11장: 운영, 모니터링, 스케일링 전략

# Qdrant: 컬렉션 생성 시 샤드 수 지정
client.create_collection(
    collection_name="large_dataset",
    vectors_config=VectorParams(size=1536, distance=Distance.COSINE),
    shard_number=6,           # 6개 샤드
    replication_factor=2,     # 각 샤드 2개 레플리카
)

# Weaviate: 자동 샤딩 설정
import weaviate.classes.config as wc
 
client.collections.create(
    name="LargeDataset",
    sharding_config=wc.Configure.sharding(
        desired_count=4,          # 4개 샤드
        virtual_per_physical=128  # 가상 샤드 수
    ),
    replication_config=wc.Configure.replication(
        factor=3  # 3개 레플리카
    ),
    # ... 나머지 설정
)

from qdrant_client.models import WriteOrdering
 
# 쓰기 일관성 설정
client.upsert(
    collection_name="critical_data",
    points=[...],
    ordering=WriteOrdering.STRONG  # 모든 레플리카에 쓰기 완료 후 반환
)
 
# 읽기 일관성 설정
results = client.query_points(
    collection_name="critical_data",
    query=query_vector,
    consistency="majority",  # 과반수 레플리카 확인
    limit=10
)

# Qdrant: 스냅샷 생성
snapshot = client.create_snapshot(collection_name="production_data")
print(f"스냅샷 생성: {snapshot.name}")
 
# 스냅샷 목록 조회
snapshots = client.list_snapshots(collection_name="production_data")
 
# 스냅샷에서 복구
client.recover_snapshot(
    collection_name="production_data",
    location=f"http://qdrant:6333/collections/production_data/snapshots/{snapshot.name}"
)

# 무중단 인덱스 리빌드 전략 (개념)
 
# 1. 새 컬렉션 생성 (최적화된 파라미터)
client.create_collection(
    collection_name="articles_v2",
    vectors_config=VectorParams(
        size=1536,
        distance=Distance.COSINE,
        hnsw_config=HnswConfigDiff(m=32, ef_construct=256)
    )
)
 
# 2. 기존 데이터 마이그레이션
# (배치 단위로 데이터 복사)
offset = None
while True:
    points, offset = client.scroll(
        collection_name="articles_v1",
        limit=1000,
        offset=offset
    )
    if not points:
        break
    client.upsert(collection_name="articles_v2", points=points)
 
# 3. 알리아스 전환 (무중단)
client.update_collection_aliases(
    change_aliases_operations=[
        {"create_alias": {"collection_name": "articles_v2", "alias_name": "articles"}},
        {"delete_alias": {"alias_name": "articles"}}  # 구 알리아스 제거
    ]
)

# Prometheus 스크래핑 설정 (Qdrant)
scrape_configs:
  - job_name: 'qdrant'
    metrics_path: '/metrics'
    static_configs:
      - targets: ['qdrant:6333']
    scrape_interval: 15s

# 애플리케이션 레벨 커스텀 메트릭
import time
from prometheus_client import Histogram, Counter
 
SEARCH_LATENCY = Histogram(
    'vector_search_latency_seconds',
    'Vector search latency',
    buckets=[0.005, 0.01, 0.025, 0.05, 0.1, 0.25, 0.5, 1.0]
)
 
SEARCH_RECALL = Histogram(
    'vector_search_recall',
    'Estimated search recall',
    buckets=[0.8, 0.85, 0.9, 0.95, 0.98, 0.99, 1.0]
)
 
SEARCH_ERRORS = Counter(
    'vector_search_errors_total',
    'Total search errors',
    ['error_type']
)
 
def monitored_search(client, collection, query_vector, top_k):
    """모니터링이 적용된 검색 함수"""
    start = time.perf_counter()
    try:
        results = client.query_points(
            collection_name=collection,
            query=query_vector,
            limit=top_k
        )
        latency = time.perf_counter() - start
        SEARCH_LATENCY.observe(latency)
        return results
    except Exception as e:
        SEARCH_ERRORS.labels(error_type=type(e).__name__).inc()
        raise

# Qdrant -> Pinecone 마이그레이션 예시
from qdrant_client import QdrantClient
from pinecone import Pinecone
 
qdrant = QdrantClient(url="http://qdrant:6333")
pc = Pinecone(api_key="PINECONE_KEY")
pine_index = pc.Index("target-index")
 
# 배치 마이그레이션
batch_size = 100
offset = None
 
while True:
    points, offset = qdrant.scroll(
        collection_name="source",
        limit=batch_size,
        offset=offset,
        with_vectors=True,
        with_payload=True
    )
 
    if not points:
        break
 
    # Pinecone 형식으로 변환
    vectors = [
        {
            "id": str(point.id),
            "values": point.vector,
            "metadata": point.payload
        }
        for point in points
    ]
 
    pine_index.upsert(vectors=vectors)
 
print("마이그레이션 완료")