=store,
owner=FakeEngineOwner(),
ingestor=FakeIngestor(),
)
job = _job()
broker.enqueue(job)
worker.drain()
import json
meta = json.loads(
tmp_path.joinpath("results", job.id, "meta.json").read_text(encoding="utf-8")
)
assert meta["status"] == "completed"
assert meta["completed_at"] is not None
def test_transcription_failure_marks_failed(self, settings, tmp_path):
"""엔진 오류 → failed + 오류 메타 저장."""
class BoomEngine:
def transcribe(self, req):
raise RuntimeError("engine exploded")
class BoomOwner:
def __init__(self):
self.engine = BoomEngine()
def transcribe(self, req):
return self.engine.transcribe(req)
def unload_all(self):
pass
broker = InProcBroker(settings)
store = ResultStore(str(tmp_path / "results"))
worker = Worker(
settings=settings,
broker=broker,
store=store,
owner=BoomOwner(),
ingestor=FakeIngestor(),
)
job = _job()
broker.enqueue(job)
worker.drain()
assert broker.get(job.id).status == JobStatus.FAILED
def test_cancel_before_claim(self, settings, tmp_path):
"""큐에서 취소된 job은 워커가 claim하지 않고 cancelled 처리."""
broker = InProcBroker(settings)
store = ResultStore(str(tmp_path / "results"))
worker = Worker(
settings=settings,
broker=broker,
store=store,
owner=FakeEngineOwner(),
ingestor=FakeIngestor(),
)
job = _job()
broker.enqueue(job)
broker.cancel(job.id)
worker.drain()
assert broker.get(job.id).status == JobStatus.CANCELLED
def test_cancel_during_processing(self, settings, tmp_path):
"""처리 중 취소 → 세그먼트 경계에서 CancelledError → cancelled (Eng P4)."""
class CancelAfterFirstOwner:
def __init__(self):
self.n = 0
self.segments = [
{"index": 0, "start": 0.0, "end": 1.0, "text": "첫 세그먼트"},
{"index": 1, "start": 1.0, "end": 2.0, "text": "두 번째"},
]
def transcribe(self, req):
# worker는 token 미전달, 취소는 세그먼트 경계에서 검사됨
return {
"segments": iter(
[
{"index": 0, "start": 0.0, "end": 1.0, "text": "첫 세그먼트"},
{"index": 1, "start": 1.0, "end": 2.0, "text": "두 번째"},
]
),
"info": {"language": "ko"},
"device": "cpu",
"compute_type": "int8",
"attempted_profiles": ["cpu/int8"],
}
# 세그먼트를 먼저 소비한 뒤 취소하도록: 첫 next 후 should_cancel True
broker = InProcBroker(settings)
store = ResultStore(str(tmp_path / "results"))
owner = CancelAfterFirstOwner()
worker = Worker(
settings=settings,
broker=broker,
store=store,
owner=owner,
ingestor=FakeIngestor(),
)
job = _job()
broker.enqueue(job)
# 처리 시작 전에 취소 요청을 미리 걸어 두면 claim 단계에서 cancelled 처리됨.
# 대신 처리 중 취소를 시뮬레이션: claim 직후 cancel 호출
claimed = broker.claim_next("w1", 60.0)
broker.cancel(claimed.id)
# 워커는 already-processing job을 직접 처리
worker._process(claimed)
assert broker.get(job.id).status == JobStatus.CANCELLED
class TestPrivacyFirst:
def test_source_deleted_after_completion(self, settings, tmp_path):
"""plan §3.7e/§6.1: 전사 완료 후 업로드 원본 오디오 즉시 삭제."""
from luke_scribe.results.store import ResultStore
settings.delete_source = True
store = ResultStore(str(tmp_path / "results"))
broker = InProcBroker(settings)
worker = Worker(
settings=settings,
broker=broker,
store=store,
owner=FakeEngineOwner(),
ingestor=FakeIngestor(),
)
job = _job()
# source가 store에 저장된 파일을 가리키도록
src = store.source_path_for(job.id, "meeting.mp3")
src.write_bytes(b"fake-audio")
job.source_path = str(src)
broker.enqueue(job)
worker.drain()
assert not src.exists() # 원본 삭제
assert store.read_result(job.id) is not None # 결과는 보존
def test_source_kept_when_delete_source_false(self, settings, tmp_path):
from luke_scribe.results.store import ResultStore
settings.delete_source = False
store = ResultStore(str(tmp_path / "results"))
broker = InProcBroker(settings)
worker = Worker(
settings=settings,
broker=broker,
store=store,
owner=FakeEngineOwner(),
ingestor=FakeIngestor(),
)
job = _job()
src = store.source_path_for(job.id, "meeting.mp3")
src.write_bytes(b"fake-audio")
job.source_path = str(src)
broker.enqueue(job)
worker.drain()
assert src.exists()
class TestWorkerProgress:
def test_progress_emitted(self, settings, tmp_path):
broker = InProcBroker(settings)
store = ResultStore(str(tmp_path / "results"))
worker = Worker(
settings=settings,
broker=broker,
store=store,
owner=FakeEngineOwner(segments=[{"index": 0, "start": 0.0, "end": 5.0, "text": "x"}]),
ingestor=FakeIngestor(duration_sec=10.0),
)
job = _job()
broker.enqueue(job)
worker.drain()
final = broker.get(job.id)
assert final.progress is not None
assert final.progress <= 1.0
assert final.processed_sec is not None
class TestCrashRecovery:
def test_stale_processing_recovered(self, settings, tmp_path):
"""Eng P11: 워커 크래시(리스 만료) → 스타트업 reconciler가 failed 처리."""
broker = InProcBroker(settings)
job = _job()
broker.enqueue(job)
claimed = broker.claim_next("w1", 60.0)
# 리스 만료 시뮬레이션
claimed.lease_expires_at = time.time() - 10
broker.save_meta(claimed)
worker = Worker(
settings=settings, broker=broker, store=ResultStore(str(tmp_path / "results"))
)
reclaimed = worker.start_reconcile()
assert job.id in reclaimed
assert broker.get(job.id).status == JobStatus.FAILED
assert broker.get(job.id).error_code == "worker_crash"