Skip to content
A simple and efficient tool to parallelize your pandas operations on all your CPUs (On Linux & macOS)
Branch: master
Clone or download
Latest commit 79ca1a5 May 13, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
docs Remove file apply_vs_parallel_apply.png Apr 1, 2019
pandarallel Fix issue #23 May 13, 2019
tests Fix issue #23 May 13, 2019
.gitignore Reorganize repository & prepare pip package Mar 11, 2019
LICENSE Create LICENSE Mar 10, 2019 Specify that Pandarallel works only on Linux & macOS Apr 4, 2019 Refactor May 5, 2019


Without parallelisation Without Pandarallel
With parallelisation With Pandarallel
Latest Release latest release
License license


$ pip install pandarallel [--user]


  • Linux or macOS (Windows is not supported at the moment)


  • Parallelization has a cost (instanciating new processes, sending data via shared memory, etc ...), so parallelization is efficiant only if the amount of calculation to parallelize is high enough. For very little amount of data, using parallezation not always worth it.
  • Functions applied should NOT be lambda functions.
from pandarallel import pandarallel
from math import sin


df.parallel_apply(lambda x: sin(x**2), axis=1)

def func(x):
    return sin(x**2)

df.parallel_apply(func, axis=1)


An example of each API is available here.


For some examples, here is the comparative benchmark with and without using Pandaral·lel.

Computer used for this benchmark:

  • OS: Linux Ubuntu 16.04
  • Hardware: Intel Core i7 @ 3.40 GHz - 4 cores


For those given examples, parallel operations run approximatively 4x faster than the standard operations (except for which runs only 3.2x faster).


First, you have to import pandarallel:

from pandarallel import pandarallel

Then, you have to initialize it.


This method takes 3 optional parameters:

  • shm_size_mb: The size of the Pandarallel shared memory in MB. If the default one is too small, it is possible to set a larger one. By default, it is set to 2 GB. (int)
  • nb_workers: The number of workers. By default, it is set to the number of cores your operating system sees. (int)
  • progress_bar: Put it to True to display a progress bar.

WARNING: Progress bar is an experimental feature. This can lead to a considerable performance loss. Not available for DataFrameGroupy.parallel_apply.

With df a pandas DataFrame, series a pandas Series, func a function to apply/map, args1, args2 some arguments & col_name a column name:

Without parallelisation With parallelisation
df.apply(func) df.parallel_apply(func)
df.applymap(func) df.parallel_applymap(func)
df.groupby(args).apply(func) df.groupby(args).parallel_apply(func)
df.groupby(args1).col_name.rolling(args2).apply(func) df.groupby(args1).col_name.rolling(args2).parallel_apply(func) series.parallel_map(func)
series.apply(func) series.parallel_apply(func)
series.rolling(args).apply(func) series.rolling(args).parallel_apply(func)

You will find a complete example here for each line of this table.


I have 8 CPUs but parallel_apply speeds up computation only about x4. Why ?

Actually Pandarallel can only speed up computation until about the number of cores your computer has. The majority of recent CPUs (like Intel core-i7) uses hyperthreading. For example, a 4 cores hyperthreaded CPU will show 8 CPUs to the Operating System, but will really have only 4 physical computation units.

On Ubuntu, you can get the number of cores with $ grep -m 1 'cpu cores' /proc/cpuinfo.

When I run from pandarallel import pandarallel, I get ModuleNotFoundError: No module named 'pyarrow._plasma. Why?

Are you using Windows? Pandarallel relies on the Pyarrow Plasma shared memory to work. Currently, Pyarrow Plasma works only on Linux & macOS (Windows in not supported). So sorry, but for now Pandarallel is supported only on Linux & macOS ...

You can’t perform that action at this time.