title: “Python API” weight: 5 type: docs aliases:

  • /api/python-api.html

Python API

PyPaimon is a Python implementation for connecting Paimon catalog, reading & writing tables. The complete Python implementation of the brand new PyPaimon does not require JDK installation.

Environment Settings

SDK is published at pypaimon. You can install by

pip install pypaimon

Create Catalog

Before coming into contact with the Table, you need to create a Catalog.

{{< tabs “create-catalog” >}} {{< tab “filesystem” >}}

from pypaimon import CatalogFactory

# Note that keys and values are all string
catalog_options = {
    'warehouse': 'file:///path/to/warehouse'
}
catalog = CatalogFactory.create(catalog_options)

{{< /tab >}} {{< tab “rest catalog” >}} The sample code is as follows. The detailed meaning of option can be found in DLF Token.

from pypaimon import CatalogFactory

# Note that keys and values are all string
catalog_options = {
  'metastore': 'rest',
  'warehouse': 'xxx',
  'uri': 'xxx',
  'dlf.region': 'xxx',
  'token.provider': 'xxx',
  'dlf.access-key-id': 'xxx',
  'dlf.access-key-secret': 'xxx'
}
catalog = CatalogFactory.create(catalog_options)

{{< /tab >}} {{< /tabs >}}

Currently, PyPaimon only support filesystem catalog and rest catalog. See [Catalog]({{< ref “concepts/catalog” >}}).

You can use the catalog to create table for writing data.

Create Database

Table is located in a database. If you want to create table in a new database, you should create it.

catalog.create_database(
    name='database_name',
    ignore_if_exists=True,  # To raise error if the database exists, set False
    properties={'key': 'value'}  # optional database properties
)

Create Table

Table schema contains fields definition, partition keys, primary keys, table options and comment. The field definition is described by pyarrow.Schema. All arguments except fields definition are optional.

Generally, there are two ways to build pyarrow.Schema.

First, you can use pyarrow.schema method directly, for example:

import pyarrow as pa

from pypaimon import Schema

pa_schema = pa.schema([
    ('dt', pa.string()),
    ('hh', pa.string()),
    ('pk', pa.int64()),
    ('value', pa.string())
])

schema = Schema.from_pyarrow_schema(
    pa_schema=pa_schema,
    partition_keys=['dt', 'hh'],
    primary_keys=['dt', 'hh', 'pk'],
    options={'bucket': '2'},
    comment='my test table')

See [Data Types]({{< ref “python-api#data-types” >}}) for all supported pyarrow-to-paimon data types mapping.

Second, if you have some Pandas data, the pa_schema can be extracted from DataFrame:

import pandas as pd
import pyarrow as pa

from pypaimon import Schema

# Example DataFrame data
data = {
    'dt': ['2024-01-01', '2024-01-01', '2024-01-02'],
    'hh': ['12', '15', '20'],
    'pk': [1, 2, 3],
    'value': ['a', 'b', 'c'],
}
dataframe = pd.DataFrame(data)

# Get Paimon Schema
record_batch = pa.RecordBatch.from_pandas(dataframe)
schema = Schema.from_pyarrow_schema(
    pa_schema=record_batch.schema,
    partition_keys=['dt', 'hh'],
    primary_keys=['dt', 'hh', 'pk'],
    options={'bucket': '2'},
    comment='my test table'
)

After building table schema, you can create corresponding table:

schema = ...
catalog.create_table(
    identifier='database_name.table_name',
    schema=schema,
    ignore_if_exists=True  # To raise error if the table exists, set False
)

# Get Table
table = catalog.get_table('database_name.table_name')

Batch Write

Paimon table write is Two-Phase Commit, you can write many times, but once committed, no more data can be written.

{{< hint warning >}} Currently, the feature of writing multiple times and committing once only supports append only table. {{< /hint >}}

table = catalog.get_table('database_name.table_name')

# 1. Create table write and commit
write_builder = table.new_batch_write_builder()
table_write = write_builder.new_write()
table_commit = write_builder.new_commit()

# 2. Write data. Support 3 methods:
# 2.1 Write pandas.DataFrame
dataframe = ...
table_write.write_pandas(dataframe)

# 2.2 Write pyarrow.Table
pa_table = ...
table_write.write_arrow(pa_table)

# 2.3 Write pyarrow.RecordBatch
record_batch = ...
table_write.write_arrow_batch(record_batch)

# 3. Commit data
commit_messages = table_write.prepare_commit()
table_commit.commit(commit_messages)

# 4. Close resources
table_write.close()
table_commit.close()

By default, the data will be appended to table. If you want to overwrite table, you should use TableWrite#overwrite API:

# overwrite whole table
write_builder = table.new_batch_write_builder().overwrite()

# overwrite partition 'dt=2024-01-01'
write_builder = table.new_batch_write_builder().overwrite({'dt': '2024-01-01'})

Update columns

You can create TableUpdate.update_by_arrow_with_row_id to update columns to data evolution tables.

The input data should include the _ROW_ID column, update operation will automatically sort and match each _ROW_ID to its corresponding first_row_id, then groups rows with the same first_row_id and writes them to a separate file.

simple_pa_schema = pa.schema([
  ('f0', pa.int8()),
  ('f1', pa.int16()),
])
schema = Schema.from_pyarrow_schema(simple_pa_schema,
                                    options={'row-tracking.enabled': 'true', 'data-evolution.enabled': 'true'})
catalog.create_table('default.test_row_tracking', schema, False)
table = catalog.get_table('default.test_row_tracking')

# write all columns
write_builder = table.new_batch_write_builder()
table_write = write_builder.new_write()
table_commit = write_builder.new_commit()
expect_data = pa.Table.from_pydict({
  'f0': [-1, 2],
  'f1': [-1001, 1002]
}, schema=simple_pa_schema)
table_write.write_arrow(expect_data)
table_commit.commit(table_write.prepare_commit())
table_write.close()
table_commit.close()

# update partial columns
write_builder = table.new_batch_write_builder()
table_update = write_builder.new_update().with_update_type(['f0'])
table_commit = write_builder.new_commit()
data2 = pa.Table.from_pydict({
  '_ROW_ID': [0, 1],
  'f0': [5, 6],
}, schema=pa.schema([
  ('_ROW_ID', pa.int64()),
  ('f0', pa.int8()),
]))
cmts = table_update.update_by_arrow_with_row_id(data2)
table_commit.commit(cmts)
table_commit.close()

# content should be:
#   'f0': [5, 6],
#   'f1': [-1001, 1002]

Batch Read

Predicate pushdown

A ReadBuilder is used to build reading utils and perform filter and projection pushdown.

table = catalog.get_table('database_name.table_name')
read_builder = table.new_read_builder()

You can use PredicateBuilder to build filters and pushdown them by ReadBuilder:

# Example filter: ('f0' < 3 OR 'f1' > 6) AND 'f3' = 'A'

predicate_builder = read_builder.new_predicate_builder()

predicate1 = predicate_builder.less_than('f0', 3)
predicate2 = predicate_builder.greater_than('f1', 6)
predicate3 = predicate_builder.or_predicates([predicate1, predicate2])

predicate4 = predicate_builder.equal('f3', 'A')
predicate_5 = predicate_builder.and_predicates([predicate3, predicate4])

read_builder = read_builder.with_filter(predicate_5)

See [Predicate]({{< ref “python-api#predicate” >}}) for all supported filters and building methods.

You can also pushdown projection by ReadBuilder:

# select f3 and f2 columns
read_builder = read_builder.with_projection(['f3', 'f2'])

Generate Splits

Then you can step into Scan Plan stage to get splits:

table_scan = read_builder.new_scan()
splits = table_scan.plan().splits()

Finally, you can read data from the splits to various data format.

Read Apache Arrow

This requires pyarrow to be installed.

You can read all the data into a pyarrow.Table:

table_read = read_builder.new_read()
pa_table = table_read.to_arrow(splits)
print(pa_table)

# pyarrow.Table
# f0: int32
# f1: string
# ----
# f0: [[1,2,3],[4,5,6],...]
# f1: [["a","b","c"],["d","e","f"],...]

You can also read data into a pyarrow.RecordBatchReader and iterate record batches:

table_read = read_builder.new_read()
for batch in table_read.to_arrow_batch_reader(splits):
    print(batch)

# pyarrow.RecordBatch
# f0: int32
# f1: string
# ----
# f0: [1,2,3]
# f1: ["a","b","c"]

Read Python Iterator

You can read the data row by row into a native Python iterator. This is convenient for custom row-based processing logic.

table_read = read_builder.new_read()
for row in table_read.to_iterator(splits):
    print(row)

# [1,2,3]
# ["a","b","c"]

Read Pandas

This requires pandas to be installed.

You can read all the data into a pandas.DataFrame:

table_read = read_builder.new_read()
df = table_read.to_pandas(splits)
print(df)

#    f0 f1
# 0   1  a
# 1   2  b
# 2   3  c
# 3   4  d
# ...

Read DuckDB

This requires duckdb to be installed.

You can convert the splits into an in-memory DuckDB table and query it:

table_read = read_builder.new_read()
duckdb_con = table_read.to_duckdb(splits, 'duckdb_table')

print(duckdb_con.query("SELECT * FROM duckdb_table").fetchdf())
#    f0 f1
# 0   1  a
# 1   2  b
# 2   3  c
# 3   4  d
# ...

print(duckdb_con.query("SELECT * FROM duckdb_table WHERE f0 = 1").fetchdf())
#    f0 f1
# 0   1  a

Read Ray

This requires ray to be installed.

You can convert the splits into a Ray Dataset and handle it by Ray Data API for distributed processing:

table_read = read_builder.new_read()
ray_dataset = table_read.to_ray(splits)

print(ray_dataset)
# MaterializedDataset(num_blocks=1, num_rows=9, schema={f0: int32, f1: string})

print(ray_dataset.take(3))
# [{'f0': 1, 'f1': 'a'}, {'f0': 2, 'f1': 'b'}, {'f0': 3, 'f1': 'c'}]

print(ray_dataset.to_pandas())
#    f0 f1
# 0   1  a
# 1   2  b
# 2   3  c
# 3   4  d
# ...

The to_ray() method supports Ray Data API parameters for distributed processing:

# Basic usage
ray_dataset = table_read.to_ray(splits)

# Specify number of output blocks
ray_dataset = table_read.to_ray(splits, override_num_blocks=4)

# Configure Ray remote arguments
ray_dataset = table_read.to_ray(
    splits,
    override_num_blocks=4,
    ray_remote_args={"num_cpus": 2, "max_retries": 3}
)

# Use Ray Data operations
mapped_dataset = ray_dataset.map(lambda row: {'value': row['value'] * 2})
filtered_dataset = ray_dataset.filter(lambda row: row['score'] > 80)
df = ray_dataset.to_pandas()

Parameters:

  • override_num_blocks: Optional override for the number of output blocks. By default, Ray automatically determines the optimal number.
  • ray_remote_args: Optional kwargs passed to ray.remote() in read tasks (e.g., {"num_cpus": 2, "max_retries": 3}).
  • concurrency: Optional max number of Ray tasks to run concurrently. By default, dynamically decided based on available resources.
  • **read_args: Additional kwargs passed to the datasource (e.g., per_task_row_limit in Ray 2.52.0+).

Ray Block Size Configuration:

If you need to configure Ray‘s block size (e.g., when Paimon splits exceed Ray’s default 128MB block size), set it before calling to_ray():

from ray.data import DataContext

ctx = DataContext.get_current()
ctx.target_max_block_size = 256 * 1024 * 1024  # 256MB (default is 128MB)
ray_dataset = table_read.to_ray(splits)

See Ray Data API Documentation for more details.

Read Pytorch Dataset

This requires torch to be installed.

You can read all the data into a torch.utils.data.Dataset or torch.utils.data.IterableDataset:

from torch.utils.data import DataLoader

table_read = read_builder.new_read()
dataset = table_read.to_torch(splits, streaming=True)
dataloader = DataLoader(
    dataset,
    batch_size=2,
    num_workers=2,  # Concurrency to read data
    shuffle=False
)

# Collect all data from dataloader
for batch_idx, batch_data in enumerate(dataloader):
    print(batch_data)

# output:
#   {'user_id': tensor([1, 2]), 'behavior': ['a', 'b']}
#   {'user_id': tensor([3, 4]), 'behavior': ['c', 'd']}
#   {'user_id': tensor([5, 6]), 'behavior': ['e', 'f']}
#   {'user_id': tensor([7, 8]), 'behavior': ['g', 'h']}

When the streaming parameter is true, it will iteratively read; when it is false, it will read the full amount of data into memory.

Incremental Read

This API allows reading data committed between two snapshot timestamps. The steps are as follows.

  • Set the option CoreOptions.INCREMENTAL_BETWEEN_TIMESTAMP on a copied table via table.copy({...}). The value must be a string: "startMillis,endMillis", where startMillis is exclusive and endMillis is inclusive.
  • Use SnapshotManager to obtain snapshot timestamps or you can determine them by yourself.
  • Read the data as above.

Example:

from pypaimon import CatalogFactory
from pypaimon.common.core_options import CoreOptions
from pypaimon.snapshot.snapshot_manager import SnapshotManager

# Prepare catalog and obtain a table
catalog = CatalogFactory.create({'warehouse': '/path/to/warehouse'})
table = catalog.get_table('default.your_table_name')

# Assume the table has at least two snapshots (1 and 2)
snapshot_manager = SnapshotManager(table)
t1 = snapshot_manager.get_snapshot_by_id(1).time_millis
t2 = snapshot_manager.get_snapshot_by_id(2).time_millis

# Read records committed between [t1, t2]
table_inc = table.copy({CoreOptions.INCREMENTAL_BETWEEN_TIMESTAMP: f"{t1},{t2}"})

read_builder = table_inc.new_read_builder()
table_scan = read_builder.new_scan()
table_read = read_builder.new_read()
splits = table_scan.plan().splits()

# To Arrow
arrow_table = table_read.to_arrow(splits)

# Or to pandas
pandas_df = table_read.to_pandas(splits)

Shard Read

Shard Read allows you to read data in parallel by dividing the table into multiple shards. This is useful for distributed processing and parallel computation.

You can specify the shard index and total number of shards to read a specific portion of the data:

# Prepare read builder
table = catalog.get_table('database_name.table_name')
read_builder = table.new_read_builder()
table_read = read_builder.new_read()

# Read the second shard (index 1) out of 3 total shards
splits = read_builder.new_scan().with_shard(1, 3).plan().splits()

# Read all shards and concatenate results
splits1 = read_builder.new_scan().with_shard(0, 3).plan().splits()
splits2 = read_builder.new_scan().with_shard(1, 3).plan().splits()
splits3 = read_builder.new_scan().with_shard(2, 3).plan().splits()

# Combine results from all shards

all_splits = splits1 + splits2 + splits3
pa_table = table_read.to_arrow(all_splits)

Example with shard read:

import pyarrow as pa
from pypaimon import CatalogFactory, Schema

# Create catalog
catalog_options = {'warehouse': 'file:///path/to/warehouse'}
catalog = CatalogFactory.create(catalog_options)
catalog.create_database("default", False)
# Define schema
pa_schema = pa.schema([
    ('user_id', pa.int64()),
    ('item_id', pa.int64()),
    ('behavior', pa.string()),
    ('dt', pa.string()),
])

# Create table and write data
schema = Schema.from_pyarrow_schema(pa_schema, partition_keys=['dt'])
catalog.create_table('default.test_table', schema, False)
table = catalog.get_table('default.test_table')

# Write data in two batches
write_builder = table.new_batch_write_builder()

# First write
table_write = write_builder.new_write()
table_commit = write_builder.new_commit()
data1 = {
    'user_id': [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14],
    'item_id': [1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009, 1010, 1011, 1012, 1013, 1014],
    'behavior': ['a', 'b', 'c', None, 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm'],
    'dt': ['p1', 'p1', 'p2', 'p1', 'p2', 'p1', 'p2', 'p1', 'p2', 'p1', 'p2', 'p1', 'p2', 'p1'],
}
pa_table = pa.Table.from_pydict(data1, schema=pa_schema)
table_write.write_arrow(pa_table)
table_commit.commit(table_write.prepare_commit())
table_write.close()
table_commit.close()

# Second write
table_write = write_builder.new_write()
table_commit = write_builder.new_commit()
data2 = {
    'user_id': [5, 6, 7, 8, 18],
    'item_id': [1005, 1006, 1007, 1008, 1018],
    'behavior': ['e', 'f', 'g', 'h', 'z'],
    'dt': ['p2', 'p1', 'p2', 'p2', 'p1'],
}
pa_table = pa.Table.from_pydict(data2, schema=pa_schema)
table_write.write_arrow(pa_table)
table_commit.commit(table_write.prepare_commit())
table_write.close()
table_commit.close()

# Read specific shard
read_builder = table.new_read_builder()
table_read = read_builder.new_read()

# Read shard 2 out of 3 total shards
splits = read_builder.new_scan().with_shard(2, 3).plan().splits()
shard_data = table_read.to_arrow(splits)

# Verify shard distribution by reading all shards
splits1 = read_builder.new_scan().with_shard(0, 3).plan().splits()
splits2 = read_builder.new_scan().with_shard(1, 3).plan().splits()
splits3 = read_builder.new_scan().with_shard(2, 3).plan().splits()

# Combine all shards should equal full table read
all_shards_data = pa.concat_tables([
    table_read.to_arrow(splits1),
    table_read.to_arrow(splits2),
    table_read.to_arrow(splits3),
])
full_table_data = table_read.to_arrow(read_builder.new_scan().plan().splits())

Key points about shard read:

  • Shard Index: Zero-based index of the shard to read (0 to total_shards-1)
  • Total Shards: Total number of shards to divide the data into
  • Data Distribution: Data is distributed evenly across shards, with remainder rows going to the last shard
  • Parallel Processing: Each shard can be processed independently for better performance
  • Consistency: Combining all shards should produce the complete table data

Data Types

Python Native TypePyArrow TypePaimon Type
intpyarrow.int8()TINYINT
intpyarrow.int16()SMALLINT
intpyarrow.int32()INT
intpyarrow.int64()BIGINT
floatpyarrow.float32()FLOAT
floatpyarrow.float64()DOUBLE
boolpyarrow.bool_()BOOLEAN
strpyarrow.string()STRING, CHAR(n), VARCHAR(n)
bytespyarrow.binary()BYTES, VARBINARY(n)
bytespyarrow.binary(length)BINARY(length)
decimal.Decimalpyarrow.decimal128(precision, scale)DECIMAL(precision, scale)
datetime.datetimepyarrow.timestamp(unit, tz=None)TIMESTAMP(p)
datetime.datepyarrow.date32()DATE
datetime.timepyarrow.time32(unit) or pyarrow.time64(unit)TIME(p)

Predicate

Predicate kindPredicate method
p1 and p2PredicateBuilder.and_predicates([p1, p2])
p1 or p2PredicateBuilder.or_predicates([p1, p2])
f = literalPredicateBuilder.equal(f, literal)
f != literalPredicateBuilder.not_equal(f, literal)
f < literalPredicateBuilder.less_than(f, literal)
f <= literalPredicateBuilder.less_or_equal(f, literal)
f > literalPredicateBuilder.greater_than(f, literal)
f >= literalPredicateBuilder.greater_or_equal(f, literal)
f is nullPredicateBuilder.is_null(f)
f is not nullPredicateBuilder.is_not_null(f)
f.startswith(literal)PredicateBuilder.startswith(f, literal)
f.endswith(literal)PredicateBuilder.endswith(f, literal)
f.contains(literal)PredicateBuilder.contains(f, literal)
f is in [l1, l2]PredicateBuilder.is_in(f, [l1, l2])
f is not in [l1, l2]PredicateBuilder.is_not_in(f, [l1, l2])
lower <= f <= upperPredicateBuilder.between(f, lower, upper)

Supported Features

The following shows the supported features of Python Paimon compared to Java Paimon:

Catalog Level

  • FileSystemCatalog
  • RestCatalog

Table Level

  • Append Tables
    • bucket = -1 (unaware)
    • bucket > 0 (fixed)
  • Primary Key Tables
    • only support deduplicate
    • bucket = -2 (postpone)
    • bucket > 0 (fixed)
    • read with deletion vectors enabled
  • Read/Write Operations
    • Batch read and write for append tables and primary key tables
    • Predicate filtering
    • Overwrite semantics
    • Incremental reading of Delta data
    • Reading and writing blob data
    • with_shard feature