User-Defined Functions (UDF) explained

User defined functions are a very important feature of OpenEO. They allow you as a user to reuse existing code, by submitting it to the back-end.

As datacubes can be very large, the back-end will only be able to run your code on a smaller chunk of the whole cube. So you need to help the back-end a bit, by designing your code to work on as small a piece of data as possible.

There are a few different types of operations where UDF’s can be used:

  1. Applying a process to each pixel: apply

  2. Applying a process to all pixels along a dimension, without changing cardinality: apply_dimension

  3. Reducing values along a dimension: reduce_dimension

  4. Applying a process to all pixels in a multidimensional neighborhood: apply_neighborhood

Not all functions will require you to write a custom process. For instance, if you want to take the absolute value of your datacube, you can simply use the predefined absolute value function. In fact, it is recommended to try and use predefined functions, as they can be more efficiently implemented.

However, when you have a large piece of code that is hard to transform into predefined openEO functions, then it makes sense to use the UDF functionality.

The section below gives an example to get you started.


Don not confuse user-defined functions (abbreviated as UDF) with user-defined processes (sometimes abbreviated as UDP) in openEO, which is a way to define and use your own process graphs as reusable building blocks. see User-Defined Processes for more information.

Example: Smoothing timeseries with a user defined function (UDF)

In this example, we start from the evi_cube that was created in the previous example, and want to apply a temporal smoothing on it. More specifically, we want to use the “Savitzky Golay” smoother that is available in the SciPy Python library.

To ensure that openEO understand your function, it needs to follow some rules, the UDF specification. This is an example that follows those rules:

Example UDF code
import xarray
from scipy.signal import savgol_filter

from openeo.udf import XarrayDataCube

def apply_datacube(cube: XarrayDataCube, context: dict) -> XarrayDataCube:
    Apply Savitzky-Golay smoothing to a timeseries datacube.
    This UDF preserves dimensionality, and assumes an input
    datacube with a temporal dimension 't' as input.
    array: xarray.DataArray = cube.get_array()
    filled = array.interpolate_na(dim='t')
    smoothed_array = savgol_filter(filled.values, 5, 2, axis=0)
    return XarrayDataCube(
        array=xarray.DataArray(smoothed_array, dims=array.dims, coords=array.coords)

The method signature of the UDF is very important, because the back-end will use it to detect the type of UDF. This particular example accepts a DataCube object as input and also returns a DataCube object. The type annotations and method name are actually used to detect how to invoke the UDF, so make sure they remain unchanged.

Once the UDF is defined in a separate file, we need to load it:

def load_udf(relative_path):
    with open(relative_path, 'r+') as f:

smoothing_udf = load_udf('')

after that, we can simply apply it along a dimension:

smoothed_evi = evi_cube_masked.apply_dimension(
    code=smoothing_udf, runtime='Python'

Example: downloading a datacube and executing an UDF locally

Sometimes it is advantageous to run a UDF on the client machine (for example when developing/testing that UDF). This is possible by using the convenience function openeo.udf.run_code.execute_local_udf(). The steps to run a UDF (like the code from above) are as follows:

For example:

my_process = connection.load_collection(...'', format='NetCDF')

smoothing_udf = load_udf('')

from openeo.udf import execute_local_udf
execute_local_udf(smoothing_udf, '', fmt='netcdf')

Note: this algorithm’s primary purpose is to aid client side development of UDFs using small datasets. It is not designed for large jobs.

UDF function names

There’s a predefined set of function signatures that you have to use to implement a UDF:

This module defines a number of function signatures that can be implemented by UDF’s. Both the name of the function and the argument types are/can be used by the backend to validate if the provided UDF is compatible with the calling context of the process graph in which it is used.

openeo.udf.udf_signatures.apply_datacube(cube, context)[source]

Map a XarrayDataCube to another XarrayDataCube.

Depending on the context in which this function is used, the XarrayDataCube dimensions have to be retained or can be chained. For instance, in the context of a reducing operation along a dimension, that dimension will have to be reduced to a single value. In the context of a 1 to 1 mapping operation, all dimensions have to be retained.

Return type


  • cube (XarrayDataCube) – input data cube

  • context (dict) – A dictionary containing user context.


output data cube

openeo.udf.udf_signatures.apply_timeseries(series, context)[source]

Process a timeseries of values, without changing the time instants.

This can for instance be used for smoothing or gap-filling.

Return type


  • series (Series) – A Pandas Series object with a date-time index.

  • context (dict) – A dictionary containing user context.


A Pandas Series object with the same datetime index.


Generic UDF function that directly manipulates a UdfData object


data (UdfData) – UdfData object to manipulate in-place

Profile a process server-side


Experimental feature - This feature only works on back-ends running the Geotrellis implementation, and has not yet been adopted in the openEO API.

Sometimes users want to ‘profile’ their UDF on the back-end. While it’s recommended to first profile it offline, in the same manner as you can debug UDF’s, back-ends may support profiling directly. Note that this will only generate statistics over the python part of the execution, therefore it is only suitable for profiling UDFs.


Only batch jobs are supported! In order to turn on profiling, set ‘profile’ to ‘true’ in job options:

... # prepare the process

When the process has finished, it will also download a file called ‘profile_dumps.tar.gz’:

  • rdd_-1.pstats is the profile data of the python driver,

  • the rest are the profiling results of the individual rdd id-s (that can be correlated with the execution using the SPARK UI).

Viewing profiling information

The simplest way is to visualize the results with a graphical visualization tool called kcachegrind. In order to do that, install kcachegrind packages (most linux distributions have it installed by default) and it’s python connector pyprof2calltree. From command line run:

pyprof2calltree rdd_<INTERESTING_RDD_ID>.pstats.

Another way is to use the builtin pstats functionality from within python:

import pstats
p = pstats.Stats('restats')


An example code can be found here .