모니터링 핵심 메트릭과 Inference Table

-- system 테이블에서 Percentile 메트릭 계산
SELECT
    endpoint_name,
    DATE(request_time) AS day,
    COUNT(*) AS total_requests,
    ROUND(PERCENTILE_CONT(0.50) WITHIN GROUP (ORDER BY execution_time_ms), 1) AS p50_ms,
    ROUND(PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY execution_time_ms), 1) AS p95_ms,
    ROUND(PERCENTILE_CONT(0.99) WITHIN GROUP (ORDER BY execution_time_ms), 1) AS p99_ms,
    ROUND(PERCENTILE_CONT(0.999) WITHIN GROUP (ORDER BY execution_time_ms), 1) AS p999_ms,
    ROUND(MAX(execution_time_ms), 1) AS max_ms
FROM system.serving.served_model_requests
WHERE DATE(request_time) >= CURRENT_DATE() - INTERVAL 7 DAYS
GROUP BY endpoint_name, DATE(request_time)
ORDER BY day DESC, endpoint_name;

-- 시간대별 평균 RPS(Requests Per Second) 계산
SELECT
    endpoint_name,
    DATE_TRUNC('hour', request_time) AS hour_bucket,
    COUNT(*) AS total_requests,
    -- 1시간(3600초) 기준 평균 RPS
    ROUND(COUNT(*) / 3600.0, 2) AS avg_rps
FROM system.serving.served_model_requests
WHERE DATE(request_time) >= CURRENT_DATE() - INTERVAL 1 DAYS
GROUP BY endpoint_name, DATE_TRUNC('hour', request_time)
ORDER BY hour_bucket DESC;

# 엔드포인트 생성 시 Inference Table 활성화
from databricks.sdk import WorkspaceClient
from databricks.sdk.service.serving import (
    EndpointCoreConfigInput,
    ServedEntityInput,
    AutoCaptureConfigInput,
)

w = WorkspaceClient()

auto_capture = AutoCaptureConfigInput(
    enabled=True,
    catalog_name="ml_prod",
    schema_name="monitoring",
    table_name_prefix="fraud_model"
)

w.serving_endpoints.create(
    name="fraud-detector",
    config=EndpointCoreConfigInput(
        served_entities=[
            ServedEntityInput(
                entity_name="ml_prod.models.fraud_detector",
                entity_version="3",
                workload_size="Small",
                scale_to_zero_enabled=True,
            )
        ],
        auto_capture_config=auto_capture,
    ),
)
# → ml_prod.monitoring.fraud_model_payload 테이블이 자동 생성됩니다

-- Inference Table 스키마 확인
DESCRIBE TABLE ml_prod.monitoring.fraud_model_payload;

-- 주요 컬럼:
-- date              : DATE        (파티션 키)
-- timestamp_ms      : BIGINT      (요청 시각, 밀리초)
-- status_code       : INT         (HTTP 상태 코드)
-- execution_time_ms : DOUBLE      (추론 소요 시간)
-- request           : STRING      (요청 JSON)
-- response          : STRING      (응답 JSON)
-- request_metadata  : MAP<STRING,STRING> (추적 메타데이터)
-- sampling_fraction : DOUBLE      (샘플링 비율)

# 샘플링 비율 설정 예시 (엔드포인트 업데이트 시)
auto_capture = AutoCaptureConfigInput(
    enabled=True,
    catalog_name="ml_prod",
    schema_name="monitoring",
    table_name_prefix="fraud_model",
    # 10% 샘플링: 고트래픽 환경에서 비용 절감
    # 기본값은 100% (sampling_fraction 미설정 시)
)

-- Inference Table에 데이터 보존 정책 설정 (90일)
ALTER TABLE ml_prod.monitoring.fraud_model_payload
SET TBLPROPERTIES (
  'delta.deletedFileRetentionDuration' = 'interval 7 days',
  'delta.logRetentionDuration' = 'interval 90 days'
);

-- 90일 이전 데이터 정기 삭제 (Lakeflow Jobs으로 예약)
DELETE FROM ml_prod.monitoring.fraud_model_payload
WHERE date < CURRENT_DATE() - INTERVAL 90 DAYS;

-- VACUUM으로 물리적 파일 정리
VACUUM ml_prod.monitoring.fraud_model_payload RETAIN 168 HOURS;