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. Using a Application Default Credentials,

  2. Using a service account by specifying a key file in JSON format. Key can be specified as a path to the key file (Keyfile Path), as a key payload (Keyfile JSON) or as secret in Secret Manager (Keyfile secret name). Only one way of defining the key can be used at a time. If you need to manage multiple keys then you should configure multiple connections.

  3. Using a credential configuration file, by specifying the path to or the content of a valid credential configuration file. A credential configuration file is a configuration file that typically contains non-sensitive metadata to instruct the google-auth library on how to retrieve external subject tokens and exchange them for service account access tokens.


    Additional permissions might be needed

    Connection which uses key from the Secret Manager requires that Application Default Credentials (ADC) have permission to access payloads of secrets.


    Alternative way of storing connections

    Besides storing only key in Secret Manager there is an option for storing entire connection. For more details take a look at Google Secret Manager Backend.

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.

Credential Configuration File

Credential configuration file JSON or path to a credential configuration file on the filesystem.

Not required if using application default credentials.

Secret name which holds Keyfile JSON

Name of the secret in Secret Manager which contains a service account key.

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.

Impersonation Chain

Optional service account to impersonate using short-term credentials, or chained list of accounts required to get the access_token of the last account in the list, which will be impersonated in all requests leveraging this connection. If set as a string, the account must grant the originating account the Service Account Token Creator IAM role. If set as a comma-separated list, the identities from the list must grant Service Account Token Creator IAM role to the directly preceding identity, with first account from the list granting this role to the originating account.

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:

    • project - Project Id

    • key_path - Keyfile Path

    • keyfile_dict - Keyfile JSON

    • key_secret_name - Secret name which holds Keyfile JSON

    • key_secret_project_id - Project Id which holds Keyfile JSON

    • scope - Scopes

    • num_retries - Number of Retries

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

For example, with URI format:

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

And using JSON format:

export AIRFLOW_CONN_GOOGLE_CLOUD_DEFAULT='{"conn_type": "google_cloud_platform", "extra": {"key_path": "/keys/key.json", "scope": "", "project": "airflow", "num_retries": 5}}'

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). The impersonation chain can also be configured directly on the Google Cloud Connection as described above, though the impersonation_chain passed to the operator takes precedence.

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.


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.

Domain-wide delegation

Some Google operators, hooks and sensors support domain-wide delegation, in addition to direct impersonation of a service account. Delegation allows a user or service account to grant another service account the ability to act on their behalf. This means that the user or service account that is delegating their permissions can continue to access and manage their own resources, while the delegated service account can also access and manage those resources.

For example:

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

    "properties": {"title": "Test1"},
    "sheets": [{"properties": {"title": "Sheet1"}}],

from import (

create_spreadsheet_operator = GoogleSheetsCreateSpreadsheetOperator(

Note that as domain-wide delegation is currently supported by most of the Google operators and hooks, its usage should be limited only to Google Workspace (gsuite) and marketing platform operators and hooks or by accessing these services through the GoogleDiscoveryAPI hook. It is deprecated in the following usages:

Was this entry helpful?