fix(P0): gpu-encoder 冷启动防护 — pre-warm + 首字节短超时兜底 #2072
@@ -7,13 +7,21 @@
|
||||
3. 生成 relay 一次性 key,构造两个带 token 的 URL:
|
||||
- put_url:给 P4000 回传结果,走 relay_base_url(Tailscale host:8092)
|
||||
- get/del_url:worker 自己下载+清理用,走 relay_internal_base_url(Docker DNS 直连 API)
|
||||
4. POST P4000 /api/render/sync:inputs={"in.mp4": "<mezzanine-get-url>"}, output_url="<put_url>"
|
||||
4. 【冷启动防护】距上次成功通信 >60s 时,先 GET /health 预热 Tailscale 链路(短超时快速失败)
|
||||
5. POST P4000 /api/render/sync:inputs={"in.mp4": "<mezzanine-get-url>"}, output_url="<put_url>"
|
||||
ffmpeg_args: -i in.mp4 [-vf <vf>] -c:v h264_nvenc ... -an/-c:a aac -f mp4 pipe:1
|
||||
5. P4000 从 relay GET mezzanine → h264_nvenc 编码 → PUT 最终 mp4 到 put_url
|
||||
6. 本客户端通过 get_url(Docker 内网)下载最终文件到 output_path,然后 DELETE 清理
|
||||
7. 删除 relay 上的 mezzanine 临时文件(以及 OSS fallback 的 key)
|
||||
- 首字节用短超时(默认20s),避免链路卡死空等上百秒;首字节到达后放宽到 ffmpeg_timeout+60s
|
||||
6. P4000 从 relay GET mezzanine → h264_nvenc 编码 → PUT 最终 mp4 到 put_url
|
||||
7. 本客户端通过 get_url(Docker 内网)下载最终文件到 output_path,然后 DELETE 清理
|
||||
8. 删除 relay 上的 mezzanine 临时文件(以及 OSS fallback 的 key)
|
||||
|
||||
任何环节失败抛 GpuEncodeError,调用方应 fallback 到 CPU libx264。
|
||||
|
||||
冷启动/链路卡顿背景(2026-09-27 实测):P4000 与 staging 之间走 Tailscale,长时间空闲
|
||||
(>7h)后首次请求曾出现 150s 延迟才真正开始下载 mezzanine,期间 ffmpeg 尚未启动、GPU 空闲。
|
||||
根因在服务端/网络层(可能是 Tailscale DERP 打洞或 httpx 连接池重建),本客户端通过
|
||||
pre_warm + 首字节短超时做兜底:预热打通链路 + 20s 内收不到首字节就快速失败让 CPU fallback,
|
||||
不再让用户等满 150s+。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
@@ -63,6 +71,13 @@ class GpuEncoderClient:
|
||||
mezzanine_transport: str = "relay",
|
||||
sync_timeout: int = 300,
|
||||
health_timeout: float = 3.0,
|
||||
# 提交编码任务前先发一次 /health 预热 Tailscale 链路,避免长时间空闲后首次请求
|
||||
# 因 DERP 打洞/NAT 映射过期/Tailscale 连接重建而阻塞上百秒。
|
||||
pre_warm: bool = True,
|
||||
# POST 首次响应超时:P4000 已收到请求后应该在数秒内开始下载 inputs;
|
||||
# 如果超过这个值还没收到任何响应字节,说明链路/服务卡住,快速失败让调用方 fallback CPU。
|
||||
# 注意:ffmpeg 编码本身靠 body.timeout 控制(300s),不应该被这个超时影响。
|
||||
post_first_byte_timeout: float = 20.0,
|
||||
vcodec: str = "h264_nvenc",
|
||||
preset: str = "p4",
|
||||
crf: int = 23,
|
||||
@@ -80,12 +95,16 @@ class GpuEncoderClient:
|
||||
self.mezzanine_transport = mezzanine_transport.lower() # "relay" | "oss"
|
||||
self.sync_timeout = sync_timeout
|
||||
self.health_timeout = health_timeout
|
||||
self.pre_warm = pre_warm
|
||||
self.post_first_byte_timeout = post_first_byte_timeout
|
||||
self.vcodec = vcodec
|
||||
self.preset = preset
|
||||
self.crf = crf
|
||||
self.bitrate = bitrate
|
||||
self._relay_secret = relay_secret
|
||||
self.oss_tmp_prefix = oss_tmp_prefix.rstrip("/") + "/" if oss_tmp_prefix else "tmp/gpu-mezzanine/"
|
||||
# 上次与 P4000 成功通信的时间戳(用于判断是否需要 pre_warm 预热)
|
||||
self._last_ok_ts: float = 0.0
|
||||
|
||||
RELAY_PATH_PREFIX = "/api/v1/internal/gpu-relay"
|
||||
|
||||
@@ -128,13 +147,16 @@ class GpuEncoderClient:
|
||||
except (urllib.error.URLError, socket.timeout, TimeoutError, json.JSONDecodeError, ConnectionError) as e:
|
||||
return GpuHealth(healthy=False, error=f"health probe failed: {e}")
|
||||
try:
|
||||
return GpuHealth(
|
||||
h = GpuHealth(
|
||||
healthy=data.get("status") == "healthy",
|
||||
worker=str(data.get("worker", "")),
|
||||
gpu_name=(data.get("gpu") or {}).get("name", ""),
|
||||
nvenc_h264=bool((data.get("nvenc") or {}).get("h264_nvenc")),
|
||||
nvenc_hevc=bool((data.get("nvenc") or {}).get("hevc_nvenc")),
|
||||
)
|
||||
if h.healthy:
|
||||
self._last_ok_ts = time.time()
|
||||
return h
|
||||
except Exception as e: # noqa: BLE001
|
||||
return GpuHealth(healthy=False, error=f"malformed health response: {e}")
|
||||
|
||||
@@ -227,7 +249,8 @@ class GpuEncoderClient:
|
||||
ffmpeg_args.append("-an")
|
||||
ffmpeg_args.extend(["-f", "mp4", "pipe:1"])
|
||||
|
||||
# 4. call P4000 sync render
|
||||
# 4. pre-warm then call P4000 sync render
|
||||
self._warm_up_if_needed()
|
||||
body = {
|
||||
"inputs": {"in.mp4": input_url},
|
||||
"ffmpeg_args": ffmpeg_args,
|
||||
@@ -235,6 +258,7 @@ class GpuEncoderClient:
|
||||
"timeout": int(timeout),
|
||||
}
|
||||
job = self._post_sync(body)
|
||||
self._last_ok_ts = time.time()
|
||||
logger.info(
|
||||
"[gpu-encoder] P4000 done: job_id=%s rc=%s size=%s dur=%ss transport=%s",
|
||||
job.get("job_id"),
|
||||
@@ -299,9 +323,38 @@ class GpuEncoderClient:
|
||||
raise GpuEncodeError("GPU_ENCODE_RELAY_SECRET not set")
|
||||
return secret
|
||||
|
||||
def _warm_up_if_needed(self) -> None:
|
||||
"""POST 前预热:如果距上次成功通信超过 idle 阈值,先打 /health 打通 Tailscale 链路。
|
||||
|
||||
背景:Tailscale 在长时间空闲(几小时)后,到对端的直连 NAT 映射可能过期,
|
||||
首次请求会走 DERP 中继打洞;极少数情况下打洞/重连会卡住上百秒(曾观测到 150s 延迟)。
|
||||
预热请求本身走短超时快速失败,不会阻塞主流程;预热成功后再发 POST。
|
||||
"""
|
||||
if not self.pre_warm:
|
||||
return
|
||||
idle = time.time() - self._last_ok_ts
|
||||
# 空闲超过 60s 才预热(正常流水线里相邻任务间隔通常 <10s,没必要每次都打)
|
||||
if idle < 60:
|
||||
return
|
||||
url = f"{self.endpoint}/health"
|
||||
t0 = time.time()
|
||||
try:
|
||||
with urllib.request.urlopen(url, timeout=min(self.health_timeout, 3.0)) as resp:
|
||||
resp.read()
|
||||
self._last_ok_ts = time.time()
|
||||
logger.debug("[gpu-encoder] pre-warm ok: took=%.2fs idle=%.0fs", time.time() - t0, idle)
|
||||
except (urllib.error.URLError, socket.timeout, TimeoutError, ConnectionError, OSError) as e:
|
||||
# 预热失败不致命——主 POST 会带自己的超时,再失败就抛 GpuEncodeError 让调用方 fallback
|
||||
logger.warning("[gpu-encoder] pre-warm probe failed (will try POST anyway): %s", e)
|
||||
|
||||
def _post_sync(self, body: dict[str, Any]) -> dict[str, Any]:
|
||||
url = f"{self.endpoint}/api/render/sync"
|
||||
req_timeout = body.get("timeout", self.sync_timeout) + 60
|
||||
ffmpeg_timeout = body.get("timeout", self.sync_timeout)
|
||||
# 连接 + 首字节用短超时(防链路卡死数百秒);首字节到达后给 ffmpeg 留足编码+上传时间
|
||||
# Python urllib 的 timeout 是整个请求总超时,所以用"两段式":
|
||||
# 阶段1:先 read(1) 拿首字节,用短超时;
|
||||
# 阶段2:再 read() 读完整 body,用 ffmpeg_timeout+60。
|
||||
connect_timeout = min(max(self.post_first_byte_timeout, 5.0), 30.0)
|
||||
payload = json.dumps(body).encode("utf-8")
|
||||
req = urllib.request.Request(
|
||||
url,
|
||||
@@ -310,14 +363,45 @@ class GpuEncoderClient:
|
||||
method="POST",
|
||||
)
|
||||
t0 = time.time()
|
||||
first_byte_ok = False
|
||||
resp = None
|
||||
try:
|
||||
with urllib.request.urlopen(req, timeout=req_timeout) as resp:
|
||||
raw = resp.read().decode("utf-8")
|
||||
resp = urllib.request.urlopen(req, timeout=connect_timeout)
|
||||
# 读首字节 —— 如果 P4000/链路卡死,这里会在 connect_timeout 内抛超时
|
||||
first_chunk = resp.read(1)
|
||||
first_byte_ok = True
|
||||
logger.debug(
|
||||
"[gpu-encoder] P4000 first byte in %.2fs (connect_timeout=%.1fs)",
|
||||
time.time() - t0,
|
||||
connect_timeout,
|
||||
)
|
||||
# 剩余用长超时(给底层socket放宽时限;如果是mock/不支持,则跳过)
|
||||
try:
|
||||
resp.fp._sock.settimeout(ffmpeg_timeout + 60)
|
||||
except (AttributeError, OSError):
|
||||
pass
|
||||
rest = resp.read()
|
||||
raw = (first_chunk + rest).decode("utf-8")
|
||||
resp.close()
|
||||
resp = None
|
||||
except urllib.error.HTTPError as e:
|
||||
detail = e.read().decode("utf-8", errors="replace")[:1000]
|
||||
raise GpuEncodeError(f"P4000 HTTP {e.code}: {detail}") from e
|
||||
except (urllib.error.URLError, socket.timeout, TimeoutError, ConnectionError) as e:
|
||||
raise GpuEncodeError(f"P4000 connection error: {e}") from e
|
||||
except (urllib.error.URLError, socket.timeout, TimeoutError, ConnectionError, OSError) as e:
|
||||
waited = time.time() - t0
|
||||
hint = "first-byte" if not first_byte_ok else "ffmpeg/upload"
|
||||
# 统一以 "connection error" 开头,便于上层 fallback 逻辑用关键词识别;
|
||||
# 末尾再附带具体错误(timed out / refused ...)供排障
|
||||
raise GpuEncodeError(
|
||||
f"P4000 {hint} connection error after {waited:.1f}s "
|
||||
f"(connect_timeout={connect_timeout:.0f}s, ffmpeg_timeout={ffmpeg_timeout}s): {e}"
|
||||
) from e
|
||||
finally:
|
||||
if resp is not None:
|
||||
try:
|
||||
resp.close()
|
||||
except Exception:
|
||||
pass
|
||||
try:
|
||||
result = json.loads(raw)
|
||||
except json.JSONDecodeError as e:
|
||||
@@ -451,6 +535,8 @@ def _build_client_from_settings() -> Optional[GpuEncoderClient]:
|
||||
mezzanine_transport=getattr(settings, "gpu_encode_mezzanine_transport", "relay") or "relay",
|
||||
sync_timeout=getattr(settings, "gpu_encode_sync_timeout", 300),
|
||||
health_timeout=getattr(settings, "gpu_encode_health_timeout", 3.0),
|
||||
pre_warm=getattr(settings, "gpu_encode_pre_warm", True),
|
||||
post_first_byte_timeout=getattr(settings, "gpu_encode_post_first_byte_timeout", 20.0),
|
||||
vcodec=getattr(settings, "gpu_encode_vcodec", "h264_nvenc"),
|
||||
preset=getattr(settings, "gpu_encode_preset", "p4"),
|
||||
crf=getattr(settings, "gpu_encode_crf", 23),
|
||||
|
||||
Reference in New Issue
Block a user