> ## Documentation Index
> Fetch the complete documentation index at: https://clickhouse.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# DataStore I/O 操作

> 使用 DataStore 读写数据——支持的所有格式和目标端

DataStore 支持从多种文件格式和数据源读取数据，也支持将数据写入其中。

## 读取数据

### CSV 文件

```python theme={null}
read_csv(filepath_or_buffer, sep=',', header='infer', names=None, 
         usecols=None, dtype=None, nrows=None, skiprows=None,
         compression=None, encoding=None, **kwargs)
```

**示例：**

```python theme={null}
from chdb import datastore as pd

# Basic CSV read
ds = pd.read_csv("data.csv")

# With options
ds = pd.read_csv(
    "data.csv",
    sep=";",                    # Custom delimiter
    header=0,                   # Header row index
    names=['a', 'b', 'c'],      # Custom column names
    usecols=['a', 'b'],         # Only read specific columns
    dtype={'a': 'Int64'},       # Specify dtypes
    nrows=1000,                 # Read only first 1000 rows
    skiprows=1,                 # Skip first row
    compression='gzip',         # Compressed file
    encoding='utf-8'            # Encoding
)

# From URL
ds = pd.read_csv("https://example.com/data.csv")
```

### Parquet 文件

推荐用于大型数据集——采用列式格式，压缩效果更佳。

```python theme={null}
read_parquet(path, columns=None, **kwargs)
```

**示例：**

```python theme={null}
# Basic Parquet read
ds = pd.read_parquet("data.parquet")

# Read specific columns only (efficient - only reads needed data)
ds = pd.read_parquet("data.parquet", columns=['col1', 'col2', 'col3'])

# From S3
ds = pd.read_parquet("s3://bucket/data.parquet")
```

### JSON 文件

```python theme={null}
read_json(path_or_buf, orient=None, lines=False, **kwargs)
```

**示例：**

```python theme={null}
# Standard JSON
ds = pd.read_json("data.json")

# JSON Lines (newline-delimited)
ds = pd.read_json("data.jsonl", lines=True)

# JSON with specific orientation
ds = pd.read_json("data.json", orient='records')
```

### Excel 文件

```python theme={null}
read_excel(io, sheet_name=0, header=0, names=None, **kwargs)
```

**示例：**

```python theme={null}
# Read first sheet
ds = pd.read_excel("data.xlsx")

# Read specific sheet
ds = pd.read_excel("data.xlsx", sheet_name="Sheet1")
ds = pd.read_excel("data.xlsx", sheet_name=2)  # Third sheet

# Read multiple sheets (returns dict)
sheets = pd.read_excel("data.xlsx", sheet_name=['Sheet1', 'Sheet2'])
```

### SQL 数据库

```python theme={null}
read_sql(sql, con, **kwargs)
```

**示例：**

```python theme={null}
# Read from SQL query
ds = pd.read_sql("SELECT * FROM users", connection)
ds = pd.read_sql("SELECT * FROM orders WHERE date > '2024-01-01'", connection)
```

### 其他格式

```python theme={null}
# Feather (Arrow)
ds = pd.read_feather("data.feather")

# ORC
ds = pd.read_orc("data.orc")

# Pickle
ds = pd.read_pickle("data.pkl")

# Fixed-width formatted
ds = pd.read_fwf("data.txt", widths=[10, 20, 15])

# HTML tables
ds = pd.read_html("https://example.com/table.html")[0]
```

***

## 写入数据

### to\_csv

导出为 CSV 格式。

```python theme={null}
to_csv(path_or_buf=None, sep=',', na_rep='', header=True, 
       index=True, mode='w', compression=None, **kwargs)
```

**示例：**

```python theme={null}
ds = pd.read_parquet("data.parquet")

# Basic export
ds.to_csv("output.csv")

# With options
ds.to_csv(
    "output.csv",
    sep=";",                    # Custom delimiter
    index=False,                # Don't include index
    header=True,                # Include header
    na_rep='NULL',              # Represent NaN as 'NULL'
    compression='gzip'          # Compress output
)

# To string
csv_string = ds.to_csv()
```

### to\_parquet

导出为 Parquet 格式 (建议用于大规模数据) 。

```python theme={null}
to_parquet(path, engine='pyarrow', compression='snappy', **kwargs)
```

**示例：**

```python theme={null}
# Basic export
ds.to_parquet("output.parquet")

# With compression options
ds.to_parquet("output.parquet", compression='gzip')
ds.to_parquet("output.parquet", compression='zstd')

# Partitioned output
ds.to_parquet(
    "output/",
    partition_cols=['year', 'month']
)
```

### to\_json

导出为 JSON 格式。

```python theme={null}
to_json(path_or_buf=None, orient='records', lines=False, **kwargs)
```

**示例：**

```python theme={null}
# Standard JSON (array of records)
ds.to_json("output.json", orient='records')

# JSON Lines (one JSON object per line)
ds.to_json("output.jsonl", lines=True)

# Different orientations
ds.to_json("output.json", orient='split')    # {columns, data, index}
ds.to_json("output.json", orient='records')  # [{col: val}, ...]
ds.to_json("output.json", orient='columns')  # {col: {idx: val}}

# To string
json_string = ds.to_json()
```

### to\_excel

导出为 Excel 格式。

```python theme={null}
to_excel(excel_writer, sheet_name='Sheet1', index=True, **kwargs)
```

**示例：**

```python theme={null}
# Single sheet
ds.to_excel("output.xlsx")
ds.to_excel("output.xlsx", sheet_name="Data", index=False)

# Multiple sheets
with pd.ExcelWriter("output.xlsx") as writer:
    ds1.to_excel(writer, sheet_name="Sales")
    ds2.to_excel(writer, sheet_name="Inventory")
```

### to\_sql

导出至 SQL 数据库，或生成 SQL 字符串。

```python theme={null}
to_sql(name=None, con=None, schema=None, if_exists='fail', **kwargs)
```

**示例：**

```python theme={null}
# Generate SQL query (no execution)
sql = ds.to_sql()
print(sql)
# SELECT ...
# FROM ...
# WHERE ...

# Write to database
ds.to_sql("table_name", connection, if_exists='replace')
```

### 其他导出方式

```python theme={null}
# To pandas DataFrame
df = ds.to_df()
df = ds.to_pandas()

# To Arrow Table
table = ds.to_arrow()

# To NumPy array
arr = ds.to_numpy()

# To dictionary
d = ds.to_dict()
d = ds.to_dict(orient='records')  # List of dicts
d = ds.to_dict(orient='list')     # Dict of lists

# To records (list of tuples)
records = ds.to_records()

# To string
s = ds.to_string()
s = ds.to_string(max_rows=100)

# To Markdown
md = ds.to_markdown()

# To HTML
html = ds.to_html()

# To LaTeX
latex = ds.to_latex()

# To clipboard
ds.to_clipboard()

# To pickle
ds.to_pickle("output.pkl")

# To feather
ds.to_feather("output.feather")
```

***

## 文件格式比较

| 格式 | 读取速度 | 写入速度 | 文件大小 | schema | 最适合 |
| - | - | - | - | - | - |
| **Parquet** | 快 | 快 | 小 | 是 | 大型数据集、分析 |
| **CSV** | 中等 | 快 | 大 | 否 | 兼容性、简单数据 |
| **JSON** | 慢 | 中等 | 大 | 部分支持 | API、嵌套数据 |
| **Excel** | 慢 | 慢 | 中等 | 部分支持 | 与非技术用户共享 |
| **Feather** | 非常快 | 非常快 | 中等 | 是 | 进程间、Pandas |

### 建议

1. **对于分析型工作负载：** 使用 Parquet
   * 列式格式允许仅读取所需的列
   * 压缩效果出色
   * 保留数据类型

2. **对于数据交换：** 使用 CSV 或 JSON
   * 兼容性通用
   * 可读性强

3. **对于 pandas 互操作：** 使用 Feather 或 Arrow
   * 序列化速度最快
   * 保留类型

***

## 压缩支持

### 读取压缩文件

```python theme={null}
# Auto-detect from extension
ds = pd.read_csv("data.csv.gz")
ds = pd.read_csv("data.csv.bz2")
ds = pd.read_csv("data.csv.xz")
ds = pd.read_csv("data.csv.zst")

# Explicit compression
ds = pd.read_csv("data.csv", compression='gzip')
```

### 写入压缩后的文件

```python theme={null}
# CSV with compression
ds.to_csv("output.csv.gz", compression='gzip')
ds.to_csv("output.csv.bz2", compression='bz2')

# Parquet (always compressed)
ds.to_parquet("output.parquet", compression='snappy')  # Default
ds.to_parquet("output.parquet", compression='gzip')
ds.to_parquet("output.parquet", compression='zstd')    # Best ratio
ds.to_parquet("output.parquet", compression='lz4')     # Fastest
```

### 压缩选项

| 压缩 | 速度 | 压缩率 | 用例 |
| - | - | - | - |
| `snappy` | 非常快 | 低 | Parquet 默认选项 |
| `lz4` | 非常快 | 低 | 优先考虑速度 |
| `gzip` | 中等 | 高 | 兼容性优先 |
| `zstd` | 快 | 非常高 | 平衡性最佳 |
| `bz2` | 慢 | 非常高 | 最大限度压缩 |

***

## 流式 I/O

对于大到无法装入内存的文件：

### 分块读取

```python theme={null}
# Read in chunks
for chunk in pd.read_csv("large.csv", chunksize=100000):
    # Process each chunk
    process(chunk)

# Using iterator
reader = pd.read_csv("large.csv", iterator=True)
chunk = reader.get_chunk(10000)
```

### 使用 ClickHouse 进行流式处理

```python theme={null}
from chdb.datastore import DataStore

# Stream from file without loading all into memory
ds = DataStore.from_file("huge.parquet")

# Operations are lazy - only computes what's needed
result = ds.filter(ds['amount'] > 1000).head(100)
```

***

## 远程数据源

### HTTP/HTTPS

```python theme={null}
# Read from URL
ds = pd.read_csv("https://example.com/data.csv")
ds = pd.read_parquet("https://example.com/data.parquet")
```

### S3

```python theme={null}
from chdb.datastore import DataStore

# Anonymous access
ds = DataStore.uri("s3://bucket/data.parquet?nosign=true")

# With credentials
ds = DataStore.from_s3(
    "s3://bucket/data.parquet",
    access_key_id="KEY",
    secret_access_key="SECRET"
)
```

### GCS、Azure、HDFS

云存储选项请参见 [工厂方法](https://clickhouse.com/docs/zh/chdb/datastore/factory-methods.md)。

***

## 最佳实践

### 1. 大文件使用 Parquet

```python theme={null}
# Convert CSV to Parquet for better performance
ds = pd.read_csv("large.csv")
ds.to_parquet("large.parquet")

# Future reads are much faster
ds = pd.read_parquet("large.parquet")
```

### 2. 只选择所需列

```python theme={null}
# Efficient - only reads col1 and col2
ds = pd.read_parquet("data.parquet", columns=['col1', 'col2'])

# Inefficient - reads all columns then filters
ds = pd.read_parquet("data.parquet")[['col1', 'col2']]
```

### 3. 启用压缩

```python theme={null}
# Smaller file size, usually faster due to less I/O
ds.to_parquet("output.parquet", compression='zstd')
```

### 4. 批次写入

```python theme={null}
# Write once, not in a loop
result = process_all_data(ds)
result.to_parquet("output.parquet")

# NOT this (inefficient)
for chunk in chunks:
    chunk.to_parquet(f"output_{i}.parquet")
```