orchestrator/runner.py

"""
Orchestrator — async fan-out across providers.

Responsibilities:
  - Receive a PipelineOutput (never raw user input).
  - For each target provider:
      * Skip if circuit breaker is open.
      * Check cache; if hit, return cached ProviderResult.
      * Otherwise invoke provider (sync → asyncio.to_thread).
      * Apply retry policy.
      * Record metrics (latency, success/failure, retry count).
      * Cache successful results.
  - Return a dict mapping provider name → ProviderResult.

All dependencies (registry, cache, metrics, health, retry policy) are
injected via the constructor.
"""

from __future__ import annotations

import asyncio
from typing import Awaitable, Callable, Dict, Iterable, Optional

from loguru import logger

from config.settings import Settings
from metrics.collector import MetricsCollector
from orchestrator.health import HealthMonitor
from orchestrator.retry import RetryPolicy, with_retry_sync
from pipeline.feature_extraction import PipelineOutput
from providers.base import Provider, ProviderCapability, ProviderResult
from providers.registry import ProviderRegistry
from storage.cache import Cache
from utils.logging import execution_context, new_execution_id


class Orchestrator:
    """Async fan-out across providers.  Injectable."""

    def __init__(
        self,
        registry: ProviderRegistry,
        cache: Cache,
        metrics: MetricsCollector,
        health: HealthMonitor,
        settings: Settings,
        retry_policy: Optional[RetryPolicy] = None,
    ) -> None:
        self._registry = registry
        self._cache = cache
        self._metrics = metrics
        self._health = health
        self._settings = settings
        self._retry = retry_policy or RetryPolicy(
            max_attempts=settings.retry_max_attempts,
            initial_backoff_seconds=settings.retry_initial_backoff_seconds,
            max_backoff_seconds=settings.retry_max_backoff_seconds,
        )

    # ------------------------------------------------------------------ #
    # Public API
    # ------------------------------------------------------------------ #
    async def run(
        self,
        pipeline_output: PipelineOutput,
        capabilities: Iterable[ProviderCapability],
        provider_whitelist: Optional[list[str]] = None,
        execution_id: Optional[str] = None,
    ) -> Dict[str, ProviderResult]:
        """Fan out across all providers matching the given capabilities.

        Args:
            pipeline_output: normalized image + face crops.
            capabilities: which capability buckets to invoke.
            provider_whitelist: optional list of provider names; if set,
                only those providers are invoked.
            execution_id: optional trace id for structured logging.

        Returns:
            dict mapping provider name → ProviderResult.
        """
        eid = execution_id or new_execution_id()
        targets: list[Provider] = []
        for cap in capabilities:
            for p in self._registry.list_by_capability(cap):
                if provider_whitelist and p.name not in provider_whitelist:
                    continue
                if not self._health.is_available(p.name):
                    logger.info(f"[orchestrator] skipping {p.name}: circuit open")
                    continue
                targets.append(p)

        if not targets:
            logger.warning("[orchestrator] no providers to invoke")
            return {}

        semaphore = asyncio.Semaphore(self._settings.orchestrator_max_concurrency)

        async def _wrapped(provider: Provider) -> tuple[str, ProviderResult]:
            async with semaphore:
                return provider.name, await self._invoke_one(provider, pipeline_output, eid)

        tasks = [_wrapped(p) for p in targets]
        gathered = await asyncio.gather(*tasks, return_exceptions=False)
        return dict(gathered)

    # ------------------------------------------------------------------ #
    # Per-provider invocation (with cache + retry + metrics)
    # ------------------------------------------------------------------ #
    async def _invoke_one(
        self,
        provider: Provider,
        pipeline_output: PipelineOutput,
        execution_id: str,
    ) -> ProviderResult:
        cache_key = self._cache_key(provider, pipeline_output)

        # Cache hit?
        if self._settings.cache_enabled:
            cached = self._cache.get(cache_key)
            if cached is not None:
                self._metrics.counters.inc("cache.hits")
                logger.debug(f"[orchestrator] cache hit for {provider.name}")
                cached.metadata["cache_hit"] = True
                return cached
        self._metrics.counters.inc("cache.misses")

        # Invoke with retry
        retry_count = 0

        def _on_retry(attempt: int, exc: Exception) -> None:
            nonlocal retry_count
            retry_count = attempt
            self._metrics.providers.record_retry(provider.name)
            self._metrics.counters.inc(f"retries.{provider.name}")

        with execution_context(execution_id=execution_id,
                               provider_id=provider.name,
                               retry_count=retry_count):
            logger.info(f"[orchestrator] invoking {provider.name}")

            def _call_sync() -> ProviderResult:
                return with_retry_sync(
                    lambda: provider.execute(pipeline_output),
                    self._retry,
                    label=provider.name,
                    on_retry=_on_retry,
                )

            try:
                result = await asyncio.wait_for(
                    asyncio.to_thread(_call_sync),
                    timeout=self._settings.orchestrator_timeout_seconds,
                )
            except asyncio.TimeoutError:
                result = ProviderResult(
                    provider=provider.name,
                    capability=provider.capability,
                    success=False,
                    elapsed_ms=self._settings.orchestrator_timeout_seconds * 1000,
                    error="Orchestrator timeout",
                    error_type="TimeoutError",
                    retry_count=retry_count,
                )
            except Exception as e:
                result = ProviderResult(
                    provider=provider.name,
                    capability=provider.capability,
                    success=False,
                    elapsed_ms=0.0,
                    error=str(e),
                    error_type=type(e).__name__,
                    retry_count=retry_count,
                )

        result.retry_count = retry_count

        # Record metrics
        self._metrics.providers.record_invocation(provider.name)
        if result.success:
            self._metrics.providers.record_success(provider.name, result.elapsed_ms)
            self._health.record_success(provider.name, result.elapsed_ms)
            self._metrics.timings.record(f"provider.{provider.name}", result.elapsed_ms)
            # Cache
            if self._settings.cache_enabled:
                self._cache.set(cache_key, result)
        else:
            self._metrics.providers.record_failure(provider.name, result.error or "")
            self._health.record_failure(provider.name)
            self._metrics.counters.inc(f"failures.{provider.name}")

        return result

    # ------------------------------------------------------------------ #
    # Cache key
    # ------------------------------------------------------------------ #
    @staticmethod
    def _cache_key(provider: Provider, pipeline_output: PipelineOutput) -> str:
        return f"{provider.name}:{pipeline_output.image_hash}"