Spaces:
Sleeping
Sleeping
File size: 9,371 Bytes
bac4901 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 | """
space_client.py — HTTP client for distributed mode.
Calls the remote Face Space and Object Space for AI inference.
Handles:
- Persistent aiohttp session (created once at startup)
- Exponential-backoff retry for cold-start wakeups (HF free tier)
- Keep-alive background pinger to prevent spaces sleeping
- Structured exceptions that search.py / upload.py already handle
"""
import asyncio
import io
import logging
import os
import aiohttp
# ── Space URLs (set these as secrets in the Gateway HF Space) ────────────────
FACE_SPACE_URL = os.getenv("FACE_SPACE_URL", "").rstrip("/")
OBJECT_SPACE_URL = os.getenv("OBJECT_SPACE_URL", "").rstrip("/")
# ── Timeouts ─────────────────────────────────────────────────────────────────
# HF free-tier spaces can take 60-90 s to cold-start.
# Keep TOTAL high so we wait through the wake-up; CONNECT stays short.
_CONNECT_TIMEOUT = 10 # seconds — TCP handshake
_TOTAL_TIMEOUT = 120 # seconds — full request including model inference
# ── Retry config (exponential backoff for cold starts) ────────────────────────
_MAX_RETRIES = 3
_RETRY_BACKOFF = [5, 15, 30] # seconds to wait before each retry
logger = logging.getLogger("space_client")
# ── Singleton aiohttp session ─────────────────────────────────────────────────
_session: aiohttp.ClientSession | None = None
async def get_session() -> aiohttp.ClientSession:
"""Return (or lazily create) the shared aiohttp session."""
global _session
if _session is None or _session.closed:
connector = aiohttp.TCPConnector(
limit=20, # max simultaneous connections
limit_per_host=10,
keepalive_timeout=60,
)
timeout = aiohttp.ClientTimeout(
connect=_CONNECT_TIMEOUT,
total=_TOTAL_TIMEOUT,
)
_session = aiohttp.ClientSession(connector=connector, timeout=timeout)
return _session
async def close_session() -> None:
"""Close the session gracefully (called from lifespan shutdown)."""
global _session
if _session and not _session.closed:
await _session.close()
_session = None
# ── Core POST helper with retry ───────────────────────────────────────────────
async def _post_form_data(url: str, form: aiohttp.FormData, label: str) -> dict:
"""
POST multipart/form-data to `url` with retry + backoff.
Raises RuntimeError on final failure so callers can log a warning
and degrade gracefully (e.g. return [] instead of crashing).
"""
session = await get_session()
last_exc: Exception | None = None
for attempt in range(_MAX_RETRIES):
try:
async with session.post(url, data=form) as resp:
if resp.status == 200:
return await resp.json()
# 503 = space still waking up — retry
if resp.status == 503:
body = await resp.text()
logger.warning(
"[space_client] %s returned 503 (attempt %d/%d): %s",
label, attempt + 1, _MAX_RETRIES, body[:200],
)
last_exc = RuntimeError(f"{label} 503: {body[:200]}")
else:
body = await resp.text()
raise RuntimeError(f"{label} HTTP {resp.status}: {body[:300]}")
except asyncio.TimeoutError as e:
logger.warning(
"[space_client] %s timeout (attempt %d/%d)", label, attempt + 1, _MAX_RETRIES
)
last_exc = e
except aiohttp.ClientConnectionError as e:
logger.warning(
"[space_client] %s connection error (attempt %d/%d): %s",
label, attempt + 1, _MAX_RETRIES, e,
)
last_exc = e
# Wait before retrying (skip wait after last attempt)
if attempt < _MAX_RETRIES - 1:
wait = _RETRY_BACKOFF[attempt]
logger.info("[space_client] Retrying %s in %ds…", label, wait)
await asyncio.sleep(wait)
raise RuntimeError(f"{label} failed after {_MAX_RETRIES} attempts: {last_exc}")
# ── Public API: embed_face ────────────────────────────────────────────────────
async def embed_face(image_bytes: bytes, quality_gate: float = 0.35) -> list[dict]:
"""
Send image bytes to the Face Space and return a list of face dicts.
Returns [] on failure (caller logs search.face_space_unavailable).
"""
if not FACE_SPACE_URL:
logger.error("[space_client] FACE_SPACE_URL is not set.")
return []
url = f"{FACE_SPACE_URL}/embed/face"
try:
form = aiohttp.FormData()
form.add_field(
"file",
io.BytesIO(image_bytes),
filename="image.jpg",
content_type="image/jpeg",
)
form.add_field("quality_gate", str(quality_gate))
result = await _post_form_data(url, form, "face-space")
faces = result.get("faces", [])
logger.info("[space_client] embed_face → %d face(s)", len(faces))
return faces
except Exception as e:
logger.warning("[space_client] embed_face failed: %s", e)
return [] # Degrade gracefully — search.py handles empty list
# ── Public API: embed_object ──────────────────────────────────────────────────
async def embed_object(image_bytes: bytes) -> list[dict]:
"""
Send image bytes to the Object Space and return a list of object dicts.
Returns [] on failure (caller raises 503 to the user).
"""
if not OBJECT_SPACE_URL:
logger.error("[space_client] OBJECT_SPACE_URL is not set.")
return []
url = f"{OBJECT_SPACE_URL}/embed/object"
try:
form = aiohttp.FormData()
form.add_field(
"file",
io.BytesIO(image_bytes),
filename="image.jpg",
content_type="image/jpeg",
)
result = await _post_form_data(url, form, "object-space")
objects = result.get("objects", [])
logger.info("[space_client] embed_object → %d object vec(s)", len(objects))
return objects
except Exception as e:
logger.warning("[space_client] embed_object failed: %s", e)
return []
# ── Health check / ping ───────────────────────────────────────────────────────
async def ping_spaces() -> dict:
"""
Ping both spaces at startup (wakes them from sleep).
Returns {"face": bool, "object": bool}.
"""
session = await get_session()
results = {"face": False, "object": False}
async def _ping(url: str, key: str) -> None:
try:
async with session.get(
f"{url}/health",
timeout=aiohttp.ClientTimeout(total=90), # allow cold-start wake
) as resp:
results[key] = resp.status == 200
logger.info("[space_client] ping %s → %d", key, resp.status)
except Exception as e:
logger.warning("[space_client] ping %s failed: %s", key, e)
ping_tasks = []
if FACE_SPACE_URL:
ping_tasks.append(_ping(FACE_SPACE_URL, "face"))
if OBJECT_SPACE_URL:
ping_tasks.append(_ping(OBJECT_SPACE_URL, "object"))
await asyncio.gather(*ping_tasks)
return results
# ── Keep-warm background task ─────────────────────────────────────────────────
async def keep_spaces_warm(interval_seconds: int = 540) -> None:
"""
Ping both spaces every `interval_seconds` (default 9 min) to prevent
HF free-tier sleep (spaces sleep after ~15 min of inactivity).
Start this as a background task from your lifespan:
asyncio.create_task(keep_spaces_warm())
"""
session = await get_session()
while True:
await asyncio.sleep(interval_seconds)
urls = []
if FACE_SPACE_URL:
urls.append((FACE_SPACE_URL, "face"))
if OBJECT_SPACE_URL:
urls.append((OBJECT_SPACE_URL, "object"))
async def _warm(url: str, label: str) -> None:
try:
async with session.get(
f"{url}/health",
timeout=aiohttp.ClientTimeout(total=30),
) as resp:
logger.debug("[space_client] keep-warm %s → %d", label, resp.status)
except Exception as e:
logger.debug("[space_client] keep-warm %s error: %s", label, e)
await asyncio.gather(*[_warm(u, l) for u, l in urls], return_exceptions=True) |