Modular pipelines

What are modular pipelines?

In many typical Kedro projects, a single (“main”) pipeline increases in complexity as the project evolves. To keep your project fit for purpose, we recommend that you create modular pipelines, which are logically isolated and can be reused. Modular pipelines are easier to develop, test and maintain, and are portable so they can be copied and reused between projects.

Modular pipelines allow you to instantiate pipelines multiple times, while allowing the user to override inputs/outputs/parameters. They are reusable within the same codebase, and shareable across projects via micro-packaging. This is the modern way to use Kedro, and will change the way you think about your pipelines.


The Kedro project visualised below is representative of one that might be seen in the real world. It takes full advantage of modular pipelines for Data Ingestion, Feature Engineering, Reporting and Train Evaluation (which even includes nested instances).

Key concepts

In this section, you will learn about how to take advantage of modular pipelines. The key points are listed below:

  1. A modular pipeline is defined by its folder structure

    • You can generate this file structure with the CLI command kedro pipeline create <pipeline_name>.

    • The folder structure keeps things isolated and encourages portability.

  2. Modular pipelines are designed to be portable and reusable

    • It’s possible to re-use the same pipeline multiple times within the same project (with different inputs/outputs or parameters).

    • You can also share pipelines across codebases via micro-packaging.

  3. The kedro.pipeline.modular_pipeline.pipeline wrapper method unlocks the real power of modular pipelines

    • Applying namespaces allows you to simplify your mental model and isolate ‘within pipeline’ processing steps.

    • Kedro-Viz is able to accelerate development by rendering namespaced pipelines as collapsible ‘super nodes’.

How do I create a modular pipeline?

You can use a project-specific CLI command to create a modular pipeline. The pipeline name must adhere to Python convention.

kedro pipeline create <pipeline_name>


For the full list of available CLI options, you can always run kedro pipeline create --help for more information.

What does the kedro pipeline create do?

Running the kedro pipeline create command adds boilerplate pipeline folders and files for the created pipeline to your project. For your convenience, Kedro gives you a pipeline-specific,, parameters and appropriate tests structure. You also don’t have to add those pesky files yourself, which is handy 😅. You can see the generated folder structure below:

Click to see the generated folder structure
├── conf
│   └── base
│       └── parameters
│           └── {{pipeline_name}}.yml  <-- Pipeline specific parameters
└── src
    ├── my_project
    │   ├──
    │   ├── pipelines
    │   |   ├──
    │   |   └── {{pipeline_name}}      <-- This folder defines the modular pipeline
    │   |       ├──          <-- To store pipeline specific documentation
    │   |       ├──        <-- So that Python treats this pipeline as a module
    │   |       ├──           <-- To declare your nodes
    │   |       └──        <-- To structure the pipeline itself
    |   └──      <-- Does NOT automatically update the registry
    └── tests
        └── pipelines
            └── {{pipeline_name}}      <-- Pipeline specific tests

If you want to do the reverse and remove a modular pipeline, you can use kedro pipeline delete <pipeline_name> to do so.

Ensuring portability

Modular pipelines are shareable between Kedro codebases via micro-packaging, but you must follow a couple of rules to ensure portability:

  • Modular pipelines should not depend on the main Python package, as this would break portability to another project.

  • Catalog references are not packaged when sharing/consuming modular pipelines, i.e. the catalog.yml file is not packaged.

  • Kedro will only look for top-level configuration in conf/; placing a configuration folder within the pipeline folder will have no effect.

  • We recommend that you document the configuration required (parameters and catalog) in the local file for any downstream consumers.

Providing modular pipeline specific dependencies

  • A modular pipeline might have external dependencies specified in a local requirements.txt file.

  • Pipeline specific dependencies are scooped up during the micro-packaging process.

  • These dependencies are not currently installed by the kedro install command, and must be manually installed.

Using the modular pipeline() wrapper to provide overrides

This wrapper really unlocks the power of modular pipelines.

  • It allows you to start instantiating the same pipeline multiple times.

  • These will be static in terms of structure, but dynamic in terms of inputs/outputs/parameters.

  • It also allows you to simplify both your mental models, and Kedro-Viz visualisations via namespaces.

from kedro.pipeline.modular_pipeline import pipeline

The pipeline() wrapper method takes the following arguments:

Keyword argument



The Pipeline object you want to wrap


Any overrides provided to this instance of the underlying wrapped Pipeline object


Any overrides provided to this instance of the underlying wrapped Pipeline object


Any overrides provided to this instance of the underlying wrapped Pipeline object


The namespace that will be encapsulated by this pipeline instance

Combining disconnected pipelines

Sometimes two pipelines must be connected, but do not share any catalog dependencies. The wrapper can be used to solve that.

Click here to see a worked example

In this example, there is a lunch_pipeline, which makes us lunch. The ‘verbs’, defrost and eat, are Python functions and the inputs/outputs are food at different points of the process (frozen, thawed and food).

cook_pipeline = pipeline(
        node(func=defrost, inputs="frozen_veg", outputs="veg"),
        node(func=grill, inputs="veg", outputs="grilled_veg"),

lunch_pipeline = pipeline([node(func=eat, inputs="food", outputs=None)])

cook_pipeline + lunch_pipeline

This combination will visualise since it’s valid pre-runtime, but it will not run since food is not an output of the cook_pipeline because the output of the cook_pipeline is grilled_veg:


  • Combining cook_pipeline + lunch_pipeline will not work since food doesn’t exist as an output of the cook_pipeline.

  • In this case, we will need to map grilled_veg to the expected input of food.

The wrapper allows us to provide a mapping and fix this disconnect.

from kedro.pipeline.modular_pipeline import pipeline

prep_pipeline = pipeline(pipe=cook_pipeline, outputs={"grilled_veg": "food"})

meal_pipeline = prep_pipeline + lunch_pipeline

Providing this input/output override will join up the pipeline nicely:



In this example we have used the + operator to join two pipelines. You can also use sum() or pass a list of pipelines to the pipe argument.

Using a modular pipeline multiple times

Reusing pipelines for slightly different purposes can be a real accelerator for teams and organisations when they reach a certain scale. In the real world, one could imagine pipelines with responsibilities like profiling or feature engineering being reused within the same project or even across projects via micro-packaging.

  • In an ideal world, we would like to use the cook_pipeline twice as you would defrost and grill multiple meals beyond the veg currently hard-coded.

  • Namespaces allow you to ‘instantiate’ the same pipeline multiple times and keep operations isolated.

  • Like one provides arguments to a class’ constructor, you can provide overriding inputs/outputs/parameters to the pipeline() wrapper.

Click here to see a worked example
cook_pipeline = pipeline(
        node(func=defrost, inputs="frozen_veg", outputs="veg", name="defrost_node"),
        node(func=grill, inputs="veg", outputs="grilled_veg"),

eat_breakfast_pipeline = pipeline(
    [node(func=eat_breakfast, inputs="breakfast_food", outputs=None)]
eat_lunch_pipeline = pipeline([node(func=eat_lunch, inputs="lunch_food", outputs=None)])

cook_pipeline + eat_breakfast_pipeline + eat_lunch_pipeline

If we visualise the snippet above, we see a disjointed pipeline:

  • We need to “defrost” two different types of food via different pipelines.

  • We cannot use the cook_pipeline twice because the internal dataset names will conflict.

  • Mapping all datasets via the pipeline() wrapper will also cause conflicts.

cook no namespace

Adding namespaces solves this issue:

cook_breakfast_pipeline = pipeline(
    inputs="frozen_veg",  # inputs stay the same, don't namespace
    outputs={"grilled_veg": "breakfast_food"},
cook_lunch_pipeline = pipeline(
    inputs="frozen_veg",  # inputs stay the same, don't namespace
    outputs={"grilled_veg": "lunch_food"},

final_pipeline = (
    + eat_breakfast_pipeline
    + cook_lunch_pipeline
    + eat_lunch_pipeline
  • namespace="lunch" renames all datasets and nodes, prefixing them with "lunch.".

  • The datasets that we explicitly “freeze” (frozen_veg) or remap (grilled_veg) are not affected/prefixed.

  • Remapping free outputs is required since “breakfast_food” and “lunch_food” are the names expected by the eat_breakfast_pipeline and eat_lunch_pipeline respectively.

  • The resulting pipeline now has two separate nodes, breakfast.defrost_node and lunch.defrost_node.

  • Also two separate datasets breakfast.veg and lunch.veg connect the nodes inside the pipelines, causing no confusion between them.


  • Visualising the final_pipeline highlights how namespaces become ‘super nodes’ which encapsulate the wrapped pipeline.

  • This example demonstrates how we can reuse the same cook_pipeline with slightly different arguments.

  • Namespaces can also be arbitrarily nested with the . character.


Parameter references (params: and parameters) will not be namespaced.

How to use a modular pipeline with different parameters

Mapping parameter values is very similar to the way we map inputs and outputs.

Click here to see a worked example
  • We instantiate the template_pipeline twice, but pass in different parameters.

  • input1 and input2 are ‘frozen’ and thus shared in both instances.

  • params:override_me does not actually exist and is designed to be overridden in both cases.

  • Providing a namespace isolates the intermediate operation and visualises nicely.

template_pipeline = pipeline(
            inputs=["input1", "input2", "params:override_me"],

alpha_pipeline = pipeline(
    inputs={"input1", "input2"},
    parameters={"params:override_me": "params:alpha"},

beta_pipeline = pipeline(
    inputs={"input1", "input2"},
    parameters={"params:override_me": "params:beta"},

final_pipeline = alpha_pipeline + beta_pipeline