Skip to content
ClickHouse Docs
ClickHouse DocsClickHouse Docs

DataStore 执行模型

理解 DataStore 的惰性求值模型,是高效使用它并获得最佳性能的关键。

惰性求值

DataStore 使用惰性求值——操作不会立即执行,而是先记录下来并编译成优化后的 SQL 查询。只有在实际需要结果时,才会执行。

示例:惰性求值与立即求值

from pathlib import Path
Path("sales.csv").write_text("""\
region,product,category,amount,quantity,price,date,order_id
East,Widget,Electronics,5200,10,120,2024-01-15,1001
West,Gadget,Electronics,800,5,160,2024-02-20,1002
East,Gizmo,Home,6500,3,100,2024-03-10,1003
North,Widget,Electronics,4500,6,150,2024-06-18,1004
West,Gadget,Electronics,2000,8,250,2024-09-14,1005
""")

from chdb import datastore as pd

ds = pd.read_csv("sales.csv")

# These operations are NOT executed yet
result = (ds
    .filter(ds['amount'] > 1000)    # Recorded, not executed
    .select('region', 'amount')      # Recorded, not executed
    .groupby('region')               # Recorded, not executed
    .agg({'amount': 'sum'})          # Recorded, not executed
    .sort('sum', ascending=False)    # Recorded, not executed
)

# Still no execution - just building the query plan
print(result.to_sql())
# SELECT region, SUM(amount) AS sum
# FROM file('sales.csv', 'CSVWithNames')
# WHERE amount > 1000
# GROUP BY region
# ORDER BY sum DESC

# NOW execution happens
df = result.to_df()  # <-- Triggers execution

惰性求值的优势

  1. 查询优化:多个操作会编译成一条优化后的 SQL 查询
  2. 过滤器下推:过滤器会在数据源层面应用
  3. 列裁剪:只读取所需的列
  4. 延迟决策:可在运行时选择执行引擎
  5. 计划检查:可在执行前查看/调试查询

执行触发时机

在需要实际值时,系统会自动触发执行:

自动触发

触发方式 示例 描述
print() / repr() print(ds) 显示结果
len() len(ds) 获取行数
.columns ds.columns 获取列名
.dtypes ds.dtypes 获取列类型
.shape ds.shape 获取维度
.index ds.index 获取行索引
.values ds.values 获取 NumPy 数组
迭代 for row in ds 遍历行
to_df() ds.to_df() 转换为 pandas
to_pandas() ds.to_pandas() to&#95;df 的别名
to_dict() ds.to_dict() 转换为字典
to_numpy() ds.to_numpy() 转换为数组
.equals() ds.equals(other) 比较 DataStore

示例:

# All these trigger execution
print(ds)              # Display
len(ds)                # 1000
ds.columns             # Index(['name', 'age', 'city'])
ds.shape               # (1000, 3)
list(ds)               # List of values
ds.to_df()             # pandas DataFrame

保持惰性执行的操作

操作 返回值 描述
filter() DataStore 添加 WHERE 子句
select() DataStore 添加列选择
sort() DataStore 添加 ORDER BY
groupby() LazyGroupBy 为 GROUP BY 做准备
join() DataStore 添加 JOIN
ds['col'] ColumnExpr 列引用
ds[['col1', 'col2']] DataStore 列选择

示例:

# These do NOT trigger execution - they stay lazy
result = ds.filter(ds['age'] > 25)      # Returns DataStore
result = ds.select('name', 'age')        # Returns DataStore
result = ds['name']                      # Returns ColumnExpr
result = ds.groupby('city')              # Returns LazyGroupBy

三阶段执行

DataStore 操作采用三阶段执行模型:

阶段 1:SQL 查询构建 (惰性)

可用 SQL 表达的操作会先累积起来:

result = (ds
    .filter(ds['status'] == 'active')   # WHERE
    .select('user_id', 'amount')         # SELECT
    .groupby('user_id')                  # GROUP BY
    .agg({'amount': 'sum'})              # SUM()
    .sort('sum', ascending=False)        # ORDER BY
    .limit(10)                           # LIMIT
)
# All compiled into one SQL query

第 2 阶段:执行时机

当触发条件出现时,就会执行已累积的 SQL:

# Execution triggered here
df = result.to_df()  
# The single optimized SQL query runs now

阶段 3:DataFrame 操作 (如有)

如果你在执行后继续进行仅适用于 pandas 的操作:

# Mixed operations
result = (ds
    .filter(ds['amount'] > 100)          # Phase 1: SQL
    .to_df()                             # Phase 2: Execute
    .pivot_table(...)                    # Phase 3: pandas
)

查看执行计划

使用 explain() 可查看将如何执行:

Querypython
ds = pd.read_csv("sales.csv")

query = (ds
    .filter(ds['amount'] > 1000)
    .groupby('region')
    .agg({'amount': ['sum', 'mean']})
)

# View execution plan
query.explain()
Responsetext
Pipeline:
  1. Source: file('sales.csv', 'CSVWithNames')
  2. Filter: amount > 1000
  3. GroupBy: region
  4. Aggregate: sum(amount), avg(amount)

Generated SQL:
SELECT region, SUM(amount) AS sum, AVG(amount) AS mean
FROM file('sales.csv', 'CSVWithNames')
WHERE amount > 1000
GROUP BY region

使用 verbose=True 查看更多详细信息:

query.explain(verbose=True)

参见调试:explain()了解完整文档。


缓存

DataStore 会缓存执行结果,以避免重复查询。

缓存的工作原理

from pathlib import Path
Path("data.csv").write_text("""\
name,age,city,salary,department
Alice,25,NYC,55000,Engineering
Bob,30,LA,65000,Product
Charlie,35,NYC,80000,Engineering
Diana,28,SF,70000,Design
Eve,42,NYC,95000,Product
""")

ds = pd.read_csv("data.csv")
result = ds.filter(ds['age'] > 25)

# First access - executes query
print(result.shape)  # Executes and caches

# Second access - uses cache
print(result.columns)  # Uses cached result

# Third access - uses cache
df = result.to_df()  # Uses cached result

缓存失效

当操作修改 DataStore 时,缓存会失效:

result = ds.filter(ds['age'] > 25)
print(result.shape)  # Executes, caches

# New operation invalidates cache
result2 = result.filter(result['city'] == 'NYC')
print(result2.shape)  # Re-executes (different query)

手动控制缓存

# Clear cache
ds.clear_cache()

# Disable caching
from chdb.datastore.config import config
config.set_cache_enabled(False)

混合 SQL 与 Pandas 操作

DataStore 可智能处理结合 SQL 和 Pandas 的操作:

与 SQL 兼容的操作

以下操作会被编译为 SQL:

  • filter(), where()
  • select()
  • groupby(), agg()
  • sort(), orderby()
  • limit(), offset()
  • join(), union()
  • distinct()
  • 列操作 (数学运算、比较、字符串方法)

仅限 Pandas 的操作

以下操作会触发执行,并使用 pandas:

  • 使用自定义函数的 apply()
  • 带有复杂聚合的 pivot_table()
  • stack()unstack()
  • 对已执行的 DataFrame 执行的操作

混合式管道

# SQL phase
result = (ds
    .filter(ds['amount'] > 100)      # SQL
    .groupby('category')              # SQL
    .agg({'amount': 'sum'})           # SQL
)

# Execution + pandas phase
result = (result
    .to_df()                          # Execute SQL
    .pivot_table(...)                 # pandas operation
)

执行引擎选择

DataStore 可使用不同的引擎来执行操作:

自动模式 (默认)

from chdb.datastore.config import config

config.set_execution_engine('auto')  # Default
# Automatically selects best engine per operation

强制使用 chDB 引擎

config.set_execution_engine('chdb')
# All operations use ClickHouse SQL

强制使用 pandas 引擎

config.set_execution_engine('pandas')
# All operations use pandas

详见配置:执行引擎


性能影响

好:尽早过滤

# Good: Filter in SQL, then aggregate
result = (ds
    .filter(ds['date'] >= '2024-01-01')  # Reduces data early
    .groupby('category')
    .agg({'amount': 'sum'})
)

不佳:过晚进行过滤

# Bad: Aggregate all, then filter
result = (ds
    .groupby('category')
    .agg({'amount': 'sum'})
    .to_df()
    .query('sum > 1000')  # Pandas filter after aggregation
)

推荐:尽早选择列

# Good: Select columns in SQL
result = (ds
    .select('user_id', 'amount', 'date')
    .filter(ds['date'] >= '2024-01-01')
    .groupby('user_id')
    .agg({'amount': 'sum'})
)

推荐做法:让 SQL 来处理

# Good: Complex aggregation in SQL
result = (ds
    .groupby('category')
    .agg({
        'amount': ['sum', 'mean', 'count'],
        'quantity': 'sum'
    })
    .sort('sum', ascending=False)
    .limit(10)
)
# One SQL query does everything

# Bad: Multiple separate queries
sums = ds.groupby('category')['amount'].sum().to_df()
means = ds.groupby('category')['amount'].mean().to_df()
# Two queries instead of one

最佳实践摘要

  1. 执行前先串联各项操作 - 先构建完整查询,再统一触发一次
  2. 尽早过滤 - 在源头减少数据量
  3. 只选择所需列 - 列裁剪可提升性能
  4. 使用 explain() 理解执行方式 - 运行前先调试
  5. 让 SQL 处理聚合 - ClickHouse 已对此进行了优化
  6. 留意执行触发时机 - 避免意外过早执行
  7. 合理使用缓存 - 了解缓存何时会失效
Navigation