Разработка системы постепенного внедрения AI-агентов (Canary Deployment)

Разработка системы canary-деплоя AI-агентов

Canary deployment для AI-агентов — постепенное увеличение трафика на новую версию с автоматическим откатом при деградации метрик. Снижает риск mass-rollout дефектных промптов или моделей.

Canary pipeline

v1.3.1 (100% трафика)
    ↓ деплой v1.3.2
v1.3.1 (95%) + v1.3.2 (5%) — наблюдение 30 мин
    ↓ всё ОК
v1.3.1 (75%) + v1.3.2 (25%) — наблюдение 1 час
    ↓ всё ОК
v1.3.1 (50%) + v1.3.2 (50%) — наблюдение 2 часа
    ↓ всё ОК
v1.3.2 (100%) — полный rollout
    ↓ при любом этапе деградация
автоматический rollback на v1.3.1

Реализация canary controller

@dataclass
class CanaryDeployment:
    deployment_id: str
    agent_name: str
    stable_version: str
    canary_version: str
    stages: list[CanaryStage]   # [(5%, 30min), (25%, 60min), (50%, 120min), (100%, 0)]
    current_stage_index: int = 0
    status: str = "in_progress"

@dataclass
class CanaryStage:
    canary_traffic_pct: float
    observation_minutes: int
    started_at: datetime | None = None

class CanaryController:
    def __init__(self, router: ExperimentRouter, analyzer: MetricsAnalyzer):
        self.router = router
        self.analyzer = analyzer

    async def advance_canary(self, deployment: CanaryDeployment):
        """Вызывается по расписанию для проверки и продвижения canary."""
        current_stage = deployment.stages[deployment.current_stage_index]

        # Проверяем что наблюдение завершено
        if not current_stage.started_at:
            current_stage.started_at = datetime.utcnow()
            return

        elapsed = (datetime.utcnow() - current_stage.started_at).total_seconds() / 60
        if elapsed < current_stage.observation_minutes:
            return  # ещё наблюдаем

        # Анализ метрик за период наблюдения
        health = await self.analyzer.compare_versions(
            deployment.agent_name,
            deployment.stable_version,
            deployment.canary_version,
            since=current_stage.started_at
        )

        if health.canary_is_unhealthy:
            await self.rollback(deployment, reason=health.degradation_reason)
            return

        # Переходим к следующей стадии
        next_index = deployment.current_stage_index + 1
        if next_index >= len(deployment.stages):
            await self.complete_rollout(deployment)
        else:
            deployment.current_stage_index = next_index
            next_stage = deployment.stages[next_index]
            await self.router.update_traffic_split(
                deployment.agent_name,
                stable_pct=100 - next_stage.canary_traffic_pct,
                canary_pct=next_stage.canary_traffic_pct,
                canary_version=deployment.canary_version
            )
            logger.info(f"Canary advanced to {next_stage.canary_traffic_pct}% for {deployment.agent_name}")

    async def rollback(self, deployment: CanaryDeployment, reason: str):
        await self.router.update_traffic_split(
            deployment.agent_name, stable_pct=100, canary_pct=0,
            canary_version=deployment.canary_version
        )
        deployment.status = "rolled_back"
        await notify_team(f"Canary rollback for {deployment.agent_name}: {reason}")
        logger.error(f"Canary rolled back: {deployment.agent_name} v{deployment.canary_version} → v{deployment.stable_version}")

Метрики для canary health check

class CanaryHealthChecker:
    THRESHOLDS = {
        "error_rate": {"max_absolute": 0.05, "max_relative_increase": 2.0},
        "p99_latency_ms": {"max_relative_increase": 1.5},
        "task_success_rate": {"min_absolute": 0.90, "max_relative_decrease": 0.1},
        "quality_score": {"max_relative_decrease": 0.05},
    }

    def is_healthy(self, stable_metrics: dict, canary_metrics: dict) -> HealthCheckResult:
        issues = []
        for metric, thresholds in self.THRESHOLDS.items():
            stable_val = stable_metrics.get(metric, 0)
            canary_val = canary_metrics.get(metric, 0)

            if "max_absolute" in thresholds and canary_val > thresholds["max_absolute"]:
                issues.append(f"{metric} too high: {canary_val:.3f} > {thresholds['max_absolute']}")

            if stable_val > 0 and "max_relative_increase" in thresholds:
                relative = canary_val / stable_val
                if relative > thresholds["max_relative_increase"]:
                    issues.append(f"{metric} increased {relative:.1f}x vs stable")

        return HealthCheckResult(is_healthy=len(issues) == 0, issues=issues)

Интеграция с Kubernetes

# Flagger (progressive delivery controller) для K8s
apiVersion: flagger.app/v1beta1
kind: Canary
metadata:
  name: vllm-agent
  namespace: ai-serving
spec:
  targetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: vllm-agent
  progressDeadlineSeconds: 3600
  service:
    port: 8000
  analysis:
    interval: 5m
    threshold: 5              # max failures before rollback
    maxWeight: 100
    stepWeight: 10            # +10% каждые 5 минут
    metrics:
      - name: request-success-rate
        thresholdRange:
          min: 99
        interval: 1m
      - name: request-duration
        thresholdRange:
          max: 5000
        interval: 1m