ForestFlow is a scalable policy-based cloud-native machine learning model server. ForestFlow strives to strike a balance between the flexibility it offers data scientists and the adoption of standards while reducing friction between Data Science, Engineering and Operations teams.

ForestFlow is policy-based because we believe automation for Machine Learning/Deep Learning operations is critical to scaling human resources. ForestFlow lends itself well to workflows based on automatic retraining, version control, A/B testing, Canary Model deployments, Shadow testing, automatic time or performance-based model deprecation and time or performance-based model routing in real-time.

Our aim with ForestFlow is to provide data scientists a simple means to deploy models to a production system with minimal friction accelerating the development to production value proposition.

To achieve these goals, ForestFlow looks to address the proliferation of model serving formats and standards for inference API specifications by adopting, what we believe, are currently, or are becoming widely adopted open source frameworks, formats, and API specifications. We do this in a pluggable format such that we can continue to evolve ForestFlow as the industry and space matures and we see a need for additional support.

Why ForestFlow?

Continuous deployment and lifecycle management of Machine Learning/Deep Learning models is currently widely accepted as a primary bottleneck for gaining value out of ML projects.

We first set out to find a solution to deploy our own models. The model server implementations we found were either proprietary, closed-source solutions or had too many limitations in what we wanted to achieve.
The main concerns for creating ForestFlow can be summarized as:

  • We wanted to reduce friction between our data science, engineering and operations teams
  • We wanted to give data scientists the flexibility to use the tools they wanted (H2O, TensorFlow, Spark export to PFA etc..)
  • We wanted to automate certain lifecycle management aspects of model deployments like automatic performance or time-based routing and retirement of stale models
  • We wanted a model server that allows easy A/B testing, Shadow (listen only) deployments and and Canary deployments. This allows our Data Scientists to experiment with real production data without impacting production and using the same tooling they would when deployment to production.
  • We wanted something that was easy to deploy and scale for different deployment scenarios (on-prem local data center single instance, cluster of instances, Kubernetes managed, Cloud native etc..)
  • We wanted the ability to treat inference requests as a stream and log predictions as a stream. This allows us to test new models against a stream of older infer requests.
  • We wanted to avoid the "super-hero" data scientist that knows how to dockerize an application, apply the science, build an API and deploy to production. This does not scale well and is difficult to support and maintain.
  • Most of all, we wanted repeatability. We didn't want to re-invent the wheel once we had support for a specific framework.

Model Deployment

For model deployment, ForestFlow supports models described via MLfLow Model format which allows for different flavors i..e, frameworks & storage formats.

ForestFlow also supports a BASIC REST API for model deployment as well that mimics the MLflow Model format but does not require it.


For inference, we’ve adopted a similar approach. ForestFlow provides 2 interfaces for maximum flexibility;
a BASIC REST API in addition to
standardizing on the GraphPipe
API specification.

Relying on standards, for example using GraphPipe’s specification means immediate availability of client libraries in a variety of languages that already support working with ForestFlow; see GraphPipe clients.

Please visit the quickstart guide to get a quick overview of setting up ForestFlow and an example on inference.
Also please visit the Inference documentation for a deeper dive.