Skip to content
ClickHouse Docs
ClickHouse DocsClickHouse Docs
49.6kGet Started

Datastore I/O operations

View as MarkdownEdit this pageReport an issueSet up docs MCP

DataStore supports reading from and writing to various file formats and data sources.

Reading Data

CSV Files

read_csv(filepath_or_buffer, sep=',', header='infer', names=None, 
         usecols=None, dtype=None, nrows=None, skiprows=None,
         compression=None, encoding=None, **kwargs)

Examples:

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 Files

Recommended for large datasets - columnar format with better compression.

read_parquet(path, columns=None, **kwargs)

Examples:

# 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 Files

read_json(path_or_buf, orient=None, lines=False, **kwargs)

Examples:

# 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 Files

read_excel(io, sheet_name=0, header=0, names=None, **kwargs)

Examples:

# 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 Databases

read_sql(sql, con, **kwargs)

Examples:

# 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)

Other Formats

# 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]

Writing Data

to_csv

Export to CSV format.

to_csv(path_or_buf=None, sep=',', na_rep='', header=True, 
       index=True, mode='w', compression=None, **kwargs)

Examples:

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

Export to Parquet format (recommended for large data).

to_parquet(path, engine='pyarrow', compression='snappy', **kwargs)

Examples:

# 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

Export to JSON format.

to_json(path_or_buf=None, orient='records', lines=False, **kwargs)

Examples:

# 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

Export to Excel format.

to_excel(excel_writer, sheet_name='Sheet1', index=True, **kwargs)

Examples:

# 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

Export to SQL database or generate SQL string.

to_sql(name=None, con=None, schema=None, if_exists='fail', **kwargs)

Examples:

# 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')

Other Export Methods

# 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")

File Format Comparison

Format Read Speed Write Speed File Size Schema Best For
Parquet Fast Fast Small Yes Large datasets, analytics
CSV Medium Fast Large No Compatibility, simple data
JSON Slow Medium Large Partial APIs, nested data
Excel Slow Slow Medium Partial Sharing with non-tech users
Feather Very Fast Very Fast Medium Yes Inter-process, pandas

Recommendations

  1. For analytics workloads: Use Parquet

    • Columnar format allows reading only needed columns
    • Excellent compression
    • Preserves data types

  2. For data exchange: Use CSV or JSON

    • Universal compatibility
    • Human-readable

  3. For pandas interop: Use Feather or Arrow

    • Fastest serialization
    • Type preservation

Compression Support

Reading Compressed Files

# 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')

Writing Compressed Files

# 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

Compression Options

Compression Speed Ratio Use Case
snappy Very Fast Low Default for Parquet
lz4 Very Fast Low Speed priority
gzip Medium High Compatibility
zstd Fast Very High Best balance
bz2 Slow Very High Maximum compression

Streaming I/O

For very large files that don’t fit in memory:

Chunked Reading

# 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)

Using ClickHouse Streaming

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)

Remote Data Sources

HTTP/HTTPS

# Read from URL
ds = pd.read_csv("https://example.com/data.csv")
ds = pd.read_parquet("https://example.com/data.parquet")

S3

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

See Factory Methods for cloud storage options.

Best Practices

1. Use Parquet for Large Files

# 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. Select Only Needed Columns

# 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. Use Compression

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

4. Batch Writes

# 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")
Navigation