Google Cloud Connection

The Google Cloud connection type enables the Google Cloud Integrations.

Authenticating to Google Cloud

There are three ways to connect to Google Cloud using Airflow.

  1. Use Application Default Credentials,

  2. Use a service account key file (JSON format) on disk - Keyfile Path.

  3. Use a service account key file (JSON format) from connection configuration - Keyfile JSON.

Only one authorization method can be used at a time. If you need to manage multiple keys then you should configure multiple connections.

Default Connection IDs

All hooks and operators related to Google Cloud use google_cloud_default by default.

Note On Application Default Credentials

Application Default Credentials are inferred by the GCE metadata server when running Airflow on Google Compute Engine or the GKE metadata server when running on GKE which allows mapping Kubernetes Service Accounts to GCP service accounts Workload Identity. This can be useful when managing minimum permissions for multiple Airflow instances on a single GKE cluster which each have a different IAM footprint. Simply assign KSAs for your worker / webserver deployments and workload identity will map them to separate GCP Service Accounts (rather than sharing a cluster-level GCE service account). From a security perspective it has the benefit of not storing Google Service Account keys on disk nor in the Airflow database, making it impossible to leak the sensitive long lived credential key material.

From an Airflow perspective Application Default Credentials can be used for a connection by specifying an empty URI.

For example:

export AIRFLOW_CONN_GOOGLE_CLOUD_DEFAULT='google-cloud-platform://'

Configuring the Connection

Project Id (optional)

The Google Cloud project ID to connect to. It is used as default project id by operators using it and can usually be overridden at the operator level.

Keyfile Path

Path to a service account key file (JSON format) on disk.

Not required if using application default credentials.

Keyfile JSON

Contents of a service account key file (JSON format) on disk.

Not required if using application default credentials.

Scopes (comma separated)

A list of comma-separated Google Cloud scopes to authenticate with.

Number of Retries

Integer, number of times to retry with randomized exponential backoff. If all retries fail, the googleapiclient.errors.HttpError represents the last request. If zero (default), we attempt the request only once.

When specifying the connection in environment variable you should specify it using URI syntax, with the following requirements:

  • scheme part should be equals google-cloud-platform (Note: look for a hyphen character)

  • authority (username, password, host, port), path is ignored

  • query parameters contains information specific to this type of connection. The following keys are accepted:

    • extra__google_cloud_platform__project - Project Id

    • extra__google_cloud_platform__key_path - Keyfile Path

    • extra__google_cloud_platform__keyfile_dict - Keyfile JSON

    • extra__google_cloud_platform__scope - Scopes

    • extra__google_cloud_platform__num_retries - Number of Retries

Note that all components of the URI should be URL-encoded.

For example:

export AIRFLOW_CONN_GOOGLE_CLOUD_DEFAULT='google-cloud-platform://?extra__google_cloud_platform__key_path=%2Fkeys%2Fkey.json&'

Direct impersonation of a service account

Google operators support direct impersonation of a service account via impersonation_chain argument (google_impersonation_chain in case of operators that also communicate with services of other cloud providers).

For example:

import os

from import (


create_dataset = BigQueryCreateEmptyDatasetOperator(

In order for this example to work, the account impersonated_account must grant the Service Account Token Creator IAM role to the service account specified in the google_cloud_default Connection. This will allow to generate impersonated_account's access token, which will allow to act on its behalf using its permissions. impersonated_account does not even need to have a generated key.


GKEStartPodOperator, DataflowCreateJavaJobOperator and DataflowCreatePythonJobOperator do not support direct impersonation as of now.

In case of operators that connect to multiple Google services, all hooks use the same value of impersonation_chain (if applicable). You can also impersonate accounts from projects other than the project of the originating account. In that case, the project id of the impersonated account will be used as the default project id in operator's logic, unless you have explicitly specified the Project Id in Connection's configuration or in operator's arguments.

Impersonation can also be used in chain: if the service account specified in Connection has Service Account Token Creator role granted on account A, and account A has this role on account B, then we are able to impersonate account B.

For example, with the following terraform setup...

terraform {
  required_version = "> 0.11.14"
provider "google" {
variable "project_id" {
  type = "string"
resource "google_service_account" "sa_1" {
  account_id   = "impersonation-chain-1"
  project = "${var.project_id}"
resource "google_service_account" "sa_2" {
  account_id   = "impersonation-chain-2"
  project = "${var.project_id}"
resource "google_service_account" "sa_3" {
  account_id   = "impersonation-chain-3"
  project = "${var.project_id}"
resource "google_service_account" "sa_4" {
  account_id   = "impersonation-chain-4"
  project = "${var.project_id}"
resource "google_service_account_iam_member" "sa_4_member" {
  service_account_id = "${}"
  role               = "roles/iam.serviceAccountTokenCreator"
  member             = "serviceAccount:${}"
resource "google_service_account_iam_member" "sa_3_member" {
  service_account_id = "${}"
  role               = "roles/iam.serviceAccountTokenCreator"
  member             = "serviceAccount:${}"
resource "google_service_account_iam_member" "sa_2_member" {
  service_account_id = "${}"
  role               = "roles/iam.serviceAccountTokenCreator"
  member             = "serviceAccount:${}"

...we should configure Airflow Connection to use impersonation-chain-1 account's key and provide following value for impersonation_chain argument...

PROJECT_ID = os.environ.get("TF_VAR_project_id", "your_project_id")

...then requests will be executed using impersonation-chain-4 account's privileges.

Was this entry helpful?