[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6FGWB41YRPVTYX7T2VRBHKS":3,"skill-same:01M6FGWB41YRPVTYX7T2VRBHKS":97},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":28,"publishedAt":49,"linkOnly":50,"markdownContent":51,"markdownBody":52,"frontMatter":53,"refName":56,"commitSha":57,"skillDirectory":58,"forksCount":59,"repositoryHtmlUrl":60,"repositoryHomepageUrl":61,"ownerHtmlUrl":62,"ownerBlogUrl":6,"ownerDescription":63,"updatedAt":64,"install":65,"resources":78},"01M6FGWB41YRPVTYX7T2VRBHKS","recommendation-system",null,"Deploy production recommendation systems with feature stores, caching, A\u002FB testing. Use for personalization APIs, low latency serving, or encountering cache invalidation, experiment tracking, quality monitoring issues.","ComeOnOliver\u002Fskillshub","ComeOnOliver","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F43215937?v=4","01M6E5JNFECJAAVFE0W7SMWWD2","Haoran Shu","01M6E5JNNDQS7AJZDYFQ8TPZCK",67,"MIT","https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fsecondsky\u002Fclaude-skills\u002Frecommendation-system\u002FSKILL.md",[18,23],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000008","测试与质量","SCENARIO","使用场景",{"categoryNo":24,"categoryName":25,"taxonomyType":26,"taxonomyTypeDesc":27},"01JBSEEDCAT000000000000030","开发工程师","PROFESSION","适用职业",[29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48],"recommendation system","recommendation","system","deploy production","Experiment Tracking","ctr","recommendations","caching","time","personalization","conversion","testing","sampling","experiment","coverage","model","api","monitoring","metrics","issues","2026-11-10 10:12:30",false,"---\nname: recommendation-system\ndescription: Deploy production recommendation systems with feature stores, caching, A\u002FB testing. Use for personalization APIs, low latency serving, or encountering cache invalidation, experiment tracking, quality monitoring issues.\nlicense: MIT\nmetadata:\n  keywords: \"recommendation system, personalization, feature store, model serving, caching strategy, Redis, A\u002FB testing, Thompson sampling, recommendation metrics, CTR, conversion rate, catalog coverage, diversity, Prometheus monitoring, recommendation API, real-time recommendations, collaborative filtering integration, production recommendations, experiment tracking\"\n---\n\n# Recommendation System\n\nProduction-ready architecture for scalable recommendation systems with feature stores, multi-tier caching, A\u002FB testing, and comprehensive monitoring.\n\n## When to Use This Skill\n\nLoad this skill when:\n- **Building Recommendation APIs**: Serving personalized recommendations at scale\n- **Implementing Caching**: Multi-tier caching for sub-millisecond latency\n- **Running A\u002FB Tests**: Experimenting with recommendation algorithms\n- **Monitoring Quality**: Tracking CTR, conversion, diversity, coverage\n- **Optimizing Performance**: Reducing latency, increasing throughput\n- **Feature Engineering**: Managing user\u002Fitem features with feature stores\n\n## Quick Start: Recommendation API in 5 Steps\n\n```bash\n# 1. Install dependencies\npip install fastapi==0.109.0 redis==5.0.0 prometheus-client==0.19.0\n\n# 2. Start Redis (for caching and feature store)\ndocker run -d -p 6379:6379 redis:alpine\n\n# 3. Create recommendation service: app.py\ncat > app.py \u003C\u003C 'EOF'\nfrom fastapi import FastAPI\nfrom pydantic import BaseModel\nfrom typing import List\nimport redis\nimport json\n\napp = FastAPI()\ncache = redis.Redis(host='localhost', port=6379, decode_responses=True)\n\nclass RecommendationResponse(BaseModel):\n    user_id: str\n    items: List[str]\n    cached: bool\n\n@app.post(\"\u002Frecommendations\", response_model=RecommendationResponse)\nasync def get_recommendations(user_id: str, n: int = 10):\n    # Check cache\n    cache_key = f\"recs:{user_id}:{n}\"\n    cached = cache.get(cache_key)\n\n    if cached:\n        return RecommendationResponse(\n            user_id=user_id,\n            items=json.loads(cached),\n            cached=True\n        )\n\n    # Generate recommendations (simplified)\n    items = [f\"item_{i}\" for i in range(n)]\n\n    # Cache for 5 minutes\n    cache.setex(cache_key, 300, json.dumps(items))\n\n    return RecommendationResponse(\n        user_id=user_id,\n        items=items,\n        cached=False\n    )\n\n@app.get(\"\u002Fhealth\")\nasync def health():\n    return {\"status\": \"healthy\"}\nEOF\n\n# 4. Run API\nuvicorn app:app --host 0.0.0.0 --port 8000\n\n# 5. Test\ncurl -X POST \"http:\u002F\u002Flocalhost:8000\u002Frecommendations?user_id=user_123&n=10\"\n```\n\n**Result**: Working recommendation API with caching in under 5 minutes.\n\n## System Architecture\n\n```\n┌─────────────┐     ┌─────────────┐     ┌─────────────┐\n│ User Events │────▶│ Feature     │────▶│ Model       │\n│ (clicks,    │     │ Store       │     │ Serving     │\n│  purchases) │     │ (Redis)     │     │             │\n└─────────────┘     └─────────────┘     └─────────────┘\n                           │                    │\n                           ▼                    ▼\n                    ┌─────────────┐     ┌─────────────┐\n                    │ Training    │     │ API         │\n                    │ Pipeline    │     │ (FastAPI)   │\n                    └─────────────┘     └─────────────┘\n                                               │\n                                               ▼\n                                        ┌─────────────┐\n                                        │ Monitoring  │\n                                        │ (Prometheus)│\n                                        └─────────────┘\n```\n\n## Core Components\n\n### 1. Feature Store\n\nCentralized storage for user and item features:\n\n```python\nimport redis\nimport json\n\nclass FeatureStore:\n    \"\"\"Fast feature access with Redis caching.\"\"\"\n\n    def __init__(self, redis_client):\n        self.redis = redis_client\n        self.ttl = 3600  # 1 hour\n\n    def get_user_features(self, user_id: str) -> dict:\n        cache_key = f\"user_features:{user_id}\"\n        cached = self.redis.get(cache_key)\n\n        if cached:\n            return json.loads(cached)\n\n        # Fetch from database\n        features = fetch_from_db(user_id)\n\n        # Cache\n        self.redis.setex(cache_key, self.ttl, json.dumps(features))\n        return features\n```\n\n### 2. Model Serving\n\nServe multiple models for A\u002FB testing:\n\n```python\nclass ModelServing:\n    \"\"\"Serve multiple recommendation models.\"\"\"\n\n    def __init__(self):\n        self.models = {}\n\n    def register_model(self, name: str, model, is_default: bool = False):\n        self.models[name] = model\n        if is_default:\n            self.default_model = name\n\n    def predict(self, user_features: dict, item_features: list, model_name: str = None):\n        model = self.models.get(model_name or self.default_model)\n        return model.predict(user_features, item_features)\n```\n\n### 3. Caching Layer\n\nMulti-tier caching for low latency:\n\n```python\nclass TieredCache:\n    \"\"\"L1 (memory) -> L2 (Redis) -> L3 (database).\"\"\"\n\n    def __init__(self, redis_client):\n        self.l1_cache = {}  # In-memory\n        self.redis = redis_client  # L2\n\n    def get(self, key: str):\n        # L1: In-memory (fastest)\n        if key in self.l1_cache:\n            return self.l1_cache[key]\n\n        # L2: Redis\n        cached = self.redis.get(key)\n        if cached:\n            value = json.loads(cached)\n            self.l1_cache[key] = value  # Promote to L1\n            return value\n\n        # L3: Miss (fetch from database)\n        return None\n```\n\n## Key Metrics\n\n| Metric | Description | Target |\n|--------|-------------|--------|\n| **CTR** | Click-through rate | >5% |\n| **Conversion Rate** | Purchases from recs | >2% |\n| **P95 Latency** | 95th percentile response time | \u003C200ms |\n| **Cache Hit Rate** | % served from cache | >80% |\n| **Coverage** | % of catalog recommended | >50% |\n| **Diversity** | Variety in recommendations | >0.7 |\n\n## Known Issues Prevention\n\n### 1. Cold Start for New Users\n**Problem**: No recommendations for users without history, poor initial experience.\n\n**Solution**: Use popularity-based fallback:\n```python\ndef get_recommendations(user_id: str, n: int = 10):\n    user_features = feature_store.get_user_features(user_id)\n\n    # Check if new user (no purchase history)\n    if user_features.get('total_purchases', 0) == 0:\n        # Fallback to popular items\n        return get_popular_items(n)\n\n    # Personalized recommendations\n    return generate_personalized_recs(user_id, n)\n```\n\n### 2. Cache Invalidation on User Actions\n**Problem**: User makes purchase, cache still shows purchased item in recommendations.\n\n**Solution**: Invalidate cache on relevant actions:\n```python\nINVALIDATING_ACTIONS = {'purchase', 'rating', 'add_to_cart'}\n\ndef on_user_action(user_id: str, action: str):\n    if action in INVALIDATING_ACTIONS:\n        cache_key = f\"recs:{user_id}:*\"\n        redis_client.delete(cache_key)\n        logger.info(f\"Invalidated cache for {user_id} due to {action}\")\n```\n\n### 3. Thundering Herd on Cache Expiry\n**Problem**: Many users' caches expire simultaneously, overload database\u002Fmodel.\n\n**Solution**: Add random jitter to TTL:\n```python\nimport random\n\ndef set_cache(key: str, value: dict, base_ttl: int = 300):\n    # Add ±10% jitter\n    jitter = random.uniform(-0.1, 0.1) * base_ttl\n    ttl = int(base_ttl + jitter)\n    redis_client.setex(key, ttl, json.dumps(value))\n```\n\n### 4. Poor Diversity = Filter Bubble\n**Problem**: Recommendations too similar, users only see same category.\n\n**Solution**: Implement diversity constraint:\n```python\ndef rank_with_diversity(items: list, scores: list, n: int = 10):\n    selected = []\n    category_counts = {}\n\n    for item, score in sorted(zip(items, scores), key=lambda x: -x[1]):\n        category = item['category']\n\n        # Limit 3 items per category\n        if category_counts.get(category, 0) >= 3:\n            continue\n\n        selected.append(item)\n        category_counts[category] = category_counts.get(category, 0) + 1\n\n        if len(selected) >= n:\n            break\n\n    return selected\n```\n\n### 5. No Monitoring = Silent Degradation\n**Problem**: Recommendation quality drops, nobody notices until users complain.\n\n**Solution**: Continuous monitoring with alerts:\n```python\nfrom prometheus_client import Counter, Histogram\n\nrecommendation_clicks = Counter('recommendation_clicks_total')\nrecommendation_latency = Histogram('recommendation_latency_seconds')\n\n@app.post(\"\u002Frecommendations\")\nasync def get_recommendations(user_id: str):\n    start = time.time()\n\n    recs = generate_recs(user_id)\n\n    latency = time.time() - start\n    recommendation_latency.observe(latency)\n\n    return recs\n\n@app.post(\"\u002Ftrack\u002Fclick\")\nasync def track_click(user_id: str, item_id: str):\n    recommendation_clicks.inc()\n    # Alert if CTR drops below 3%\n```\n\n### 6. Stale Features = Outdated Recommendations\n**Problem**: User preferences change but features don't update, recommendations irrelevant.\n\n**Solution**: Set appropriate TTLs and update triggers:\n```python\nclass FeatureStore:\n    def __init__(self, redis_client):\n        self.redis = redis_client\n        # Shorter TTL for frequently changing features\n        self.user_ttl = 300  # 5 minutes\n        self.item_ttl = 3600  # 1 hour\n\n    def update_on_event(self, user_id: str, event: str):\n        # Invalidate on important events\n        if event in ['purchase', 'rating']:\n            self.redis.delete(f\"user_features:{user_id}\")\n            logger.info(f\"Refreshed features for {user_id}\")\n```\n\n### 7. A\u002FB Test Sample Size Too Small\n**Problem**: Declare winner too early, results not statistically significant.\n\n**Solution**: Calculate required sample size first:\n```python\ndef calculate_sample_size(\n    baseline_rate: float,\n    min_detectable_effect: float,\n    alpha: float = 0.05,\n    power: float = 0.8\n) -> int:\n    \"\"\"Calculate required sample size per variant.\"\"\"\n    from scipy import stats\n\n    z_alpha = stats.norm.ppf(1 - alpha\u002F2)\n    z_beta = stats.norm.ppf(power)\n\n    p1 = baseline_rate\n    p2 = baseline_rate * (1 + min_detectable_effect)\n    p_avg = (p1 + p2) \u002F 2\n\n    n = (\n        (z_alpha + z_beta)**2 * 2 * p_avg * (1 - p_avg) \u002F\n        (p2 - p1)**2\n    )\n\n    return int(n)\n\n# Example: detect 10% lift with baseline CTR=5%\nn_required = calculate_sample_size(\n    baseline_rate=0.05,\n    min_detectable_effect=0.10\n)\nprint(f\"Required sample size: {n_required} per variant\")\n# Wait until both variants reach this size before concluding\n```\n\n## When to Load References\n\nLoad reference files for detailed production implementations:\n\n- **Production Architecture**: Load `references\u002Fproduction-architecture.md` for complete FeatureStore, ModelServing, and RecommendationService implementations with batch fetching, caching integration, and FastAPI deployment patterns.\n\n- **Caching Strategies**: Load `references\u002Fcaching-strategies.md` when implementing multi-tier caching (L1\u002FL2\u002FL3), cache warming, invalidation strategies, probabilistic refresh, or thundering herd prevention.\n\n- **A\u002FB Testing Framework**: Load `references\u002Fab-testing-framework.md` for deterministic variant assignment, Thompson sampling (multi-armed bandits), Bayesian and frequentist significance testing, and experiment tracking.\n\n- **Monitoring & Alerting**: Load `references\u002Fmonitoring-alerting.md` for Prometheus metrics integration, dashboard endpoints, alert rules, and quality monitoring (diversity, coverage).\n\n## Best Practices\n\n1. **Feature Precomputation**: Compute features offline, serve from cache\n2. **Batch Fetching**: Use Redis MGET for multiple users\u002Fitems\n3. **Cache Aggressively**: 5-15 minute TTL for user recommendations\n4. **Fail Gracefully**: Return popular items if personalization fails\n5. **Monitor Everything**: Track CTR, latency, diversity, coverage\n6. **A\u002FB Test Continuously**: Always be experimenting with new algorithms\n7. **Diversity Constraint**: Ensure varied recommendations\n8. **Explain Recommendations**: Provide reasons (\"Highly rated\", \"Popular\")\n\n## Common Patterns\n\n### Recommendation Service\n```python\nclass RecommendationService:\n    def __init__(self, feature_store, model_serving, cache):\n        self.feature_store = feature_store\n        self.model_serving = model_serving\n        self.cache = cache\n\n    def get_recommendations(self, user_id: str, n: int = 10):\n        # 1. Check cache\n        cached = self.cache.get(f\"recs:{user_id}:{n}\")\n        if cached:\n            return cached\n\n        # 2. Get features\n        user_features = self.feature_store.get_user_features(user_id)\n        candidates = self.get_candidates(user_id)\n\n        # 3. Score candidates\n        scores = self.model_serving.predict(user_features, candidates)\n\n        # 4. Rank with diversity\n        recommendations = self.rank_with_diversity(candidates, scores, n)\n\n        # 5. Cache\n        self.cache.set(f\"recs:{user_id}:{n}\", recommendations, ttl=300)\n\n        return recommendations\n```\n\n### A\u002FB Testing\n```python\ndef assign_variant(user_id: str, experiment_id: str) -> str:\n    \"\"\"Deterministic assignment - same user always gets same variant.\"\"\"\n    import hashlib\n\n    hash_input = f\"{user_id}:{experiment_id}\"\n    hash_value = int(hashlib.md5(hash_input.encode()).hexdigest(), 16)\n\n    # 50\u002F50 split\n    return 'control' if hash_value % 2 == 0 else 'treatment'\n\n# Usage\nvariant = assign_variant('user_123', 'rec_algo_v2')\nmodel_name = 'main' if variant == 'control' else 'experimental'\nrecs = get_recommendations(user_id, model_name=model_name)\n```\n\n### Monitoring\n```python\nfrom prometheus_client import Counter, Histogram\n\nrequests_total = Counter('recommendation_requests_total', ['status'])\nlatency_seconds = Histogram('recommendation_latency_seconds')\n\n@app.post(\"\u002Frecommendations\")\nasync def get_recommendations(user_id: str):\n    with latency_seconds.time():\n        try:\n            recs = generate_recs(user_id)\n            requests_total.labels(status='success').inc()\n            return recs\n        except Exception as e:\n            requests_total.labels(status='error').inc()\n            raise\n```\n","# Recommendation System\n\nProduction-ready architecture for scalable recommendation systems with feature stores, multi-tier caching, A\u002FB testing, and comprehensive monitoring.\n\n## When to Use This Skill\n\nLoad this skill when:\n- **Building Recommendation APIs**: Serving personalized recommendations at scale\n- **Implementing Caching**: Multi-tier caching for sub-millisecond latency\n- **Running A\u002FB Tests**: Experimenting with recommendation algorithms\n- **Monitoring Quality**: Tracking CTR, conversion, diversity, coverage\n- **Optimizing Performance**: Reducing latency, increasing throughput\n- **Feature Engineering**: Managing user\u002Fitem features with feature stores\n\n## Quick Start: Recommendation API in 5 Steps\n\n```bash\n# 1. Install dependencies\npip install fastapi==0.109.0 redis==5.0.0 prometheus-client==0.19.0\n\n# 2. Start Redis (for caching and feature store)\ndocker run -d -p 6379:6379 redis:alpine\n\n# 3. Create recommendation service: app.py\ncat > app.py \u003C\u003C 'EOF'\nfrom fastapi import FastAPI\nfrom pydantic import BaseModel\nfrom typing import List\nimport redis\nimport json\n\napp = FastAPI()\ncache = redis.Redis(host='localhost', port=6379, decode_responses=True)\n\nclass RecommendationResponse(BaseModel):\n    user_id: str\n    items: List[str]\n    cached: bool\n\n@app.post(\"\u002Frecommendations\", response_model=RecommendationResponse)\nasync def get_recommendations(user_id: str, n: int = 10):\n    # Check cache\n    cache_key = f\"recs:{user_id}:{n}\"\n    cached = cache.get(cache_key)\n\n    if cached:\n        return RecommendationResponse(\n            user_id=user_id,\n            items=json.loads(cached),\n            cached=True\n        )\n\n    # Generate recommendations (simplified)\n    items = [f\"item_{i}\" for i in range(n)]\n\n    # Cache for 5 minutes\n    cache.setex(cache_key, 300, json.dumps(items))\n\n    return RecommendationResponse(\n        user_id=user_id,\n        items=items,\n        cached=False\n    )\n\n@app.get(\"\u002Fhealth\")\nasync def health():\n    return {\"status\": \"healthy\"}\nEOF\n\n# 4. Run API\nuvicorn app:app --host 0.0.0.0 --port 8000\n\n# 5. Test\ncurl -X POST \"http:\u002F\u002Flocalhost:8000\u002Frecommendations?user_id=user_123&n=10\"\n```\n\n**Result**: Working recommendation API with caching in under 5 minutes.\n\n## System Architecture\n\n```\n┌─────────────┐     ┌─────────────┐     ┌─────────────┐\n│ User Events │────▶│ Feature     │────▶│ Model       │\n│ (clicks,    │     │ Store       │     │ Serving     │\n│  purchases) │     │ (Redis)     │     │             │\n└─────────────┘     └─────────────┘     └─────────────┘\n                           │                    │\n                           ▼                    ▼\n                    ┌─────────────┐     ┌─────────────┐\n                    │ Training    │     │ API         │\n                    │ Pipeline    │     │ (FastAPI)   │\n                    └─────────────┘     └─────────────┘\n                                               │\n                                               ▼\n                                        ┌─────────────┐\n                                        │ Monitoring  │\n                                        │ (Prometheus)│\n                                        └─────────────┘\n```\n\n## Core Components\n\n### 1. Feature Store\n\nCentralized storage for user and item features:\n\n```python\nimport redis\nimport json\n\nclass FeatureStore:\n    \"\"\"Fast feature access with Redis caching.\"\"\"\n\n    def __init__(self, redis_client):\n        self.redis = redis_client\n        self.ttl = 3600  # 1 hour\n\n    def get_user_features(self, user_id: str) -> dict:\n        cache_key = f\"user_features:{user_id}\"\n        cached = self.redis.get(cache_key)\n\n        if cached:\n            return json.loads(cached)\n\n        # Fetch from database\n        features = fetch_from_db(user_id)\n\n        # Cache\n        self.redis.setex(cache_key, self.ttl, json.dumps(features))\n        return features\n```\n\n### 2. Model Serving\n\nServe multiple models for A\u002FB testing:\n\n```python\nclass ModelServing:\n    \"\"\"Serve multiple recommendation models.\"\"\"\n\n    def __init__(self):\n        self.models = {}\n\n    def register_model(self, name: str, model, is_default: bool = False):\n        self.models[name] = model\n        if is_default:\n            self.default_model = name\n\n    def predict(self, user_features: dict, item_features: list, model_name: str = None):\n        model = self.models.get(model_name or self.default_model)\n        return model.predict(user_features, item_features)\n```\n\n### 3. Caching Layer\n\nMulti-tier caching for low latency:\n\n```python\nclass TieredCache:\n    \"\"\"L1 (memory) -> L2 (Redis) -> L3 (database).\"\"\"\n\n    def __init__(self, redis_client):\n        self.l1_cache = {}  # In-memory\n        self.redis = redis_client  # L2\n\n    def get(self, key: str):\n        # L1: In-memory (fastest)\n        if key in self.l1_cache:\n            return self.l1_cache[key]\n\n        # L2: Redis\n        cached = self.redis.get(key)\n        if cached:\n            value = json.loads(cached)\n            self.l1_cache[key] = value  # Promote to L1\n            return value\n\n        # L3: Miss (fetch from database)\n        return None\n```\n\n## Key Metrics\n\n| Metric | Description | Target |\n|--------|-------------|--------|\n| **CTR** | Click-through rate | >5% |\n| **Conversion Rate** | Purchases from recs | >2% |\n| **P95 Latency** | 95th percentile response time | \u003C200ms |\n| **Cache Hit Rate** | % served from cache | >80% |\n| **Coverage** | % of catalog recommended | >50% |\n| **Diversity** | Variety in recommendations | >0.7 |\n\n## Known Issues Prevention\n\n### 1. Cold Start for New Users\n**Problem**: No recommendations for users without history, poor initial experience.\n\n**Solution**: Use popularity-based fallback:\n```python\ndef get_recommendations(user_id: str, n: int = 10):\n    user_features = feature_store.get_user_features(user_id)\n\n    # Check if new user (no purchase history)\n    if user_features.get('total_purchases', 0) == 0:\n        # Fallback to popular items\n        return get_popular_items(n)\n\n    # Personalized recommendations\n    return generate_personalized_recs(user_id, n)\n```\n\n### 2. Cache Invalidation on User Actions\n**Problem**: User makes purchase, cache still shows purchased item in recommendations.\n\n**Solution**: Invalidate cache on relevant actions:\n```python\nINVALIDATING_ACTIONS = {'purchase', 'rating', 'add_to_cart'}\n\ndef on_user_action(user_id: str, action: str):\n    if action in INVALIDATING_ACTIONS:\n        cache_key = f\"recs:{user_id}:*\"\n        redis_client.delete(cache_key)\n        logger.info(f\"Invalidated cache for {user_id} due to {action}\")\n```\n\n### 3. Thundering Herd on Cache Expiry\n**Problem**: Many users' caches expire simultaneously, overload database\u002Fmodel.\n\n**Solution**: Add random jitter to TTL:\n```python\nimport random\n\ndef set_cache(key: str, value: dict, base_ttl: int = 300):\n    # Add ±10% jitter\n    jitter = random.uniform(-0.1, 0.1) * base_ttl\n    ttl = int(base_ttl + jitter)\n    redis_client.setex(key, ttl, json.dumps(value))\n```\n\n### 4. Poor Diversity = Filter Bubble\n**Problem**: Recommendations too similar, users only see same category.\n\n**Solution**: Implement diversity constraint:\n```python\ndef rank_with_diversity(items: list, scores: list, n: int = 10):\n    selected = []\n    category_counts = {}\n\n    for item, score in sorted(zip(items, scores), key=lambda x: -x[1]):\n        category = item['category']\n\n        # Limit 3 items per category\n        if category_counts.get(category, 0) >= 3:\n            continue\n\n        selected.append(item)\n        category_counts[category] = category_counts.get(category, 0) + 1\n\n        if len(selected) >= n:\n            break\n\n    return selected\n```\n\n### 5. No Monitoring = Silent Degradation\n**Problem**: Recommendation quality drops, nobody notices until users complain.\n\n**Solution**: Continuous monitoring with alerts:\n```python\nfrom prometheus_client import Counter, Histogram\n\nrecommendation_clicks = Counter('recommendation_clicks_total')\nrecommendation_latency = Histogram('recommendation_latency_seconds')\n\n@app.post(\"\u002Frecommendations\")\nasync def get_recommendations(user_id: str):\n    start = time.time()\n\n    recs = generate_recs(user_id)\n\n    latency = time.time() - start\n    recommendation_latency.observe(latency)\n\n    return recs\n\n@app.post(\"\u002Ftrack\u002Fclick\")\nasync def track_click(user_id: str, item_id: str):\n    recommendation_clicks.inc()\n    # Alert if CTR drops below 3%\n```\n\n### 6. Stale Features = Outdated Recommendations\n**Problem**: User preferences change but features don't update, recommendations irrelevant.\n\n**Solution**: Set appropriate TTLs and update triggers:\n```python\nclass FeatureStore:\n    def __init__(self, redis_client):\n        self.redis = redis_client\n        # Shorter TTL for frequently changing features\n        self.user_ttl = 300  # 5 minutes\n        self.item_ttl = 3600  # 1 hour\n\n    def update_on_event(self, user_id: str, event: str):\n        # Invalidate on important events\n        if event in ['purchase', 'rating']:\n            self.redis.delete(f\"user_features:{user_id}\")\n            logger.info(f\"Refreshed features for {user_id}\")\n```\n\n### 7. A\u002FB Test Sample Size Too Small\n**Problem**: Declare winner too early, results not statistically significant.\n\n**Solution**: Calculate required sample size first:\n```python\ndef calculate_sample_size(\n    baseline_rate: float,\n    min_detectable_effect: float,\n    alpha: float = 0.05,\n    power: float = 0.8\n) -> int:\n    \"\"\"Calculate required sample size per variant.\"\"\"\n    from scipy import stats\n\n    z_alpha = stats.norm.ppf(1 - alpha\u002F2)\n    z_beta = stats.norm.ppf(power)\n\n    p1 = baseline_rate\n    p2 = baseline_rate * (1 + min_detectable_effect)\n    p_avg = (p1 + p2) \u002F 2\n\n    n = (\n        (z_alpha + z_beta)**2 * 2 * p_avg * (1 - p_avg) \u002F\n        (p2 - p1)**2\n    )\n\n    return int(n)\n\n# Example: detect 10% lift with baseline CTR=5%\nn_required = calculate_sample_size(\n    baseline_rate=0.05,\n    min_detectable_effect=0.10\n)\nprint(f\"Required sample size: {n_required} per variant\")\n# Wait until both variants reach this size before concluding\n```\n\n## When to Load References\n\nLoad reference files for detailed production implementations:\n\n- **Production Architecture**: Load `references\u002Fproduction-architecture.md` for complete FeatureStore, ModelServing, and RecommendationService implementations with batch fetching, caching integration, and FastAPI deployment patterns.\n\n- **Caching Strategies**: Load `references\u002Fcaching-strategies.md` when implementing multi-tier caching (L1\u002FL2\u002FL3), cache warming, invalidation strategies, probabilistic refresh, or thundering herd prevention.\n\n- **A\u002FB Testing Framework**: Load `references\u002Fab-testing-framework.md` for deterministic variant assignment, Thompson sampling (multi-armed bandits), Bayesian and frequentist significance testing, and experiment tracking.\n\n- **Monitoring & Alerting**: Load `references\u002Fmonitoring-alerting.md` for Prometheus metrics integration, dashboard endpoints, alert rules, and quality monitoring (diversity, coverage).\n\n## Best Practices\n\n1. **Feature Precomputation**: Compute features offline, serve from cache\n2. **Batch Fetching**: Use Redis MGET for multiple users\u002Fitems\n3. **Cache Aggressively**: 5-15 minute TTL for user recommendations\n4. **Fail Gracefully**: Return popular items if personalization fails\n5. **Monitor Everything**: Track CTR, latency, diversity, coverage\n6. **A\u002FB Test Continuously**: Always be experimenting with new algorithms\n7. **Diversity Constraint**: Ensure varied recommendations\n8. **Explain Recommendations**: Provide reasons (\"Highly rated\", \"Popular\")\n\n## Common Patterns\n\n### Recommendation Service\n```python\nclass RecommendationService:\n    def __init__(self, feature_store, model_serving, cache):\n        self.feature_store = feature_store\n        self.model_serving = model_serving\n        self.cache = cache\n\n    def get_recommendations(self, user_id: str, n: int = 10):\n        # 1. Check cache\n        cached = self.cache.get(f\"recs:{user_id}:{n}\")\n        if cached:\n            return cached\n\n        # 2. Get features\n        user_features = self.feature_store.get_user_features(user_id)\n        candidates = self.get_candidates(user_id)\n\n        # 3. Score candidates\n        scores = self.model_serving.predict(user_features, candidates)\n\n        # 4. Rank with diversity\n        recommendations = self.rank_with_diversity(candidates, scores, n)\n\n        # 5. Cache\n        self.cache.set(f\"recs:{user_id}:{n}\", recommendations, ttl=300)\n\n        return recommendations\n```\n\n### A\u002FB Testing\n```python\ndef assign_variant(user_id: str, experiment_id: str) -> str:\n    \"\"\"Deterministic assignment - same user always gets same variant.\"\"\"\n    import hashlib\n\n    hash_input = f\"{user_id}:{experiment_id}\"\n    hash_value = int(hashlib.md5(hash_input.encode()).hexdigest(), 16)\n\n    # 50\u002F50 split\n    return 'control' if hash_value % 2 == 0 else 'treatment'\n\n# Usage\nvariant = assign_variant('user_123', 'rec_algo_v2')\nmodel_name = 'main' if variant == 'control' else 'experimental'\nrecs = get_recommendations(user_id, model_name=model_name)\n```\n\n### Monitoring\n```python\nfrom prometheus_client import Counter, Histogram\n\nrequests_total = Counter('recommendation_requests_total', ['status'])\nlatency_seconds = Histogram('recommendation_latency_seconds')\n\n@app.post(\"\u002Frecommendations\")\nasync def get_recommendations(user_id: str):\n    with latency_seconds.time():\n        try:\n            recs = generate_recs(user_id)\n            requests_total.labels(status='success').inc()\n            return recs\n        except Exception as e:\n            requests_total.labels(status='error').inc()\n            raise\n```\n",{"name":5,"license":15,"metadata":54,"description":7},{"keywords":55},"recommendation system, personalization, feature store, model serving, caching strategy, Redis, A\u002FB testing, Thompson sampling, recommendation metrics, CTR, conversion rate, catalog coverage, diversity, Prometheus monitoring, recommendation API, real-time recommendations, collaborative filtering integration, production recommendations, experiment tracking","main","def8531e65114c0fca8fb8551c1871ee0eed705c","skills\u002Fsecondsky\u002Fclaude-skills\u002Frecommendation-system",24,"https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub","https:\u002F\u002Fskillshub.wtf","https:\u002F\u002Fgithub.com\u002FComeOnOliver","AI & Blockchain Engineer\r\n","2026-06-24 22:07:19",{"source":8,"skillName":5,"commands":66,"prompt":76,"zipAvailable":77},[67,70,73],{"tool":68,"command":69},"npx","npx skills add ComeOnOliver\u002Fskillshub --skill recommendation-system",{"tool":71,"command":72},"bunx","bunx skills add ComeOnOliver\u002Fskillshub --skill recommendation-system",{"tool":74,"command":75},"pnpm","pnpm dlx skills add ComeOnOliver\u002Fskillshub --skill recommendation-system","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fsecondsky\u002Fclaude-skills\u002Frecommendation-system\u002FSKILL.md",true,[79,85,89,93],{"path":80,"resourceType":81,"resourceTypeDesc":82,"sizeBytes":83,"sourceHtmlUrl":84},"references\u002Fab-testing-framework.md","REFERENCE","参考文档",19615,"https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fsecondsky\u002Fclaude-skills\u002Frecommendation-system\u002Freferences\u002Fab-testing-framework.md",{"path":86,"resourceType":81,"resourceTypeDesc":82,"sizeBytes":87,"sourceHtmlUrl":88},"references\u002Fcaching-strategies.md",14517,"https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fsecondsky\u002Fclaude-skills\u002Frecommendation-system\u002Freferences\u002Fcaching-strategies.md",{"path":90,"resourceType":81,"resourceTypeDesc":82,"sizeBytes":91,"sourceHtmlUrl":92},"references\u002Fmonitoring-alerting.md",19142,"https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fsecondsky\u002Fclaude-skills\u002Frecommendation-system\u002Freferences\u002Fmonitoring-alerting.md",{"path":94,"resourceType":81,"resourceTypeDesc":82,"sizeBytes":95,"sourceHtmlUrl":96},"references\u002Fproduction-architecture.md",18450,"https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fsecondsky\u002Fclaude-skills\u002Frecommendation-system\u002Freferences\u002Fproduction-architecture.md",[98,109,119,128,134,141],{"skillNo":99,"name":100,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":101,"categories":102,"keywords":106,"publishedAt":49},"01M6FKJN13GMG4KY57J3QD8GY5","model-hierarchy-skill","https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fzscole\u002Fmodel-hierarchy-skill\u002Fmodel-hierarchy-skill\u002FSKILL.md",[103],{"categoryNo":104,"categoryName":105,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发",[44,107,108],"model hierarchy","model hierarchy skill",{"skillNo":110,"name":111,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":112,"categories":113,"keywords":115,"publishedAt":49},"01M6FKJJJ6GCAMR4PPPRD12M5X","zoom-mcp","https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fzoom\u002Fskills\u002Fzoom-mcp\u002FSKILL.md",[114],{"categoryNo":104,"categoryName":105,"taxonomyType":21,"taxonomyTypeDesc":22},[116,117,118],"MCP","zoom","zoom mcp",{"skillNo":120,"name":121,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":122,"categories":123,"keywords":124,"publishedAt":49},"01M6FKJE3KYT3KPEZMEEHYKA6K","zoom-apps-sdk","https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fzoom\u002Fskills\u002Fzoom-apps-sdk\u002FSKILL.md",[],[125,126,127,117],"sdk","zoom apps sdk","apps",{"skillNo":129,"name":130,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":131,"categories":132,"keywords":133,"publishedAt":49},"01M6FKJ9EDH9Q5VWDF8HAJRSP0","windows","https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fzoom\u002Fskills\u002Fwindows\u002FSKILL.md",[],[130],{"skillNo":135,"name":136,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":137,"categories":138,"keywords":140,"publishedAt":49},"01M6FKJ666BEEFC3AE9GJ72Q93","whiteboard","https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fzoom\u002Fskills\u002Fwhiteboard\u002FSKILL.md",[139],{"categoryNo":104,"categoryName":105,"taxonomyType":21,"taxonomyTypeDesc":22},[136],{"skillNo":142,"name":143,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":144,"categories":145,"keywords":146,"publishedAt":49},"01M6FKJ21WFCDGTJQ2Q91T9DVW","websockets","https:\u002F\u002Fgithub.com\u002FComeOnOliver\u002Fskillshub\u002Fblob\u002Fmain\u002Fskills\u002Fzoom\u002Fskills\u002Fwebsockets\u002FSKILL.md",[],[143]]