---

Modyn is a data-centric machine learning pipeline orchestrator, i.e., a platform for model training on growing datasets where points get added over time. Check out our blog post for a brief introduction.

⚡️ Quickstart

For deploying and running integration tests, you will need Docker. Furthermore, we use micromamba for local environments and tmuxp for easily managing components panes (optional). For local development, run

# In case you don't have micromamba yet
# macos:
brew install micromamba
# alternatively:
"${SHELL}" <(curl -L micro.mamba.pm/install.sh)

# Start here if you have micromamba already
./scripts/initial_setup.sh
micromamba env create -f ./environment.yml
micromamba activate modyn
pip install -e .
pip install -r dev-requirements.txt

and then ./scripts/python_compliance.sh to check that your local installation of Modyn is functioning.

If you want to run all Modyn components, run

./scripts/run_modyn.sh
tmuxp load tmuxp.yaml # Assumes you have tmuxp, see above for link

For running all integration tests, run

./scripts/run_integrationtests.sh

macOS Installation:: Make sure to run ./scripts/initial_setup.sh as outlined above. If not, installation might fail due to PyTorch not being found. Unfortunately, the PyTorch channel currently does not support macOS.

GPU Installation:: If you want to use a GPU, make sure to install nvidia-docker and confirm to use CUDA on first run of ./scripts/initial_setup.sh. Optionally, if you want to use Apex (require, e.g., for DLRM model), make sure to confirm to install Apex. In this case, having the NVIDIA docker runtime as Docker default runtime is required. The script will try to enable this, if we have sudo privileges on the system. The CUDA version can be adjusted in the scripts/initial_setup.sh file.

Next Steps. Checkout our Example Pipeline guide for an example on how to run a Modyn pipeline. Checkout our Technical Guidelines for some hints on developing Modyn and how to add new data selection and triggering policies. Checkout the Architecture Documentation for an overview of Modyn's components. Last, checkout our full paper on Modyn for more technical background and experiments we ran using Modyn.

Please reach out via Github, Twitter, E-Mail, or any other channel of communication if you are interested in collaborating, have any questions, or have any problems running Modyn.

How to contribute.

🔁 What are dynamic datasets and what is Modyn used for?

ML is is often applied in use cases where training data grows over time, i.e., datasets are growing instead of static. Training must incorporate data changes for high model quality, however this is often challenging and expensive due to large datasets and models. With Modyn, we are actively developing an open-source platform that manages growing datasets at scale and supports pluggable policies for when and what data to train on. Furthermore, we are developing a representative open-source benchmarking suite for ML training on dynamic datasets.

The unit of execution in Modyn is a pipeline. At minimum, a pipeline consists of (1) the model specification, (2) the training dataset and a corresponding byte parsing function that defines how to convert raw sample bytes to model input, (3) the triggering policy, (4) the data selection policy, (5) training hyperparameters such as the the learning rate and batch size, (6) training configuration such as data processing workers and number of GPUs, and (7) the model storage policy, i.e., a definition how the models are compressed and stored. Checkout our Example Pipeline guide for an example on how to run a Modyn pipeline.

Modyn allows researchers to explore triggering and data selection policies (see Technical Guidelines on how to add new policies to Modyn), while alleviating the burdens of managing large growing datasets and orchestrating recurring training jobs. However, we strive towards usage of Modyn in practical environments as well. We welcome input from both research and practice.

✉️ About

Modyn is being developed at the Efficient Architectures and Systems Lab (EASL) at the ETH Zurich Systems Group. Please reach out to mboether [at] inf [­dot] ethz [dot] ch or open an issue on Github if you have any questions or inquiry related to Modyn and its usage.

Paper / Citation

If you use Modyn, please cite our SIGMOD'25 paper:

@inproceedings{Bother2025Modyn,
  author = {B\"{o}ther, Maximilian and Robroek, Ties and Gsteiger, Viktor and Ma, Xianzhe and T\"{o}z\"{u}n, P{\i}nar and Klimovic, Ana},
  title = {Modyn: Data-Centric Machine Learning Pipeline Orchestration},
  booktitle = {Proceedings of the Conference on Management of Data (SIGMOD)},
  year = {2025},
}