PyCylon is the Python binding for LibCylon (C++ Cylon). The uniqueness of PyCylon is that it can be used as a library or a framework. As a library, PyCylon seamlessly integrates with PyArrow. This brings us the capability of providing the user the compatibility with Pandas, Numpy and Tensors. As a framework we support distributed relational algebra operations using MPI as the distributed backend.

Table API

PyCylon Table API currently offers a set of relational algebra operators to preprocess the data.


In a Cylon programme, if you use Cylon with MPI backend, the initialization must be done as follows;

ctx: CylonContext = CylonContext("mpi")

Without MPI,

ctx: CylonContext = CylonContext()
Note: In the current release, Cylon only supports MPI as a distributed backend

Load a Table

Using Cylon

from import Table
from import csv_reader
tb1: Table =, '/tmp/csv.csv', ',')

Using PyArrow and convert to PyCylon Table

from pyarrow import csv
from import Table
from pyarrow import Table as PyArrowTable
pyarrow_tb: PyArrowTable = csv.read_csv('/tmp/csv.csv')
cylon_tb = Table.from_arrow(pyarrow_tb)

Also a Cylon Table can be converted to a PyArrow Table

pyarrow_tb: PyArrowTable = Table.to_arrow(cylon_tb)


Join API supports left, right, inner, outer' joins with Hash or Sort algorithms. User can specify these configs as using Python str.

Sequential Join

tb1: Table =, '/tmp/csv.csv', ',')
tb2: Table =, '/tmp/csv.csv', ',')
tb3: Table = tb1.join(ctx, table=tb2, join_type='left', algorithm='hash', left_col=0, right_col=0)

Distributed Join

tb1: Table =, '/tmp/csv.csv', ',')
tb2: Table =, '/tmp/csv.csv', ',')
tb3: Table = tb1.distributed_join(ctx, table=tb2, join_type='left', algorithm='hash', left_col=0, right_col=0)


Sequential Union

tb4: Table = tb1.union(ctx, table=tb2)

Distributed Union

tb5: Table = tb1.distributed_union(ctx, table=tb2)


Sequential Intersect

tb4: Table = tb1.intersect(ctx, table=tb2)

Distributed Intersect

tb5: Table = tb1.distributed_intersect(ctx, table=tb2)


Sequential Subtract

tb4: Table = tb1.subtract(ctx, table=tb2)

Distributed Subtract

tb5: Table = tb1.distributed_subtract(ctx, table=tb2)


This is not yet supported from PyCylon API, but LibCylon supports this.

Python Examples

  1. Simple Data Loading Benchmark
  2. Sequential MNIST with PyTorch