Kedro is a python package which facilitates the prototyping of data pipelines. It aims at enforcing software engineering best practices (separation between I/O and compute, abstraction, templating…). It is specifically useful for machine learning projects since it provides within the same interface interactive objects for the exploration phase, and Command Line Interface (CLI) and configuration files for the production phase. This makes the transition from exploration to production as smooth as possible.
For more details, see Kedro’s official documentation.
Mlflow is a library which manages the lifecycle of machine learning models. Mlflow provides 4 modules:
Mlflow Tracking: This modules focuses on experiment versioning. Its goal is to store all the objects needed to reproduce any code execution. This includes code through version control, but also parameters and artifacts (i.e objects fitted on data like encoders, binarizers…). These elements vary wildly during machine learning experimentation phase.
Mlflowalso enable to track metrics to evaluate runs, and provides a User Interface (UI) to browse the different runs and compare them.
Mlflow Projects: This module provides a configuration files and CLI to enable reproducible execution of pipelines in production phase.
Mlflow Models: This module defines a standard way for packaging machine learning models, and provides built-in ways to serve registered models. Such standardization enable to serve these models across a wide range of tools.
Mlflow Model Registry: This modules aims at monitoring deployed models. The registry manages the transition between different versions of the same model (when the dataset is retrained on new data, or when parameters are updated) while it is in production.
For more details, see Mlflow’s official documentation.
A brief comparison between
Mlflow do not compete in the same field, they provide some overlapping functionalities.
Mlflow is specifically dedicated to machine learning and its lifecycle management, while
Kedro focusing on data pipeline development. Below chart compare the different functionalities:
|I/O configuration files||-
|Compute configuration files||-
|Parameters and data versioning||-
We discuss hereafter how the two libraries compete on the different functionalities and eventually complete each others.
Configuration and prototyping: Kedro 1 - 0 Mlflow¶
Kedro are essentially overlapping on the way they offer a dedicated configuration files for running the pipeline from CLI. However:
Mlflowprovides a single configuration file (the
MLProject) where all elements are declared (data, parameters and pipelines). Its goal is mainly to enable CLI execution of the project, but it is not very flexible. In my opinion, this file is production oriented and is not really intended to use for exploration.
Kedrooffers a bunch of files (
pipeline.py) and their associated abstraction (
Kedrois much more opinionated: each object has a dedicated place (and only one!) in the template. This makes the framework both exploration and production oriented. The downside is that it could make the learning curve a bit sharper since a newcomer has to learn all
Kedrospecifications. It also provides a
kedro-vizplugin to visualize the DAG interactively, which is particularly handy in medium-to-big projects.
Kedrois a clear winner here, since it provides more functionnalities than
Mlflow. It handles very well by design the exploration phase of data science projects when Mlflow is less flexible.
Versioning: Kedro 1 - 1 Mlflow¶
Journal aims at reproducibility, but is not focused on machine learning. The
Journal keeps track of two elements:
the CLI arguments, including on the fly parameters. This makes the command used to run the pipeline fully reproducible.
AbstractVersionedDataSetfor which versioning is activated. It consists in copying the data whom
savemethod of the
AbstractVersionedDataSetis called. This approach suffers from two main drawbacks:
the configuration is assumed immutable (including parameters), which is not realistic ni machine learning projects where they are very volatile. To fix this, the
git shahas been recently added to the
Journal, but it has still some bugs in my experience (including the fact that the current
git shais logged even if the pipeline is ran with uncommitted change, which prevents reproducibility). This is still recent and will likely evolve in the future.
there is no support for browsing old runs, which prevents cleaning the database with old and unused datasets, compare runs between each other…
On the other hand,
distinguishes between artifacts (i.e. any data file), metrics (integers that may evolve over time) and parameters. The logging is very straightforward since there is a one-liner function for logging the desired type. This separation makes further manipulation easier.
offers a way to configure the logging in a database through the
mlflow_tracking_uriparameter. This database-like logging comes with easy querying of different runs through a client (for instance “find the most recent run with a metric at least above a given threshold” is immediate with
Mlflowbut hacky in
comes with a User Interface (UI) which enable to browse / filter / sort the runs, display graphs of the metrics, render plots… This make the run management much easier than in
has a command to reproduce exactly the run from a given
git sha, which is not possible in
Mlflowis a clear winner here, because UI and run querying are must-have for machine learning projects. It is more mature than
Kedrofor versioning and more focused on machine learning.
Model packaging and service: Kedro 1 - 2 Mlflow¶
Kedro offers a way to package the code to make the pipelines callable, but does not manage specifically machine learning models.
Mlflow offers a way to store machine learning models with a given “flavor”, which is the minimal amount of information necessary to use the model for prediction:
a configuration file
all the artifacts, i.e. the necessary data for the model to run (including encoder, binarizer…)
a conda configuration through an
When a stored model meets these requirements,
Mlflow provides built-in tools to serve the model (as an API or for batch prediction) on many machine learning tools (Microsoft Azure ML, Amazon Sagemaker, Apache SparkUDF) and locally.
Mlflowis currently the only tool which adresses model serving. This is currently not the top priority for
Kedro, but may come in the future (through Kedro Server maybe?)
Conclusion: Use Kedro and add Mlflow for machine learning projects¶
In my opinion,
Kedro’s will to enforce software engineering best practice makes it really useful for machine learning teams. It is extremely well documented and the support is excellent, which makes it very user friendly even for people with no computer science background. However, it lacks some machine learning-specific functionalities (better versioning, model service), and it is where
Mlflow fills the gap.