Skip to content

install.md

Latest commit

 

History

History
87 lines (68 loc) · 2.84 KB

install.md

File metadata and controls

87 lines (68 loc) · 2.84 KB
Raw

Feast Installation Quickstart Guide

This is a quickstart guide for Feast administrators setting up a Feast deployment for the first time.

Feast is meant to run on Kubernetes, and currently requires managed services from GCP for certain operations

  • Dataflow (for loading feature data into Feast)
  • BigQuery (to act as feature warehouse)
  • Pub/Sub (for event data)
  • Cloud Storage (for staging data and logs)

This guide will assume that users are using GKE (Google Container Engine), and that these managed services are available. In addition, Redis will be used for the feature serving database.

Prerequisites

  • Kubernetes cluster
    • The user should have a GKE cluster provisioned, with kubectl set up to access this cluster
    • This cluster should have the right scopes to start jobs on Dataflow and to modify BigQuery datasets and tables. The simplest way to set this up is by setting the scope to cloud-platform when provisioning the Kubernetes cluster.
  • Helm
    • Helm should be installed locally and Tiller should be installed within this cluster. As noted here, make sure you have the cluster-admin role attached to the Helm service account.
  • Feast repository
    • You have cloned the Feast repository and your command line is active in the root of the repository

Set up the environment

Set the following environmental variables before beginning the installation

GCP_PROJECT=my-feast-project
FEAST_CLUSTER=feast
FEAST_REPO=$(pwd)
FEAST_VERSION=0.1.0
FEAST_STORAGE_BUCKET=gs://${GCP_PROJECT}-feast

Ensure that your kubectl context is set to the correct cluster

gcloud container clusters get-credentials "${FEAST_CLUSTER}" --project "${GCP_PROJECT}"

Create a storage bucket for Feast to stage data

gsutil mb "${FEAST_STORAGE_BUCKET}"

Install Feast using Helm

Create a copy of the Helm values.yaml file. This file will need to be configured with the specifics of your environment.

cp charts/feast/values.yaml .

Keys that likely need to be changed based on the user’s environment are

  • core.projectId - your GCP project
  • core.image.tag - the current Feast version
  • serving.image.tag - the current Feast version
  • core.jobs.*
    • DataflowRunner or DirectRunner or FlinkRunner settings for Beam
    • Note: DirectRunner should not be used in production it is for testing only.

See the documentation page for the complete list of options.

You can update values.yaml directly, or specify overrides on the Helm command line:

helm install --name feast charts/dist/feast-${FEAST_VERSION}.tgz \
--set core.projectId=${GCP_PROJECT} \
--set core.image.tag=${FEAST_VERSION} \
--set serving.image.tag=${FEAST_VERSION}