Architecture

Why omega|ml

A typical data science workflow consists of the following core steps:

  1. acquire data & store for subsequent processes
  2. clean data & publish for uses
  3. train & evaluate models
  4. publish models & reports
  5. execute prediction using previously trained models

In any production scenario, each step requires a scalable storage to store raw and cleaned data, models and APIs to execute models. You will also need a compute cluster that is easy to access and provides all the required packages. Engineering such a system from scratch is hard, takes considerable time and skills. omega|ml provides all of this in an integrated, scalable fashion.

omega|ml provides

  • the central storage for data and models, using MongoDB as the highly-scalable storage provider
  • a client API to out-of-core data processing that follows Pandas semantics
  • a client API to models that follows scikit-learn semantics, integrating scikit-learn and Apache Spark models
  • an integrated compute cluster runtime to train and execute models, as well as to execute arbitrary scripts and automatically publish reports
  • a sophisticated REST API to data, models, scripts and runtime
  • a user interface to access information on all of the above

Extensibility

With the exception of the REST API, all of the above are easily extensible using mixins.

In addition, omemgal provides interfaces to existing compute clusters like Anaconda’s Distributed and Apache Spark. omega|ml also provides an extensible framework to add custom backends and compute clusters through a common API.

Thanks to extensibility at the core of the architecture, omega|ml can easily accommodate any third-party storage or machine learning backend, or add new types of operations on data and models.

How omega|ml works

  • data is stored via the datasets.put API. datasets.put supports native Python objects like dicts and lists, Pandas DataFrames and Series, numpy arrays as well as externally stored files that are accessible through http, ftp or stored on cloud services like Amazon’s s3. Other datatypes can be easily added by a custom data backend.
  • machine learning models are stored via the models.put API. models.put supports scikit-learn and Spark mllib models. Other machine learning frameworks can be easily added by a custom model backend.
  • jobs (custom python scripts in the form of Jupyter notebooks) are stored via the jobs.put API.
  • the runtime cluster and any other authorized user can access the data models and jobs through the datasets.get, models.get and jobs.get methods, respectively. Using this common API any compute job e.g. to train a model can directly access the relevant data without the need to transfer the data to the worker instance first.

omega|ml is composed of the following main components:

Core components

The core components provide the storage for data and models. Models can be trained locally and stored in the cluster for prediction via the REST API.

  • Omega - the main API and programming interface to omega|ml
  • OmegaStore - the storage for data and models
  • OmegaRuntime - the celery runtime cluster to train and execute models and jobs

Enterprise Edition

The omega|ml Enterprise Edition provides a full integrated, enterprise-scale data science platform as a service. It is the best match for a multi-user environment with enterprise features and an extended set of functionality.

  • security features - security features covering all components (REST API, MongoDB, RabbitMQ etc.)
  • omegajobs - JupyterHub with per-user Notebooks
  • omegapkg - Lambda functions
  • omegaweb - the REST API, web interface and dashboard
  • omegaops - the operations API
  • omegacli - client components to the REST API
  • omegastream - Python native Streaming API
  • OmegaRuntimeDask - the dask distributed runtime cluster to execute models and jobs
  • SparkBackend - the Apache Spark mllib backend implementation
  • custom mixins - custom storage mixins to process data
  • custom backends - custom storage backends to third-party storage

Optional components

  • omegaetl - the distributed ETL engine to run ETL jobs on the cluster
  • workit - the distributed scheduler for large-scale job management

Third-party dependencies

omega|ml depends on the following third-party products (all open source):

  • MongoDB - the highly scalable NoSQL database, ideal for data science workloads
  • RabbitMQ - the most-widely used open source message broker
  • Celery - the efficient and highly-throughput Distributed Task Queue for Python applications
  • Anaconda - the most popular Python Data Science Distribution
  • MySQL - the world’s most popular open source database, backed by Oracle

Note that omegaml’s license does not include the above products. However, omega|ml provides the required docker build instructions to download, install and configure these applications for use with omegaml.

A number of smaller third-party components in the Python ecosystem are used in omegaml. Refer to the LICENSES file for details.

Positioning omega|ml

The core focus of omega|ml is to enable enterprise-grade application integration of data science workflows, at scale. While the following products provide specific elements for application integration and also integrate directly with omegaml, none of them provide the simplicity of the omega|ml API and the versatility of its standard and extensible backends.

  • Apache Spark - Spark is a JVM-based distributed compute cluster based on a in-memory data layer spread across the cluster. Spark is great if you rely on Scala or other JVM languages and have a Hadoop file system ready to store data in a native format. While there is a Python API, PySpark, it is relatively cumbersome and complex to use. Most importantly, Spark does not have support for scikit-learn models nor does it provide an enterprise- ready REST API. omega|ml provides the SparkBackend to integate with an existing Spark Cluster, providing the same easy API to Spark, shielding users from the complexities and pitfalls of the PySpark API.
  • Anaconda Dask Distributed - Dask Distributed is Anaconda’s answer to Spark in the Python ecosystem. Fully implemented in Python, Dask Distributed is a great additition to any Python-based data science workflow. However, very much like Apache Spark, Dask Distributed also does not provide an integration API nor does it provide persistent storage, neither for data nor for models. omega|ml provides a runtime implementation that integrates directly with a Dask Distributed cluster, making it easy and straight forward to use its power while keeping the advantage of omegaml’s API and flexibility.
  • mlflow - ML Flow provides tools for packaging and deploying models as a software component. While this is a good approach if there one or a few models in an organization, it does not solve the underlying meta data and management issues. omega|ml provides a platform, not just model packaging. In a future release omega|ml might make use of mlflow features for model packaging and serve as a deployment backend.