File size: 7,981 Bytes
0703e2b | 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 | """Retries, timeouts, rate limiting and concurrency caps for LLMs and retrievers.
from fgao.resilience import resilient
llm = resilient(OpenAILLM("gpt-4o-mini"), attempts=4, timeout=20, rate=50, max_concurrency=64)
kb = resilient(my_retriever, attempts=3, timeout=5)
``resilient(x)`` returns an object with the same interface as ``x`` (``complete`` /
``acomplete`` for an LLM, ``retrieve`` / ``aretrieve`` for a retriever):
* retries with exponential backoff and full jitter on any exception in ``retry_on``
* per-attempt ``timeout`` (seconds) on the async path; the sync path relies on the
client's own timeout, which you should also set
* ``rate``: token-bucket limit on calls per second (shared by all threads and tasks)
* ``max_concurrency``: cap on simultaneous calls
* ``stats``: calls, attempts, retries, timeouts, failures (thread-safe counters)
A call that still fails after the last attempt raises. Combine with
``FGAORAG(..., on_error="record")`` to let FGAO continue without that operation.
"""
from __future__ import annotations
import asyncio
import random
import threading
import time
from dataclasses import dataclass, field
from typing import Any, Dict, Optional, Tuple, Type
@dataclass
class RetryPolicy:
attempts: int = 3
backoff: float = 0.5 # base delay (s); attempt i waits U(0, min(max_backoff, backoff * 2**i))
max_backoff: float = 8.0
timeout: Optional[float] = None # per attempt, async path
retry_on: Tuple[Type[BaseException], ...] = (Exception,)
def delay(self, i: int, rng: random.Random) -> float:
return rng.uniform(0.0, min(self.max_backoff, self.backoff * (2 ** i)))
class RateLimiter:
"""Token bucket: ``rate`` calls per second, burst ``burst`` (default: ``rate``)."""
def __init__(self, rate: float, burst: Optional[float] = None) -> None:
if rate <= 0:
raise ValueError("rate must be positive")
self.rate = float(rate)
self.capacity = float(burst if burst is not None else max(1.0, rate))
self.tokens = self.capacity
self.t = time.monotonic()
self._lock = threading.Lock()
def _reserve(self) -> float:
"""Take a token; return how long the caller must wait before using it."""
with self._lock:
now = time.monotonic()
self.tokens = min(self.capacity, self.tokens + (now - self.t) * self.rate)
self.t = now
self.tokens -= 1.0
return 0.0 if self.tokens >= 0 else -self.tokens / self.rate
def acquire(self) -> None:
w = self._reserve()
if w > 0:
time.sleep(w)
async def aacquire(self) -> None:
w = self._reserve()
if w > 0:
await asyncio.sleep(w)
@dataclass
class Stats:
calls: int = 0
attempts: int = 0
retries: int = 0
timeouts: int = 0
failures: int = 0
_lock: threading.Lock = field(default_factory=threading.Lock, repr=False, compare=False)
def add(self, **kw: int) -> None:
with self._lock:
for k, v in kw.items():
setattr(self, k, getattr(self, k) + v)
def as_dict(self) -> Dict[str, int]:
return {k: getattr(self, k) for k in ("calls", "attempts", "retries", "timeouts", "failures")}
class _Resilient:
def __init__(self, inner: Any, policy: RetryPolicy, rate: Optional[float], max_concurrency: Optional[int],
name: str, seed: Optional[int]) -> None:
self.inner = inner
self.policy = policy
self.limiter = RateLimiter(rate) if rate else None
self.max_concurrency = max_concurrency
self._tsem = threading.BoundedSemaphore(max_concurrency) if max_concurrency else None
self._asems: Dict[int, asyncio.Semaphore] = {}
self.name = name
self.stats = Stats()
self._rng = random.Random(seed)
def _asem(self) -> Optional[asyncio.Semaphore]:
if not self.max_concurrency:
return None
loop = id(asyncio.get_running_loop())
sem = self._asems.get(loop)
if sem is None:
sem = self._asems[loop] = asyncio.Semaphore(self.max_concurrency)
return sem
def _sync(self, fn, *args):
p = self.policy
self.stats.add(calls=1)
for i in range(max(1, p.attempts)):
if self.limiter:
self.limiter.acquire()
self.stats.add(attempts=1)
try:
if self._tsem:
with self._tsem:
return fn(*args)
return fn(*args)
except p.retry_on:
if i + 1 >= p.attempts:
self.stats.add(failures=1)
raise
self.stats.add(retries=1)
time.sleep(p.delay(i, self._rng))
raise RuntimeError("unreachable") # pragma: no cover
async def _async(self, make, *args):
p = self.policy
self.stats.add(calls=1)
sem = self._asem()
for i in range(max(1, p.attempts)):
if self.limiter:
await self.limiter.aacquire()
self.stats.add(attempts=1)
try:
coro = make(*args)
if sem is not None:
async with sem:
return await (asyncio.wait_for(coro, p.timeout) if p.timeout else coro)
return await (asyncio.wait_for(coro, p.timeout) if p.timeout else coro)
except asyncio.TimeoutError:
self.stats.add(timeouts=1)
if i + 1 >= p.attempts:
self.stats.add(failures=1)
raise
self.stats.add(retries=1)
except asyncio.CancelledError:
raise
except p.retry_on:
if i + 1 >= p.attempts:
self.stats.add(failures=1)
raise
self.stats.add(retries=1)
await asyncio.sleep(p.delay(i, self._rng))
raise RuntimeError("unreachable") # pragma: no cover
class ResilientLLM(_Resilient):
def complete(self, prompt: str) -> str:
return self._sync(self.inner.complete, prompt)
async def acomplete(self, prompt: str) -> str:
from .rag.llms import acomplete
return await self._async(lambda p: acomplete(self.inner, p), prompt)
class ResilientRetriever(_Resilient):
def retrieve(self, query: str, k: int):
return self._sync(self.inner.retrieve, query, k)
async def aretrieve(self, query: str, k: int):
from .rag.retrievers import aretrieve
return await self._async(lambda q, kk: aretrieve(self.inner, q, kk), query, k)
def resilient(x: Any, *, attempts: int = 3, backoff: float = 0.5, max_backoff: float = 8.0,
timeout: Optional[float] = None, rate: Optional[float] = None, max_concurrency: Optional[int] = None,
retry_on: Tuple[Type[BaseException], ...] = (Exception,), name: str = "", seed: Optional[int] = None,
kind: Optional[str] = None):
"""Wrap an LLM or retriever (anything ``as_llm`` / ``as_retriever`` accepts).
``kind="llm"`` or ``"retriever"`` resolves objects that look like both
(e.g. LangChain runnables).
"""
from .rag.llms import as_llm
from .rag.retrievers import as_retriever
policy = RetryPolicy(attempts, backoff, max_backoff, timeout, retry_on)
looks_retriever = (hasattr(x, "retrieve") or isinstance(x, (list, tuple)) or hasattr(x, "similarity_search")
or hasattr(x, "_get_relevant_documents"))
if kind == "retriever" or (kind is None and looks_retriever):
return ResilientRetriever(as_retriever(x), policy, rate, max_concurrency, name or "retriever", seed)
return ResilientLLM(as_llm(x), policy, rate, max_concurrency, name or "llm", seed)
__all__ = ["resilient", "RetryPolicy", "RateLimiter", "ResilientLLM", "ResilientRetriever", "Stats"]
|