The general purpose micro-orchestration framework for building dataflows from python functions. Express data, ML, LLM pipelines/workflows, and web requests in a simple declarative manner.
Hamilton is a novel paradigm for specifying a flow of delayed execution in python. It works on python objects of any type and dataflows of any complexity. Core to the design of Hamilton is a clear mapping of function name to artifact, allowing you to quickly grok the relationship between the code you write and the data you produce.
This paradigm makes modifications easy to build and track, ensures code is self-documenting, and makes it natural to unit test your data transformations. When connected together, these functions form a Directed Acyclic Graph (DAG), which the Hamilton framework can execute, optimize, and report on.
Note: Hamilton describes DAGs. If you're looking for something to handle loops or conditional edges (say, for a human-in-the-loop application like a chatbot or agent), you might appreciate Burr -- it integrates well with any python library (including Hamilton!).
✅ Model a dataflow -- If you can model your problem as a DAG in python, Hamilton is the cleanest way to build it.
✅ Unmaintainable spaghetti code -- Hamilton dataflows are unit testable, self-documenting, and provide lineage.
✅ Long iteration/experimentation cycles -- Hamilton provides a clear, quick, and methodical path to debugging/modifying/extending your code.
✅ Reusing code across contexts -- Hamilton encourages code that is independent of infrastructure and can run regardless of execution setting.
❌ Provisioning infrastructure -- you want a macro-orchestration system (see airflow, kubeflow, sagemaker, etc...).
❌ Doing your ML for you -- we organize your code, BYOL (bring your own libraries).
❌ Tracking execution + associated artifacts -- Hamilton is lightweight, but if this is important to you see the DAGWorks product.
See the table below for more specifics/how it compares to other common tooling.
Here are common things that Hamilton is compared to, and how Hamilton compares to them.
| Feature | Hamilton | Macro orchestration systems (e.g. Airflow) | Feast | dbt | Dask |
|---|---|---|---|---|---|
| Python 3.8+ | ✅ | ✅ | ✅ | ✅ | ✅ |
| Helps you structure your code base | ✅ | ❌ | ❌ | ✅ | ❌ |
| Code is always unit testable | ✅ | ❌ | ❌ | ❌ | ❌ |
| Documentation friendly | ✅ | ❌ | ❌ | ❌ | ❌ |
| Can visualize lineage easily | ✅ | ❌ | ❌ | ✅ | ✅ |
| Is just a library | ✅ | ❌ | ❌ | ❌ | ✅ |
| Runs anywhere python runs | ✅ | ❌ | ❌ | ❌ | ✅ |
| Built for managing python transformations | ✅ | ❌ | ❌ | ❌ | ❌ |
| Can model GenerativeAI/LLM based workflows | ✅ | ❌ | ❌ | ❌ | ❌ |
| Replaces macro orchestration systems | ❌ | ✅ | ❌ | ❌ | ❌ |
| Is a feature store | ❌ | ❌ | ✅ | ❌ | ❌ |
| Can model transforms at row/column/object/dataset level | ✅ | ❌ | ❌ | ❌ | ❌ |
If you don't want to install anything to try Hamilton, we recommend trying www.tryhamilton.dev. Otherwise, here's a quick getting started guide to get you up and running in less than 15 minutes. If you need help join our slack community to chat/ask Qs/etc. For the latest updates, follow us on twitter!
Requirements:
- Python 3.8+
To get started, first you need to install hamilton. It is published to pypi under sf-hamilton:
pip install sf-hamilton
Note: to use the DAG visualization functionality, you should instead do:
pip install "sf-hamilton[visualization]"
While it is installing we encourage you to start on the next section.
Note: the content (i.e. names, function bodies) of our example code snippets are for illustrative purposes only, and don't reflect what we actually do internally.
Hamilton is a new paradigm when it comes to creating, um, dataframes (let's use dataframes as an example, otherwise you can create ANY python object). Rather than thinking about manipulating a central dataframe, as is normal in some data engineering/data science work, you instead think about the column(s) you want to create, and what inputs are required. There is no need for you to think about maintaining this dataframe, meaning you do not need to think about any "glue" code; this is all taken care of by the Hamilton framework.
For example rather than writing the following to manipulate a central dataframe object df:
df['col_c'] = df['col_a'] + df['col_b']you write
def col_c(col_a: pd.Series, col_b: pd.Series) -> pd.Series:
"""Creating column c from summing column a and column b."""
return col_a + col_bIn diagram form:
The Hamilton framework will then be able to build a DAG from this function definition.
So let's create a "Hello World" and start using Hamilton!
By now, you should have installed Hamilton, so let's write some code.
- Create a file
my_functions.pyand add the following functions:
import pandas as pd
def avg_3wk_spend(spend: pd.Series) -> pd.Series:
"""Rolling 3 week average spend."""
return spend.rolling(3).mean()
def spend_per_signup(spend: pd.Series, signups: pd.Series) -> pd.Series:
"""The cost per signup in relation to spend."""
return spend / signupsThe astute observer will notice we have not defined spend or signups as functions. That is okay,
this just means these need to be provided as input when we come to actually wanting to create a dataframe.
Note: functions can take or create scalar values, in addition to any python object type.
- Create a
my_script.pywhich is where code will live to tell Hamilton what to do:
import sys
import logging
import importlib
import pandas as pd
from hamilton import driver
logging.basicConfig(stream=sys.stdout)
initial_columns = { # load from actuals or wherever -- this is our initial data we use as input.
# Note: these do not have to be all series, they could be scalar inputs.
'signups': pd.Series([1, 10, 50, 100, 200, 400]),
'spend': pd.Series([10, 10, 20, 40, 40, 50]),
}
# we need to tell hamilton where to load function definitions from
module_name = 'my_functions'
module = importlib.import_module(module_name) # or we could just do `import my_functions`
dr = driver.Driver(initial_columns, module) # can pass in multiple modules
# we need to specify what we want in the final dataframe.
output_columns = [
'spend', # or module.spend
'signups', # or module.signups
'avg_3wk_spend', # or module.avg_3wk_spend
'spend_per_signup', # or module.spend_per_signup
]
# let's create the dataframe!
# if you only did `pip install sf-hamilton` earlier:
df = dr.execute(output_columns)
# else if you did `pip install "sf-hamilton[visualization]"` earlier:
# dr.visualize_execution(output_columns, './my-dag.dot', {})
print(df)- Run my_script.py
python my_script.py
You should see the following output:
spend signups avg_3wk_spend spend_per_signup
0 10 1 NaN 10.000
1 10 10 NaN 1.000
2 20 50 13.333333 0.400
3 40 100 23.333333 0.400
4 40 200 33.333333 0.200
5 50 400 43.333333 0.125
You should see the following image if you ran dr.visualize_execution(output_columns, './my-dag.dot', {"format": "png"}, orient="TB"):
Note: we treat displaying Inputs in a special manner for readability in our visualizations. So you'll likely see input
nodes repeated.
Congratulations - you just created your Hamilton dataflow that created a dataframe!
We have a growing list of examples showcasing how one might use Hamilton. You currently have two places to find them:
- The Hamilton Dataflow Hub -- which makes it easy to pull and then modify code.
- The
examples/folder in this repository.
For the Hub, this will contain user contributed dataflows, e.g. text_summarization, forecasting, data processing, that will be continually added to.
For the examples/ directory, you'll have to copy/fork the repository to run them.
E.g.
- Hello world
- Scaling on to Ray, Dask, or Pandas on Spark
- Training a model with scikit-learn
- Doing air quality analysis solely in numpy
We also have a docker container that contains some of these examples so you can pull that and run them locally. See the examples folder README for details.
This repository is maintained by the original creators of Hamilton, who have since founded DAGWorks inc., a company largely dedicated to building and maintaining the Hamilton library. We decided to fork the original because Stitch Fix did not want to transfer ownership to us; we had grown the star count in the original repository to 893: 
before forking.
For the backstory on how Hamilton came about, see the original Stitch Fix blog post!.
We have a small but active community on slack. Come join us!
Hamilton is released under the BSD 3-Clause Clear License.
- Stitch Fix
- UK Government Digital Services
- IBM
- British Cycling
- PNNL
- ORNL
- Federal Reserve Board
- Joby Aviation
- Two
- Transfix
- Railofy
- Habitat Energy
- KI-Insurance
- Ascena Retail
- Opendoor
- NaroHQ
- Rippling
- AB-InBev
- EquipmentShare
- Everstream.ai
- Flectere
- F33.ai
To add your company, make a pull request to add it here.
We take contributions, large and small. We operate via a Code of Conduct and expect anyone contributing to do the same.
To see how you can contribute, please read our contributing guidelines and then our developer setup guide.
- Lineage + Hamilton in 10 minutes
- (Organic Content) The perks of creating dataflows with Hamilton by Thierry Jean
- Developing Scalable Feature Engineering DAGs with Metaflow & Hamilton
- Tidy Production Pandas with Hamilton
- Towards Data Science post on backstory & introduction.
- How to use Hamilton with Pandas in 5 minutes.
- How to iterate with Hamilton in a Notebook.
- Original Stitch Fix Post.
- Extension Stitch Fix Post.
- Hamilton: a python micro-framework for data/feature engineering at Stitch Fix - 40 mins:
- Hamilton: a python micro-framework for tidy scalable pandas - ~20 mins:
We'd appreciate citing Hamilton by referencing one of the following:
@inproceedings{DBLP:conf/vldb/KrawczykI22,
author = {Stefan Krawczyk and Elijah ben Izzy},
editor = {Satyanarayana R. Valluri and Mohamed Za{\"{\i}}t},
title = {Hamilton: a modular open source declarative paradigm for high level
modeling of dataflows},
booktitle = {1st International Workshop on Composable Data Management Systems,
CDMS@VLDB 2022, Sydney, Australia, September 9, 2022},
year = {2022},
url = {https://cdmsworkshop.github.io/2022/Proceedings/ShortPapers/Paper6\_StefanKrawczyk.pdf},
timestamp = {Wed, 19 Oct 2022 16:20:48 +0200},
biburl = {https://dblp.org/rec/conf/vldb/KrawczykI22.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@inproceedings{CEURWS:conf/vldb/KrawczykIQ22,
author = {Stefan Krawczyk and Elijah ben Izzy and Danielle Quinn},
editor = {Cinzia Cappiello and Sandra Geisler and Maria-Esther Vidal},
title = {Hamilton: enabling software engineering best practices for data transformations via generalized dataflow graphs},
booktitle = {1st International Workshop on Data Ecosystems co-located with 48th International Conference on Very Large Databases (VLDB 2022)},
pages = {41--50},
url = {https://ceur-ws.org/Vol-3306/paper5.pdf},
year = {2022}
}
Hamilton is an ambitious project to provide a unified way to describe any dataflow, independent of where it runs. You can find currently support integrations and high-level roadmap below. Please reach out via slack or email (stefan / elijah at dagworks.io) to contribute or share feedback!
- Any python object type! E.g. Pandas, Spark dataframes, Dask dataframes, Ray datasets, Polars, dicts, lists, primitives, your custom objects, etc.
- data processing
- feature engineering
- model training
- LLM application workflows
- all of them together
See the data quality docs.
- Ability to define data quality check on an object.
- Pandera schema integration.
- Custom object type validators.
- Integration with other data quality libraries (e.g. Great Expectations, Deequ, whylogs, etc.)
- Open telemetry/tracing plugin.
- Checkpoint caching (e.g. save a function's result to disk, independent of input) - WIP.
- Finergrained caching (e.g. save a function's result to disk, dependent on input).
- Runs anywhere python runs. E.g. airflow, prefect, dagster, kubeflow, sagemaker, jupyter, fastAPI, snowpark, etc.
Specific integrations with other systems where we help you write code that runs on those systems.
- Delegate function execution to Ray.
- Function grouping (e.g. fuse multiple functions into a single Ray task)
- Delegate function execution to Dask.
- Function grouping (e.g. fuse multiple functions into a single Dask task)
- Pandas on spark integration (via GraphAdapter)
- PySpark native UDF map function integration (via GraphAdapter)
- PySpark native aggregation function integration
- PySpark join, filter, groupby, etc. integration
- Packaging functions for Snowpark
- Numba integration
- Generate code to execute on a custom topology, e.g. microservices, etc.
- Generating Airflow | Prefect | Metaflow | Dagster | Kubeflow Pipelines | Sagemaker Pipelines | etc from Hamilton.
- Plugins for common MLOps/DataOps tools: MLFlow, DBT, etc.
- Depth first search traversal
- Async function support via AsyncDriver
- Parallel walk over a generator
- Python multiprocessing execution (still in beta)
- Python threading support
- Grouping of nodes into tasks for efficient parallel computation
- Breadth first search traversal
- Sequential walk over a generator
- At Driver instantiation time, using configuration/modules and
@config.when. - With
@resolveduring Driver instantiation time.
In general we prescribe the following:
- Ensure you understand Hamilton Basics.
- Familiarize yourself with some of the Hamilton decorators. They will help keep your code DRY.
- Start creating Hamilton Functions that represent your work. We suggest grouping them in modules where it makes sense.
- Write a simple script so that you can easily run things end to end.
- Join our Slack community to chat/ask Qs/etc.
For the backstory on Hamilton we invite you to watch a roughly-9 minute lightning talk on it that we gave at the apply conference: video, slides.
If you're using Hamilton, it's likely that you'll need to migrate some code. Here are some useful tricks we found to speed up that process.
Live templates are a cool feature and allow you to type in a name which expands into some code.
E.g. For example, we wrote one to make it quick to stub out Hamilton functions: typing graphfunc would turn into ->
def _(_: pd.Series) -> pd.Series:
""""""
return _Where the blanks are where you can tab with the cursor and fill things in. See your pycharm preferences for setting this up.
If you are doing a lot of repetitive work, one might consider multiple cursors. Multiple cursors allow you to do things on multiple lines at once.
To use it hit option + mouse click to create multiple cursors. Esc to revert back to a normal mode.
By default, when using Hamilton, it collects anonymous usage data to help improve Hamilton and know where to apply development efforts.
We capture three types of events: one when the Driver object is instantiated, one when the execute() call on the Driver object completes, and one for most Driver object function invocations.
No user data or potentially sensitive information is or ever will be collected. The captured data is limited to:
- Operating System and Python version
- A persistent UUID to indentify the session, stored in ~/.hamilton.conf.
- Error stack trace limited to Hamilton code, if one occurs.
- Information on what features you're using from Hamilton: decorators, adapters, result builders.
- How Hamilton is being used: number of final nodes in DAG, number of modules, size of objects passed to
execute(), the name of the Driver function being invoked.
If you're worried, see telemetry.py for details.
If you do not wish to participate, one can opt-out with one of the following methods:
- Set it to false programmatically in your code before creating a Hamilton driver:
from hamilton import telemetry telemetry.disable_telemetry()
- Set the key
telemetry_enabledtofalsein ~/.hamilton.conf under theDEFAULTsection:[DEFAULT] telemetry_enabled = False - Set HAMILTON_TELEMETRY_ENABLED=false as an environment variable. Either setting it for your shell session:
or passing it as part of the run command:
export HAMILTON_TELEMETRY_ENABLED=falseHAMILTON_TELEMETRY_ENABLED=false python NAME_OF_MY_DRIVER.py
For the hamilton UI you jmust use the environment variable method prior to running docker compose.
- Stefan Krawczyk (@skrawcz)
- Elijah ben Izzy (@elijahbenizzy)
- Danielle Quinn (@danfisher-sf)
- Rachel Insoft (@rinsoft-sf)
- Shelly Jang (@shellyjang)
- Vincent Chu (@vslchusf)
- Christopher Prohm (@chmp)
- James Lamb (@jameslamb)
- Avnish Pal (@bovem)
- Sarah Haskins (@frenchfrywpepper)
- Thierry Jean (@zilto)
- Michał Siedlaczek (@elshize)
- Benjamin Hack (@benhhack)
- Bryan Galindo (@bryangalindo)
- Jordan Smith (@JoJo10Smith)
- Roel Bertens (@roelbertens)
- Swapnil Delwalkar (@swapdewalkar)
- Fran Boon (@flavour)
- Tom Barber (@buggtb)
- Konstantin Tyapochkin (@tyapochkin)
- Nils Olsson (@nilsso)
- Michał Siedlaczek (@elshize)
- Alaa Abedrabbo (@AAbedrabbo)
- Shreya Datar (@datarshreya)
- Baldo Faieta (@baldofaieta)
- Anwar Brini (@AnwarBrini)
- Gourav Kumar (@gms101)
- Amos Aikman (@amosaikman)
- Ankush Kundaliya (@akundaliya)
- David Weselowski (@j7zAhU)
- Peter Robinson (@Peter4137)
- Seth Stokes (@sT0v
- Louis Maddox (@lmmx)
- Stephen Bias (@s-ducks)
- Anup Joseph (@AnupJoseph)
- Jan Hurst (@janhurst)
- Flavia Santos (@flaviassantos)
- Nicolas Huray (@nhuray)
- Manabu Niseki (@ninoseki)
![ellipsis-dev[bot] avatar](https://avatars.githubusercontent.com/in/64358?v=4 "ellipsis-dev[bot]")
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent