Local Execution

Execute KFP pipelines locally

Overview

KFP supports executing components and pipelines locally, enabling a tight development loop before running your code remotely.

Executing components and pipelines locally is easy. Simply initialize a local session using local.init(), then call the component or pipeline like a normal Python function. KFP will log information about the execution. Once execution completes, you can access the task outputs just as you would when composing a pipeline; the only difference is that the outputs are now materialized values, not references to future outputs.

In the following example, we use the DockerRunner type. Runner types are covered in more detail below.

from kfp import local
from kfp import dsl

local.init(runner=local.DockerRunner())

@dsl.component
def add(a: int, b: int) -> int:
    return a + b

# run a single component
task = add(a=1, b=2)
assert task.output == 3

# or run it in a pipeline
@dsl.pipeline
def math_pipeline(x: int, y: int, z: int) -> int:
    t1 = add(a=x, b=y)
    t2 = add(a=t1.output, b=z)
    return t2.output

pipeline_task = math_pipeline(x=1, y=2, z=3)
assert pipeline_task.output == 6

Similarly, you can create artifacts and read the contents:

from kfp import local
from kfp import dsl
from kfp.dsl import Output, Artifact
import json

local.init(runner=local.SubprocessRunner())

@dsl.component
def add(a: int, b: int, out_artifact: Output[Artifact]):
    import json

    result = json.dumps(a + b)

    with open(out_artifact.path, 'w') as f:
        f.write(result)

    out_artifact.metadata['operation'] = 'addition'


task = add(a=1, b=2)
# can read artifact contents
with open(task.outputs['out_artifact'].path) as f:
    contents = f.read()
assert json.loads(contents) == 3
assert task.outputs['out_artifact'].metadata['operation'] == 'addition'

By default, KFP will raise an exception if your component exits with a failure status. You can toggle this behavior using raise_on_error. You can also specify a new local “pipeline root” using pipeline_root. This is the local directory to which component outputs, including artifacts, are written.

local.init(runner=...,
           raise_on_error=False,
           pipeline_root='~/my/component/outputs')

Local runners

You can choose from two local runner types which indicate how and where your component should be executed: DockerRunner and SubprocessRunner.

When invoking components locally using the DockerRunner, the task will execute in a container.

from kfp import local

local.init(runner=local.DockerRunner())

Since the local DockerRunner executes each task in a separate container, the DockerRunner:

It is recommended to use DockerRunner whenever possible.

When you use the DockerRunner, KFP mounts your local pipeline root to the container to write outputs outside of the container. This means that your component outputs will still be available for inspection even after the container exits.

The DockerRunner requires Docker to be installed, but requires essentially no knowledge of Docker to use.

Alternative: SubprocessRunner

The SubprocessRunner is recommended when executing components in local environments where Docker cannot be installed, such as in some notebook environments.

from kfp import local

local.init(runner=local.SubprocessRunner())

Since SubprocessRunner runs your code in a subprocess, the SubprocessRunner:

  • Offers less local runtime environment isolation than the DockerRunner
  • Does not support custom images or easily support tasks with complex environment dependencies
  • Only allows execution of Lightweight Python Component

By default, the SubprocessRunner will install your dependencies into a virtual environment. This is recommended, but can be disabled by setting use_venv=False:

from kfp import local

local.init(runner=local.SubprocessRunner(use_venv=False))

Limitations

Local execution is designed to help quickly test components and pipelines locally before testing in a remote environment.

Local execution comes with several limitations:

  • Local execution does not feature optimizations and additional features such as caching, retry, etc. While these feature are important for production pipelines, they are less critical for a local testing environment. You will find that task methods like .set_retry, .set_caching_options, etc. have no effect locally.
  • Local execution makes simple assumptions about the resources available on your machine. Local execution does not support specifying resource requests/limits/affinities related to memory, cores, accelerators, etc. You will find that task methods like .set_memory_limit, .set_memory_request, .set_accelerator_type etc. have no effect locally.

While local pipeline execution has full support for sequential and nested pipelines, it does not yet support dsl.Condition, dsl.ParallelFor, or dsl.ExitHandler.

Feedback

Was this page helpful?


Last modified February 21, 2024: local pipeline execution docs (#3671) (00546e4)