Trace 조회와 프로덕션 모니터링

import mlflow

# 특정 실험의 최근 트레이스 조회
traces = mlflow.search_traces(
    experiment_ids=["12345"],
    max_results=10,
    order_by=["timestamp DESC"]
)

# 특정 트레이스 상세 조회
trace = mlflow.get_trace(trace_id="abc-123-def")
for span in trace.data.spans:
    print(f"  {span.name}: {span.status} ({span.end_time - span.start_time}ms)")

-- 최근 실패한 트레이스 조회
SELECT
    trace_id,
    timestamp,
    status,
    total_tokens,
    latency_ms
FROM system.mlflow.traces
WHERE status = 'ERROR'
    AND timestamp >= CURRENT_DATE() - INTERVAL 1 DAY
ORDER BY timestamp DESC;

-- 토큰 사용량 추이
SELECT
    DATE(timestamp) AS day,
    COUNT(*) AS trace_count,
    SUM(total_tokens) AS total_tokens,
    AVG(latency_ms) AS avg_latency
FROM system.mlflow.traces
GROUP BY DATE(timestamp)
ORDER BY day DESC;

# Trace에 태그 추가 (필터링/검색에 활용)
trace = mlflow.get_last_active_trace()
mlflow.set_trace_tag("user_id", "user-12345")
mlflow.set_trace_tag("session_id", "session-abc")
mlflow.set_trace_tag("environment", "production")
mlflow.set_trace_tag("model_version", "v2.1")