DataStore 프로파일러는 실행 시간을 측정하고 성능 병목을 파악하는 데 도움이 됩니다.
빠른 시작
from chdb import datastore as pd
from chdb.datastore.config import config, get_profiler
# 프로파일링 활성화
config.enable_profiling()
# 작업 실행
ds = pd.read_csv("large_data.csv")
result = (ds
.filter(ds['amount'] > 100)
.groupby('category')
.agg({'amount': 'sum'})
.sort('sum', ascending=False)
.head(10)
.to_df()
)
# 보고서 확인
profiler = get_profiler()
print(profiler.report())프로파일링 활성화
from chdb.datastore.config import config
# 프로파일링 활성화
config.enable_profiling()
# 프로파일링 비활성화
config.disable_profiling()
# 프로파일링 활성화 여부 확인
print(config.profiling_enabled) # True or False프로파일러 API
프로파일러 가져오는 방법
from chdb.datastore.config import get_profiler
profiler = get_profiler()report()
성능 보고서를 출력합니다.
profiler.report(min_duration_ms=0.1)매개변수:
| 매개변수 | 유형 | 기본값 | 설명 |
|---|---|---|---|
min_duration_ms |
float | 0.1 |
이 값 이상의 소요 시간을 가진 단계만 표시합니다 |
예시 출력:
======================================================================
EXECUTION PROFILE
======================================================================
45.79ms (100.0%) Total Execution
23.25ms ( 50.8%) Query Planning [ops_count=2]
22.29ms ( 48.7%) SQL Segment 1 [ops=2]
20.48ms ( 91.9%) SQL Execution
1.74ms ( 7.8%) Result to DataFrame
----------------------------------------------------------------------
TOTAL: 45.79ms
======================================================================보고서에는 다음이 표시됩니다:
- 각 단계의 소요 시간(밀리초)
- 상위 단계 시간/전체 시간 대비 비율
- 작업의 계층적 중첩 구조
- 각 단계의 메타데이터(예:
ops_count,ops)
step()
코드 블록의 실행 시간을 수동으로 측정합니다.
with profiler.step("custom_operation"):
# 여기에 코드를 작성하세요
expensive_operation()clear()
모든 프로파일링 데이터를 삭제합니다.
profiler.clear()summary()
단계 이름과 소요 시간(ms)의 매핑이 담긴 딕셔너리를 가져옵니다.
summary = profiler.summary()
for name, duration in summary.items():
print(f"{name}: {duration:.2f}ms")출력 예시:
Total Execution: 45.79ms
Total Execution.Cache Check: 0.00ms
Total Execution.Query Planning: 23.25ms
Total Execution.SQL Segment 1: 22.29ms
Total Execution.SQL Segment 1.SQL Execution: 20.48ms
Total Execution.SQL Segment 1.Result to DataFrame: 1.74ms보고서 살펴보기
단계 이름
| 단계 이름 | 설명 |
|---|---|
Total Execution |
전체 실행 시간 |
Query Planning |
쿼리 계획 수립에 소요된 시간 |
SQL Segment N |
SQL 세그먼트 N 실행 |
SQL Execution |
실제 SQL 쿼리 실행 |
Result to DataFrame |
결과를 pandas로 변환 |
Cache Check |
쿼리 캐시 확인 |
Cache Write |
결과를 캐시에 쓰기 |
소요 시간
- 계획 단계 (쿼리 계획): 일반적으로 빠릅니다
- 실행 단계 (SQL 실행): 실제 작업이 수행되는 단계입니다
- 전송 단계 (결과를 DataFrame으로): 데이터를 pandas로 변환하는 단계입니다
병목 지점 파악
======================================================================
EXECUTION PROFILE
======================================================================
200.50ms (100.0%) Total Execution
10.25ms ( 5.1%) Query Planning [ops_count=4]
190.00ms ( 94.8%) SQL Segment 1 [ops=4]
185.00ms ( 97.4%) SQL Execution <- 주요 병목 지점
5.00ms ( 2.6%) Result to DataFrame
----------------------------------------------------------------------
TOTAL: 200.50ms
======================================================================프로파일링 패턴
단일 쿼리 프로파일링
config.enable_profiling()
profiler = get_profiler()
profiler.clear() # 이전 데이터 초기화
# 쿼리 실행
result = ds.filter(...).groupby(...).agg(...).to_df()
# 이 쿼리의 프로파일 확인
print(profiler.report())여러 쿼리 프로파일링하기
config.enable_profiling()
profiler = get_profiler()
profiler.clear()
# 쿼리 1
with profiler.step("Query 1"):
result1 = query1.to_df()
# 쿼리 2
with profiler.step("Query 2"):
result2 = query2.to_df()
print(profiler.report())접근 방식 비교
profiler = get_profiler()
# 방법 1: 필터 후 groupby
profiler.clear()
with profiler.step("filter_then_groupby"):
result1 = ds.filter(ds['x'] > 10).groupby('y').sum().to_df()
summary1 = profiler.summary()
time1 = summary1.get('filter_then_groupby', 0)
# 방법 2: Groupby 후 필터
profiler.clear()
with profiler.step("groupby_then_filter"):
result2 = ds.groupby('y').sum().filter(ds['x'] > 10).to_df()
summary2 = profiler.summary()
time2 = summary2.get('groupby_then_filter', 0)
print(f"Approach 1: {time1:.2f}ms")
print(f"Approach 2: {time2:.2f}ms")
print(f"Winner: {'Approach 1' if time1 < time2 else 'Approach 2'}")최적화 팁
1. SQL 실행 시간 확인
SQL execution이 병목이라면:
- 데이터 양을 줄일 수 있도록 필터를 더 추가합니다
- CSV 대신 Parquet를 사용합니다
- 적절한 인덱스가 설정되어 있는지 확인합니다(데이터베이스 소스의 경우)
2. I/O 시간 확인
read_csv 또는 read_parquet가 병목이라면:
- Parquet 사용(열 지향, 압축 형식)
- 필요한 컬럼만 읽기
- 가능하면 원본 데이터에서 필터링
3. 데이터 전송 확인
to_df가 느리다면:
- 결과 집합(result set)이 너무 클 수 있습니다
- 필터를 더 추가하거나 limit를 설정하세요
- 미리 보려면
head()를 사용하세요
4. 엔진 비교
from chdb.datastore.config import config
# chdb로 프로파일링
config.use_chdb()
profiler.clear()
result_chdb = query.to_df()
time_chdb = profiler.total_duration_ms
# pandas로 프로파일링
config.use_pandas()
profiler.clear()
result_pandas = query.to_df()
time_pandas = profiler.total_duration_ms
print(f"chdb: {time_chdb:.2f}ms")
print(f"pandas: {time_pandas:.2f}ms")모범 사례
1. 최적화하기 전에 먼저 프로파일링
# 추측하지 말고 측정하세요!
config.enable_profiling()
result = your_query.to_df()
print(get_profiler().report())2. 테스트 사이에는 초기화하세요
profiler.clear() # 이전 데이터 초기화
# 테스트 실행
print(profiler.report())3. 초점을 맞추려면 min_duration_ms를 사용하세요
# 100ms 이상 걸리는 작업만 표시
profiler.report(min_duration_ms=100)4. 대표성 있는 데이터를 프로파일링하세요
# 실제 데이터 크기로 프로파일링하세요
# 소규모 테스트 데이터로는 실제 병목 현상이 드러나지 않을 수 있습니다5. 운영 환경에서는 비활성화
# 개발
config.enable_profiling()
# 운영
config.set_profiling_enabled(False) # 오버헤드 방지예시: 전체 프로파일링 세션
from chdb import datastore as pd
from chdb.datastore.config import config, get_profiler
# Setup
config.enable_profiling()
config.enable_debug() # 현재 진행 상황도 확인
profiler = get_profiler()
# 데이터 로드
profiler.clear()
print("=== Loading Data ===")
ds = pd.read_csv("sales_2024.csv") # 1000만 행
print(profiler.report())
# 쿼리 1: 단순 필터
profiler.clear()
print("\n=== Query 1: Simple Filter ===")
result1 = ds.filter(ds['amount'] > 1000).to_df()
print(profiler.report())
# 쿼리 2: 복합 집계
profiler.clear()
print("\n=== Query 2: Complex Aggregation ===")
result2 = (ds
.filter(ds['amount'] > 100)
.groupby('region', 'category')
.agg({
'amount': ['sum', 'mean', 'count'],
'quantity': 'sum'
})
.sort('sum', ascending=False)
.head(20)
.to_df()
)
print(profiler.report())
# 요약
print("\n=== Summary ===")
print(f"Query 1: {len(result1)} rows")
print(f"Query 2: {len(result2)} rows")