Raptor frees data scientists and ML engineers to build and deploy operational models and ML-driven functionality, without learning backend engineering.
It compiles your python research code to production artifacts, and takes care of the engineering concerns such as scalability and reliability using best-practices on Kubernetes.
Getting started in 5 minutes » · Report a Bug · Request a Feature
Raptor frees data scientists and ML engineers to focus on the data science and research work, and build operational models and ML-driven functionality without learning backend engineering. Focus on what you're good at, increase your end-to-end velocity, and close the gap between research and production.
With Raptor, you can export your Python research code as standard production artifacts, and deploy them to Kubernetes. Once they deployed, Raptor optimizes data processing and feature calculation for production, deploys models to Sagemaker or Docker containers, and connects to your production data sources, scaling, high availability, caching, monitoring, and all other backend concerns.
Raptor is made by and for data scientists and ML engineers. We know how hard it is to build and deploy models to be an integral part of your products, and we want to make it easier.
Before Raptor, data scientists had to work closely with backend engineers to build a "production version" of their work: connect to data sources, transform their data with Flink/Spark or even Java, create APIs, dockerizing the model, handle scaling and high availability, and more.
With Raptor, data scientists can focus only on their research and model development, then export their work to production. Raptor takes care of the rest, including connecting to data sources, transforming the data, deploying and connecting the model, etc. This means data scientists can focus on what they do best, and Raptor handles the rest.
- Focus on your work: Raptor frees data scientists and ML engineers to focus on the model, without learning backend engineering. Stop worrying about the engineering concerns, and focus on what you're good at.
- Eliminate serving/training skew: You can use the same code for training and production to avoid training serving skew.
- Real-time/on-demand: Raptor optimizes feature calculations and predictions to be performed at the time of the request.
- Seamless caching and storage: Raptor uses an integrated caching system, and store your historical data for training purposes. So you won't need any other data storage system such as "Feature Store".
- Turns data science work into production artifacts: Raptor implements best-practice functionalities of Kubernetes solutions, such as scaling, health, auto-recovery, monitoring, logging, and more.
- Integrates with R&D team: Raptor extends existing DevOps tools and infrastructure and allows you to connect your ML research to the rest of your organization's R&D ecosystem, utilizing tools such as CI/CD and monitoring.
To start, install Raptor LabSDK. The LabSDK is a Python package that help you develop models and features in notebooks or IDEs.
pip install raptor-labsdkimport pandas as pd
from raptor import *
from typing_extensions import TypedDict
@data_source(
training_data=pd.read_csv(
'https://gist.githubusercontent.com/AlmogBaku/8be77c2236836177b8e54fa8217411f2/raw/hello_world_transactions.csv'),
production_config=StreamingConfig()
)
class BankTransaction(TypedDict):
customer_id: str
amount: float
timestamp: str
# Define features 🧪
@feature(keys='customer_id', data_source=BankTransaction)
@aggregation(function=AggregationFunction.Sum, over='10h', granularity='1h')
def total_spend(this_row: BankTransaction, ctx: Context) -> float:
"""total spend by a customer in the last hour"""
return this_row['amount']
@feature(keys='customer_id', data_source=BankTransaction)
@freshness(max_age='5h', max_stale='1d')
def amount(this_row: BankTransaction, ctx: Context) -> float:
"""total spend by a customer in the last hour"""
return this_row['amount']
# Train the model 🤓
@model(
keys='customer_id',
input_features=['total_spend+sum'],
input_labels=[amount],
model_framework='sklearn',
model_server='sagemaker-ack',
)
@freshness(max_age='1h', max_stale='100h')
def amount_prediction(ctx: TrainingContext):
from sklearn.linear_model import LinearRegression
df = ctx.features_and_labels()
trainer = LinearRegression()
trainer.fit(df[ctx.input_features], df[ctx.input_labels])
return trainer
amount_prediction.export() # Export to production 🎉This will generate a bunch of artifacts in the out directory. The out directory also includes a Makefile that can
be used for integration in any CI/CD pipeline, or even invoked manually.
Traditional MLOps platforms are focused on managing the ML resources lifecycle and are not designed for building operational models and features. Raptor is designed for building operational models and features, and can be integrated with MLOps platforms.
Feature store is a data storage system that stores pre-computed features for training and online purposes. That means you need to orchestrate the pre-computation of the features, store them, connect them to your model, and write ad-hoc backend code.
Raptor takes a radically different approach. You focus on the model, and Raptor takes care of the rest. Raptor has a built-in caching system that allows you to achieve similar results to a feature store but without the need to orchestrate the data pipeline and the model deployment directly.
Model servers are designed for serving models in production. They are not designed for building models and features for production. In fact, Raptor integrates seamlessly with Model Servers(such as Sagemaker, BentoML, etc.) to serve your models.
The work with Raptor starts in your research phase in your notebook or IDE. Raptor allows you to write your ML work in a translatable way for production purposes.
Models and Features in Raptor are composed of a declarative part(via Python's decorators) and a function code. This way, Raptor can translate the heavy-lifting engineering concerns(such as aggregations or caching) by implementing the "declarative part", and optimizing the implementation for production.
After you are satisfied with your research results, "export" these definitions, and deploy it to Kubernetes using standard tools; Once deployed, Raptor Core(the server-side part) is extending Kubernetes with the ability to implement them. It takes care of the engineering concerns by managing and controlling Kubernetes-native resources such as deployments to connect your production data sources and run your business logic at scale.
You can read more about Raptor's architecture in the docs.
Raptor installation is not required for training purposes. You only need to install Raptor when deploying to production (or staging).
Learn more about production installation at the docs.
- Kubernetes cluster (including EKS, GKE, etc.)
- Redis server (> 2.8.9)
- Optional: Snowflake or S3 bucket (to record historical data for retraining purposes)
- S3 historical storage plugins
- S3 storing
- S3 fetching data - Spark
- Deploy models to model servers
- Sagemaker ACK
- VertexAI
- Seldon
- Kubeflow
- KFServing
- Standalone
- Large-scale training
- Support more data sources
- Kafka
- GCP Pub/Sub
- Rest
- Snowflake
- BigQuery
- gRPC
- Redis
- Postgres
- GraphQL
See the open issues for a full list of proposed features (and known issues) .
Contributions make the open-source community a fantastic place to learn, inspire, and create. Any contributions you make are greatly appreciated (not only code! but also documenting, blogging, or giving us feedback) 😍.
Please fork the repo and create a pull request if you have a suggestion. You can also simply open an issue and choose " Feature Request" to give us some feedback.
Don't forget to give the project a star! ⭐️
For more information about contributing code to the project, read the CONTRIBUTING.md file.
Distributed under the Apache2 License. Read the LICENSE file for more information.
You can join the Raptor community on Slack, follow us on Twitter, and participate in the issues and pull requests.
Don't forget to give the project a star! ⭐️