"""Global rate limiter for API requests.""" import asyncio import random import time from collections.abc import AsyncIterator, Callable from contextlib import asynccontextmanager from typing import Any, ClassVar, TypeVar import httpx import openai from loguru import logger from core.rate_limit import StrictSlidingWindowLimiter T = TypeVar("T") class AdaptiveRateLimiter: """Adaptive rate limiter that backs off on 429s and recovers gradually. Starts at a high throughput and auto-adjusts based on upstream feedback. This gives maximum throughput in normal conditions while self-correcting when rate limits are hit. """ _limiter_count: ClassVar[int] = 0 def __init__( self, initial_rate: int = 100, min_rate: int = 10, window: float = 60.0, backoff_factor: float = 0.5, recovery_factor: float = 1.2, ) -> None: self._initial_rate = initial_rate self._current_rate = initial_rate self._min_rate = min_rate self._window = window self._backoff_factor = backoff_factor self._recovery_factor = recovery_factor self._limiter = StrictSlidingWindowLimiter(initial_rate, window) self._lock = asyncio.Lock() self._success_streak: int = 0 self._instance_id = AdaptiveRateLimiter._limiter_count AdaptiveRateLimiter._limiter_count += 1 async def acquire(self) -> None: await self._limiter.acquire() def record_429(self) -> None: """Called when a 429 is received — reduce rate immediately.""" self._current_rate = max( self._min_rate, int(self._current_rate * self._backoff_factor) ) self._limiter = StrictSlidingWindowLimiter(self._current_rate, self._window) self._success_streak = 0 logger.warning( "ADAPTIVE_RATE: instance={} backed off to {} req/min (429 received)", self._instance_id, self._current_rate, ) def record_success(self) -> None: """Called on success — gradually recover rate if below initial.""" if self._current_rate >= self._initial_rate: self._success_streak = 0 return self._success_streak += 1 # Recover after 3 consecutive successes if self._success_streak >= 3: self._current_rate = min( self._initial_rate, int(self._current_rate * self._recovery_factor), ) self._limiter = StrictSlidingWindowLimiter(self._current_rate, self._window) self._success_streak = 0 logger.info( "ADAPTIVE_RATE: instance={} recovered to {} req/min", self._instance_id, self._current_rate, ) class ModelHealthTracker: """Track per-model health based on recent failures.""" _instance: ClassVar[ModelHealthTracker | None] = None def __init__( self, failure_ttl: float = 30.0, max_failures: int = 3, *, failure_ttl_nim: float = 15.0, max_failures_nim: int = 2, failure_ttl_zen: float = 60.0, max_failures_zen: int = 5, ) -> None: self._failure_ttl = failure_ttl self._max_failures = max_failures self._failure_ttl_nim = failure_ttl_nim self._max_failures_nim = max_failures_nim self._failure_ttl_zen = failure_ttl_zen self._max_failures_zen = max_failures_zen self._failures: dict[str, list[float]] = {} self._failure_ttls: dict[str, float] = {} self._max_failures_map: dict[str, int] = {} @classmethod def get_instance(cls) -> ModelHealthTracker: if cls._instance is None: cls._instance = cls() return cls._instance def _params_for(self, model_ref: str) -> tuple[float, int]: """Return (failure_ttl, max_failures) for a model based on provider.""" if model_ref in self._failure_ttls: return self._failure_ttls[model_ref], self._max_failures_map[model_ref] if model_ref.startswith("zen/"): return self._failure_ttl_zen, self._max_failures_zen if model_ref.startswith("nvidia_nim/"): return self._failure_ttl_nim, self._max_failures_nim return self._failure_ttl, self._max_failures def record_failure(self, model_ref: str) -> None: """Record a failure timestamp for a model.""" now = time.monotonic() if model_ref not in self._failures: self._failures[model_ref] = [] self._failures[model_ref].append(now) logger.debug("HEALTH: recorded failure for '{}'", model_ref) def is_healthy(self, model_ref: str) -> bool: """Check if model has had fewer than max_failures in the TTL window.""" if model_ref not in self._failures: return True ttl, max_f = self._params_for(model_ref) cutoff = time.monotonic() - ttl recent = [t for t in self._failures[model_ref] if t > cutoff] self._failures[model_ref] = recent healthy = len(recent) < max_f if not healthy: logger.debug( "HEALTH: model '{}' is unhealthy ({} failures in {}s)", model_ref, len(recent), ttl, ) return healthy def get_failure_count(self, model_ref: str) -> int: """Get number of recent failures for a model.""" if model_ref not in self._failures: return 0 ttl, _ = self._params_for(model_ref) cutoff = time.monotonic() - ttl return len([t for t in self._failures[model_ref] if t > cutoff]) def clear_failures(self, model_ref: str) -> None: """Clear failure history for a model (on success).""" if model_ref in self._failures: self._failures.pop(model_ref) class GlobalRateLimiter: """ Global singleton rate limiter that blocks all requests when a rate limit error is encountered (reactive) and throttles requests (proactive) using a strict rolling window. Optionally enforces a max_concurrency cap: at most N provider streams may be open simultaneously, independent of the sliding window. Proactive limits - throttles requests to stay within API limits. Reactive limits - pauses all requests when a 429 is hit. Concurrency limit - caps simultaneously open streams. """ _instance: ClassVar[GlobalRateLimiter | None] = None _scoped_instances: ClassVar[dict[str, GlobalRateLimiter]] = {} def __init__( self, rate_limit: int = 40, rate_window: float = 60.0, max_concurrency: int = 5, adaptive_rate: int | None = None, adaptive_min_rate: int = 10, ): # Prevent re-initialization on singleton reuse if hasattr(self, "_initialized"): return if rate_limit <= 0: raise ValueError("rate_limit must be > 0") if rate_window <= 0: raise ValueError("rate_window must be > 0") if max_concurrency <= 0: raise ValueError("max_concurrency must be > 0") self._rate_limit = rate_limit self._rate_window = float(rate_window) self._max_concurrency = max_concurrency self._adaptive_rate = adaptive_rate self._adaptive_min_rate = adaptive_min_rate if adaptive_rate is not None: self._proactive_limiter = AdaptiveRateLimiter( initial_rate=adaptive_rate, min_rate=adaptive_min_rate, window=float(rate_window), ) else: self._proactive_limiter = StrictSlidingWindowLimiter( rate_limit, float(rate_window) ) self._blocked_until: float = 0 self._concurrency_sem = asyncio.Semaphore(max_concurrency) self._initialized = True limiter_type = ( f"Adaptive({adaptive_rate}→{adaptive_min_rate})" if adaptive_rate is not None else f"Strict({rate_limit})" ) logger.info( f"GlobalRateLimiter initialized {limiter_type} / {rate_window}s, max_concurrency={max_concurrency}" ) @classmethod def get_instance( cls, rate_limit: int | None = None, rate_window: float | None = None, max_concurrency: int = 5, ) -> GlobalRateLimiter: """Get or create the singleton instance. Args: rate_limit: Requests per window (only used on first creation) rate_window: Window in seconds (only used on first creation) max_concurrency: Max simultaneous open streams (only used on first creation) """ if cls._instance is None: cls._instance = cls( rate_limit=rate_limit or 40, rate_window=rate_window or 60.0, max_concurrency=max_concurrency, ) return cls._instance @classmethod def get_scoped_instance( cls, scope: str, *, rate_limit: int | None = None, rate_window: float | None = None, max_concurrency: int = 5, adaptive_rate: int | None = None, adaptive_min_rate: int = 10, ) -> GlobalRateLimiter: """Get or create a provider-scoped limiter instance. Zen gets unlimited adaptive rate (9999) since it has no rate limits. NIM gets adaptive rate from nim_rate_limit setting. """ if not scope: raise ValueError("scope must be non-empty") desired_rate_limit = 9999 if scope == "zen" else rate_limit or 40 desired_rate_window = float(rate_window or 60.0) existing = cls._scoped_instances.get(scope) if existing and existing.matches_config( desired_rate_limit, desired_rate_window, max_concurrency ): return existing if existing: logger.info( "Rebuilding provider rate limiter for updated scope '{}'", scope ) # Adaptive rate only for NIM (not Zen which is unlimited) use_adaptive = adaptive_rate if scope == "nvidia_nim" else None cls._scoped_instances[scope] = cls( rate_limit=desired_rate_limit, rate_window=desired_rate_window, max_concurrency=max_concurrency, adaptive_rate=use_adaptive, adaptive_min_rate=adaptive_min_rate, ) return cls._scoped_instances[scope] @classmethod def reset_instance(cls) -> None: """Reset singleton (for testing).""" cls._instance = None cls._scoped_instances = {} async def wait_if_blocked(self) -> bool: """ Wait if currently rate limited or throttle to meet quota. Returns: True if was reactively blocked and waited, False otherwise. """ # 1. Reactive check: Wait if someone hit a 429 waited_reactively = False now = time.monotonic() if now < self._blocked_until: wait_time = self._blocked_until - now logger.warning( f"Global provider rate limit active (reactive), waiting {wait_time:.1f}s..." ) await asyncio.sleep(wait_time) waited_reactively = True # 2. Proactive check: strict rolling window (no bursts beyond N in last W seconds) await self._acquire_proactive_slot() return waited_reactively async def _acquire_proactive_slot(self) -> None: """ Acquire a proactive slot enforcing a strict rolling window. Guarantees: at most `self._rate_limit` acquisitions in any interval of length `self._rate_window` (seconds). """ await self._proactive_limiter.acquire() def set_blocked(self, seconds: float = 60) -> None: """ Set global block for specified seconds (reactive). Args: seconds: How long to block (default 60s) """ self._blocked_until = time.monotonic() + seconds logger.warning(f"Global provider rate limit set for {seconds:.1f}s (reactive)") def is_blocked(self) -> bool: """Check if currently reactively blocked.""" return time.monotonic() < self._blocked_until def matches_config( self, rate_limit: int, rate_window: float, max_concurrency: int ) -> bool: """Return whether this limiter matches the requested runtime config.""" return ( self._rate_limit == rate_limit and self._rate_window == float(rate_window) and self._max_concurrency == max_concurrency ) def remaining_wait(self) -> float: """Get remaining reactive wait time in seconds.""" return max(0.0, self._blocked_until - time.monotonic()) def record_failure(self, model_ref: str | None = None) -> None: """Record a failure for rate limit tracking. Args: model_ref: Optional model identifier for health tracking. """ # Record in the shared health tracker if model provided if model_ref: health = ModelHealthTracker.get_instance() health.record_failure(model_ref) def is_healthy(self, model_ref: str | None = None) -> bool: """Check if provider/model is healthy based on failure history. Args: model_ref: Optional model identifier for health tracking. Returns: True if no recent failures or model_ref is None. """ if model_ref is None: return True health = ModelHealthTracker.get_instance() return health.is_healthy(model_ref) @asynccontextmanager async def concurrency_slot(self) -> AsyncIterator[None]: """Async context manager that holds one concurrency slot for a stream. Blocks until a slot is available (controlled by max_concurrency). """ await self._concurrency_sem.acquire() try: yield finally: self._concurrency_sem.release() async def execute_with_retry( self, fn: Callable[..., Any], *args: Any, max_retries: int = 3, base_delay: float = 0.3, max_delay: float = 20.0, jitter: float = 0.1, **kwargs: Any, ) -> Any: """Execute an async callable with rate limiting and retry on 429. Waits for the proactive limiter before each attempt. On 429, applies adaptive backoff and notifies the adaptive rate limiter. Snappier recovery than fixed delays. Args: fn: Async callable to execute. max_retries: Maximum number of retry attempts after the first failure. base_delay: Base delay in seconds for exponential backoff. max_delay: Maximum delay cap in seconds. jitter: Maximum random jitter in seconds added to each delay. Returns: The result of the callable. Raises: The last exception if all retries are exhausted. """ last_exc: Exception | None = None for attempt in range(1 + max_retries): await self.wait_if_blocked() try: result = await fn(*args, **kwargs) # Notify adaptive limiter of success (triggers gradual recovery) self._record_success_for_adaptive() return result except openai.RateLimitError as e: last_exc = e self._record_429_for_adaptive() if attempt >= max_retries: logger.warning( f"Rate limit retry exhausted after {max_retries} retries" ) break delay = min(base_delay * (2**attempt), max_delay) delay += random.uniform(0, jitter) logger.warning( f"Rate limited (429), attempt {attempt + 1}/{max_retries + 1}. " f"Retrying in {delay:.1f}s..." ) self.set_blocked(delay) await asyncio.sleep(delay) except httpx.HTTPStatusError as e: if e.response.status_code != 429: raise last_exc = e self._record_429_for_adaptive() if attempt >= max_retries: logger.warning( f"HTTP 429 retry exhausted after {max_retries} retries" ) break delay = min(base_delay * (2**attempt), max_delay) delay += random.uniform(0, jitter) logger.warning( f"HTTP 429 from upstream, attempt {attempt + 1}/{max_retries + 1}. " f"Retrying in {delay:.1f}s..." ) self.set_blocked(delay) await asyncio.sleep(delay) assert last_exc is not None raise last_exc def _record_429_for_adaptive(self) -> None: """Notify adaptive limiter of a 429 — triggers rate backoff.""" if isinstance(self._proactive_limiter, AdaptiveRateLimiter): self._proactive_limiter.record_429() def _record_success_for_adaptive(self) -> None: """Notify adaptive limiter of success — triggers gradual rate recovery.""" if isinstance(self._proactive_limiter, AdaptiveRateLimiter): self._proactive_limiter.record_success()