**The recent addition of the *Firestore* Lookup variable in GTM Server opens up a whole new world of possibilities in terms of data enrichment. But how to get data into the *Firestore* database? This tutorial will show you how to load data from *BigQuery* into *Firestore* (including code samples).**

Contents

- Set up a Google Workflow that exports BigQuery data to a Storage Bucket as JSON.
- Set up a Google Cloud Function that uploads the JSON file to Google Firestore.
- Connect Firestore to Google Tag Manager Server and use the Firestore Lookup Variable

## Google Tag Manager Server & Firestore

The release of the Firestore Lookup and [asynchronous variables](https://www.simoahava.com/analytics/asynchronous-variables-server-side-google-tag-manager/) within GTM Server variables opens up a whole new world of possibilities.

By using the Firestore Lookup variable (or the asynchronous variable to lookup data from some other system) one could think of the following use cases:

- Enrich product data / transactions (e.g. adding margins or product attributes)
- Enrich user data (e.g. adding user properties, segments, CLV scores)
- User-ID enrichment / identity stitching (retrieve alternative identifiers)

The `Firestore API` that powers the *Firestore Lookup variable*, can also be leveraged to write data to Firestore. Use cases:

- A persistent storage system (e.g. storing user-level data over different events)
- Enrich user data (e.g. adding user properties, segments, CLV scores)
- *And much more*

Many clients I work for, store the necesarry data for the use cases mentioned above in `Google BigQuery` (think of single user profiles or product feeds). However, BigQuery is not suitable to connect directly to GTM Server.

> A while ago I wrote an [comparable article](https://stacktonic.com/article/create-a-user-profile-and-recommender-service-using-big-query-redis-and-gtm-server) explaining how to connect a Redis cache to GTM Server. However the Firestore method seems the way to go now (although the Redis / GTM Server connection can be optimized using asyncrhoniys variables)

## BigQuery vs. Firestore

BigQuery is meant as an analytics database, not as a database that can perform a lot of reads and writes in low latency environments. But what databases do? 

The following graph is a nice representation of the different databases and use cases. (of course biased to the options available in the Google Cloud)

![which-database-google-cloud.jpeg](https://static.stacktonic.com/medium_which_database_google_cloud_e5031028bd.jpeg)
*[Google Cloud database options](https://cloud.google.com/blog/topics/developers-practitioners/your-google-cloud-database-options-explained). Image by [thecloudgirl.dev](https://thecloudgirl.dev)*

There a multiple contenders for our use case (think of Redis / Memcache used in a [related article](https://stacktonic.com/article/create-a-user-profile-and-recommender-service-using-big-query-redis-and-gtm-server)), but in this case we'll opt for Firestore.

## The solution - Google Workflows!

The solution we've build and are sharing in this article leverages a `Google Worfklow` and a `Google Cloud Function`. 

Google Workflow enables you to set up an automation with just a "simple" `.yaml` file (also see code below). The Google Workflow will:

1. Export the specified query to a new BigQuery table 
2. Export the table to a Storage Bucket in JSON format
3. Creates a message in a Google Pub/Sub topic (holding the configuration for the Cloud Function to run like the bucket name and project-ID)

Lastly, a Google Cloud Function will trigger, opens the JSON export in the Storage Bucket and loops the rows into the Firestore database.

[![export-google-bigquery-to-firestore-plain.PNG](https://static.stacktonic.com/medium_export_google_bigquery_to_firestore_plain_e1c86f907a.PNG)](https://static.stacktonic.com/export_google_bigquery_to_firestore_plain_e1c86f907a.PNG)

> Read more about [Google Workflows](https://stacktonic.com/article/google-cloud-workflows-export-data-from-big-query-to-cloud-storage) in my other tutorial.

## Prerequisites

So let's start! 

Before proceeding, make sure you have:

- An active Google Cloud Project
- Some data within Google BigQuery, ready to send Firestore
- A query that fetches the data you want to send to Firestore (and think of a unique key you want to match with event data in GTM Server)
- A Google Cloud Storage bucket (to store the BigQuery exports)

And don't forget to enable the following services in your GCP project (and have some basic knowledge about the services):

- Enable Google Workflows
- Enable Google Cloud Functions
- Enable Google Pub/Sub

## Step 1: Create the Google Cloud Function

First step, create a Cloud Function that triggers on a Pub/Sub topic. Also create the Pub/Sub topic (can be done directly when creating the Cloud Function). 

- **Environment:** 1st gen
- **Trigger type:** Cloud Pub/Sub
- **Memory Allocated:** 1GB (depends on the amount of data you want to process, so adjust accordingly).
- **Timeout:** Depends on the file size you want to process (see limits). For files that are not that large, you could start with the default 60 seconds.
- **Runtime:** Node.js 16
- Attach a Service Account that has access to the Google Cloud Bucket + Firestore + Pub/Sub

Configure the following roles for the Service Account that is running the the Cloud Function (and is also added to the project where Firestore is enabled)

- Storage Admin
- Cloud Datastore User
- Pub/Sub Publisher

Copy paste the JS code below ([also available in the GitHub repo](https://github.com/stacktonic-com/gcp-export-bigquery-to-firestore/tree/master/cloudfunction-cloudstorage-to-firestore))

> You don't have to adjust or configure anything in the scripts itself. The Google Workflow will pass the configuration settings (like the bucket / path of the JSON export) in the Pub/Sub message.

**package.json**

```json
{
    "name": "gcf-cloudstorage-to-firestore",
    "version": "1.0.0",
    "dependencies": {
        "@google-cloud/storage": "^5.20.3",
        "firebase-admin": "^10.2.0",
        "split": "^1.0.1"
    }
}
```

**index.js**

```javascript
'use strict';

const admin = require('firebase-admin');
const {Storage} = require('@google-cloud/storage');
const split = require('split');
const { pipeline } = require('stream/promises');

// Upload functioin
async function jsontoFirestore(file, firestoreKey, firestoreCollection) {  
    
    let keysWritten = 0;
    
    return new Promise(resolve => {
        file.createReadStream()
        .on('error', error => reject(error))
        .on('response', (response) => {
            // connection to GCS opened
        }).pipe(split())
        .on('data',  async record => {
            if (!record || record === "") return;
                keysWritten++;

                const data = JSON.parse(record);
                const key = data[firestoreKey].replace(/[/]|\./g, '');

                try {
                    await admin.firestore().collection(firestoreCollection).doc(key).set(data)
                } catch(e) {
                    console.log(`Error setting document:  ${e}`);
                }
        })
        .on('end', () => {
            console.log(`Successfully written ${keysWritten} keys to Firestore.`);
        })
        .on('error', error => reject(error));
    });
}    

/**
 * Triggered from a Pub/Sub message.
 *
 * @param {!Object} event Event payload.
 * @param {!Object} context Metadata for the event.
 */
exports.loadCloudStorageToFirestore = async(event, context) => {

    const pubSubMessage = event.data ? Buffer.from(event.data, 'base64').toString(): '{}';
    const config = JSON.parse(pubSubMessage);

    console.log(config)

    if (typeof config.projectId != 'undefined') {

        const projectId = config.projectId;
        const bucketName = config.bucketName;
        const bucketPath = config.bucketPath;
        const firestoreCollection = config.firestoreCollection;
        const firestoreKey = config.firestoreKey;

        console.log(`Initiated new import to Firebase: gs://${bucketName}/${bucketPath}`)

        // Init Firebase
        if (admin.apps.length === 0) {
            admin.initializeApp({ projectId: projectId })
        }

        // Init Storage
        const storage = new Storage()
        const bucket = storage.bucket(bucketName);
        const file = bucket.file(bucketPath);

        try {
        
            // TO-DO: Remove old records

            // Read file and send to Firestore
            await jsontoFirestore(file, firestoreKey, firestoreCollection);

        
        } catch(e) {
            console.log(`Error importing ${bucketPath} to Firestore: ${e}`);
        }
    }

};
```
*Some parts taken from this [Cloud Function](https://futurice.com/blog/bigquery-to-memorystore).*

## Step 2: Create the Google Workflow

Create a new Google Workflow and make sure the Service Account you assigned to the Workflow has access to the following services:

- Google BigQuery (*BigQuery Data Editor + BigQuery Job User*)
- Google Storage (*Storage Admin*)
- Google Cloud Scheduler (to regularly execute the Workflow)

Copy / paste the code from below (or from this [GitHub repository](https://github.com/stacktonic-com/gcp-export-bigquery-to-firestore/tree/master/workflow-bigquery-to-cloudstorage)

> This time, **you do need to add your configuration values** (like `project_id` or `gcs_bucket`).

The GCP console will render a visual representation of your Workflow and should look something like this:

![google-workflow-bigquery-to-cloud-storage-json.PNG](https://static.stacktonic.com/google_workflow_bigquery_to_cloud_storage_json_07da6e2aba.PNG)

**workflow-bigquery-to-cloudstorage.yaml**

```yaml
- init:
    assign:
    - project_id: "<your-project-id>"
    - bq_dataset_export: "<your-bq-dataset-for-export-table>"
    - bq_table_export: "<your-bq-tablename-for-export-table>"
    - bq_query: >
            select
                user_id,
                device_first,
                channel_grouping_first
            from
                `stacktonic-cloud.st_core.dim_customer`
    - gcs_bucket: "<your-export-bucket>"
    - gcs_filepath: "firestore-export/firestore-export.json"
    - pubsub_topic: "<your-pubsub-topic-name>"
    - pubsub_message: {
            "projectId": "<your-firestore-project-id>",
            "bucketName": "<your-export-bucket>",
            "bucketPath": "firestore-export/firestore-export.json",
            "firestoreCollection": "<your-firestore-collection>",
            "firestoreKey": "<your-key-to-use-as-firestore-document-id>"
        }
- bigquery-create-export-table:
                call: googleapis.bigquery.v2.jobs.insert
                args:
                    projectId: ${project_id}
                    body:
                        configuration:
                            query:
                                query: ${bq_query}
                                destinationTable:
                                    projectId: ${project_id}
                                    datasetId: ${bq_dataset_export}
                                    tableId: ${bq_table_export}
                                create_disposition: "CREATE_IF_NEEDED"
                                write_disposition: "WRITE_TRUNCATE"
                                allowLargeResults: true
                                useLegacySql: false

- bigquery-table-to-gcs:
    call: googleapis.bigquery.v2.jobs.insert
    args:
        projectId: ${project_id}
        body:
            configuration:
                extract:
                    compression: NONE
                    destinationFormat: "NEWLINE_DELIMITED_JSON"
                    destinationUris: ['${"gs://" + gcs_bucket + "/" + gcs_filepath}']
                    sourceTable:
                        projectId: ${project_id}
                        datasetId: ${bq_dataset_export}
                        tableId: ${bq_table_export}
- publish_message_to_pubsub:
    call: googleapis.pubsub.v1.projects.topics.publish
    args:
      topic: ${"projects/" + project_id + "/topics/" + pubsub_topic}
      body:
        messages:
          - data: ${base64.encode(json.encode(pubsub_message))}
```

**Now run the Google Workflow!** 

The workflow should export the BigQuery query results (yes!) to Cloudstorage and load it into Firestore:

![google-firestore-example.PNG](https://static.stacktonic.com/medium_google_firestore_example_43d3520240.PNG)

## Step 3: Connect GTM Server with Firestore

The last step, make sure the GTM environment can connect with the Firebase database.  Simo Ahava already wrote a great article about [Firestore and GTM Server](https://www.simoahava.com/analytics/enrich-server-side-data-with-cloud-firestore/), so check out that article. 

Some remarks:

- **Make sure you set up Firestore in Native mode (can't be undone)**
- When you run GTM Server in App Engine, make sure the Service Account that runs App Engine (probably App Engine default service account) has access to Firestore (*Datastore User*).
- If not in the same project, add the service account that runs App Engine to the project where Firestore lives.

## Limits & cost

Some last considerations:

**Limits**

This setup is using full loads:

- Only new records are added or existing records are updated (so old records are not deleted).
- This process can be optimized by using delta’s (for example set up a process in BigQuery calculating the daily delta’s) or deleting all the records in the Firestore collection first.

Possible issues with large datasets:
- Google Cloud Functions timeout limit (configurable, but a maximum runtime of 9 minutes)
- Google Cloud Function memory limit (although files from Cloud Storage are not fully loaded into memory but processed streaming)

**Costs**

Google Workflows and Cloud Functions are serverless. Likely these will not incur a lot of costs since the Workfow and Function probably won't be executed a lot. Firestore heavily depends on volume + there is also a free tier. Check the [official pricing overview](https://cloud.google.com/firestore/pricing) for the full overview.

The recent addition of the Firestore Lookup variable in GTM Server opens up a whole new world of possibilities in terms of data enrichment. But how to get data into the Firestore database? This tutorial will show you how to load data from BigQuery into Firestore (including code samples).

large_google-bigquery-to-firestore.PNG

medium_google-bigquery-to-firestore.PNG

small_google-bigquery-to-firestore.PNG

thumbnail_google-bigquery-to-firestore.PNG

google-bigquery-to-firestore.PNG

Export BigQuery data to Google Firestore & GTM Server

**For a lot of data projects Apache Airflow is the orchestration tool of choice. However, Airflow can be a bit overkill for smaller projects and tasks. One alternative is Google Workflows where you can create orchestration flows using simple *.yaml* files. In this tutorial I will show you how to export data from BigQuery to a Google Cloud Storage bucket using Google Workflows.**

Contents:

- Introduction Google Workflows
- Workflow example: Exporting Google BigQuery data to Google Cloud Storage as a CSV (two methods).

## Why Google Workflows?

Google Workflows is a lightweight and serverless workflow orchestration service where you can connect Google Cloud and external services (using APIs). As mentioned, Apache Airflow is used in a lot of projects, but  it can be a bit overkill since you have to maintain the Airflow cluster (or pay to maintain it). It can also have some overlap with other tools like *dbt*. 

 Triggered by a post from Robert Sahlin related to [modern data workflow orchestration](https://robertsahlin.com/modern-data-workflow-orchestration-part-1/) and [The Unbundling](https://blog.fal.ai/the-unbundling-of-airflow-2/) of Airflow by Gorkem Yurtseven I've dicided to give [Google Workflows](https://cloud.google.com/workflows) a try for some workflows mainy related to Google Cloud. Currently, I'm using Workflows successfully in some smaller projects.

## Google Workflows basics

Basically, Google workflows uses a *.yaml* file where you can build your workflow. The file contains tasks that can be executed in order. You can upload this *.yaml* file and Google will create a visual representation of your flow. Flows can be scheduled using Cloud Scheduler, executed manually or using an API. Logs are also collected.

 There are a bunch of connectors / APIs available already and the list is growing. For an overview [checkout the official connector list](https://cloud.google.com/workflows/docs/reference/googleapis). Make sure you also check out the [introduction by Google](https://cloud.google.com/workflows/docs/overview).

> As of writing, Google Workflows is fairly new, so example code and documentation is somewhat limited.

Exporting from Google BigQuery to a Cloud Storage bucket is an operation that is present in a lot of pipelines I developed. So I'm taking this operation to build an example Workflow. Two methods:

## Method 1: BigQuery to Cloud Storage (intermediate table)

The first method will create an intermediate table from your query, and is exporting that table as a CSV to Google Cloud Storage. Adding this step enables you to export large amounts of data into a single file (max 1GB to a single file). If you have a limited amount of data or it doesn't matter that data is always exported to multiple CSV files / file parts, you could also use option 2.

**Workflow steps**

1. Run a query a save the result in an intermediate table
2. Export the intermediate table to a CSV

![google-workflows-bigquery-to-google-cloud-storage-option-1.png](https://static.stacktonic.com/google_workflows_bigquery_to_google_cloud_storage_option_1_78241bbfc8.png)

**Workflow .yaml code**

```yaml
- init:
    assign:
    - project_id: "<your-project-id>"
    - dataset_export: "<your-bigquery-dataset>"
    - table_export: "<your-bigquery-table>"
    - query: >
        select * from <your-table>
    - bucket: "<your-cloud-bucket>"
    - folder: "<your-export-folder-within-cloud-bucket>"
    - filename: ${string(sys.now()) + "-<your-filename>.csv"}

- bigquery-create-export-table:
                call: googleapis.bigquery.v2.jobs.insert
                args:
                    projectId: ${project_id}
                    body:
                        configuration:
                            query:
                                query: ${query}
                                destinationTable:
                                    projectId: ${project_id}
                                    datasetId: ${dataset_export}
                                    tableId: ${table_export}
                                create_disposition: "CREATE_IF_NEEDED"
                                write_disposition: "WRITE_TRUNCATE"
                                allowLargeResults: true
                                useLegacySql: false

- bigquery-table-to-gcs:
    call: googleapis.bigquery.v2.jobs.insert
    args:
        projectId: ${project_id}
        body:
            configuration:
                extract:
                    compression: NONE
                    destinationFormat: "CSV"
                    destinationUris: ['${"gs://" + bucket + "/" + folder + "/" + filename}']
                    fieldDelimiter: ","
                    printHeader: true
                    sourceTable:
                        projectId: ${project_id}
                        datasetId: ${dataset_export}
                        tableId: ${table_export}

```

## Method 2: BigQuery to Cloud Storage (direct)

This method will export a query result directly to a CSV. However, it will create multiple file parts in most situations. 

**Workflow .yaml code**

```yaml
- init:
    assign:
    - project_id: "<your-project-id>"
     - query: >
        select * from <your-table>
    -  - bucket: "<your-cloud-bucket>"
    - folder: "<your-export-folder-within-cloud-bucket>"
    - filename: "<your-filename)"
- bigquery-table-to-gcs:
    call: googleapis.bigquery.v2.jobs.query
    args:
      projectId: ${project_id}
      body:
        query: ${"EXPORT DATA OPTIONS( uri='gs://" + bucket + "/" + folder + "/'||current_date()||'-" + filename + "-part-*.csv', format='CSV', overwrite=true, header=true) AS " + query}
        useLegacySql: false
```

## Closing thoughts

Google Workflows is an excellent service for small projects and tasks. But it can also be used in larger projects. You can create Workflows using the interface, the CLI / API or [Terraform](https://cloud.google.com/workflows/docs/create-workflow-terraform), which makes it more manageable in larger projects. 

There are also some remarks. Think of pipelines that need to be able to perform backfills. By default, this is harder to achieve within Workflows. Or more complex flows that can't be easily writtin in a *.yaml* configuration file. In those cases, Airflow of Prefect are also good alternatives.



For a lot of data projects Apache Airflow is the orchestration tool of choice. However, Airflow can be a bit overkill for smaller projects and tasks. One alternative is Google Workflows where you can create orchestration flows using simple *.yaml* files. In this tutorial I will show you how to export data from BigQuery to a Google Cloud Storage bucket using Google Workflows.

large_workflows-bigquery-to-cloud-storage-leader.png

medium_workflows-bigquery-to-cloud-storage-leader.png

small_workflows-bigquery-to-cloud-storage-leader.png

thumbnail_workflows-bigquery-to-cloud-storage-leader.png

workflows-bigquery-to-cloud-storage-leader.png

Google Cloud Workflows: Export data from BigQuery to Cloud Storage

**There is some debate about the effectiveness and future potential of Customer Data Platforms (CDPs). One aspect is the increasing competition from cloud data (warehouse) platforms. Overlap seems to be growing as it became much easier to build (some) of the functionality that existing CDPs are offering. Referencing this approach, a term I hear more often is the *Headless* (Customer) Data Platform. What is this *headless* approach all about and how does it compare to existing, stand-alone tools and suites?**

The Headless Customer Data Platform (CDP) or Data Management Platform (DMP) refers to an increasingly popular solution (as there is no exact definition and probably never will be): 

- A cloud data warehouse acts as the foundation (think of GCP / AWS / Azure). Different data sources are ingested, transformed and joined. 
- Create insights and actions on all kinds of levels (**not limited to customer data**). Think of product or campaign level data.
- Activate data in channels, platforms and channels **directly from the data warehouse environment**
- Without a stand-alone CDP tool or suite, but a combination of cloud (data) services and specialized tools or frameworks.

## Headless?

*Headless* architecture gained popularity in the Content Management (CMS) world and is widely used these days. The concept: decoupling the back-end (content management) and the front-end (website). The CMS will act as a central content hub. Through an API, content can be shared with all sorts of applications, such as websites, apps or even dynamically generated emails . The front-end is no longer intertwined with the content management system. The result: all content centralized in one system and front-end applications can be developed independently from the CMS.

The headless CDP or DMP acts as a **(marketing) data hub** where insights and actions can be created on all kinds of levels **and** are activated / integrated with channels, tools or platforms (think of audiences, events or triggers). This is done directly from the cloud data warehouse, without the need for a stand-alone CDP tool or suite. However, there is no centralized user interface. The solution consists of a combination of services within the cloud environment (building blocks) and other (specialized) tools that fulfill a specific need (e.g. connectivity or decision making). 

![headless-cdp-architecture_white.PNG](https://static.stacktonic.com/medium_headless_cdp_architecture_white_cb9e8e79bc.PNG)


## Not limited to customer data only

I'm comparing the headless approach to existing stand-alone CDP solutions, which are mainly build around customer data. However, a headless data management system is not limited to customer data only. Some examples:

- Identity management, building a single (360°) customer view and (customer) audiences.
- Enriching product data / product feeds. Gaining insights into product performance.
- Create models, such as product recommenders, purchase of churn prediction models or forecasting sales (both in batch as real-time).
- Creating business rules or “triggers” (decisioning) for follow-up / next-best actions.
- Sales and engagement reports, based on multiple data sources. Creating notifications and alerts based on deviations within these reports.

The headless approach has a lot of overlap with the idea behind [The Modern Datastack](https://snowplowanalytics.com/blog/2021/05/12/modern-data-stack/). However, I will try to make the comparison with existing CDP tools, so limiting myself to functionalities these platforms generally offer.

## A data warehouse 'with benefits'

Many organizations recently invested in a data warehouse in one of the well-known cloud providers such as AWS, Azure or GCP. Data can be retrieved from various sources and joined together. Usually this is based on periodic batches (once per day or hour), but real-time data ingestion is also perfectly possible (due to the rise of real-time architecture). Think of (live) orders or clickstream data from websites / apps.

Working with such a data warehouse and all the different services within these platforms became much easier. Services within the cloud platforms can be stacked on top of each other (**building blocks**), making it easier to import, transform and analyze large amounts of data. Skills that are important for working with such a data environment, such as SQL, R or Python, are also increasingly present within organizations (but yes, in most cases not nearly enough).

*Activating* this data in channels and/or tools directly from the data warehouse is what increases the overlap with existing CDPs, thus becoming an alternative. 

## The headless approach vs. CDP tools: overlap in functionality

There are many CDP 'flavors'. For example, one platform is focusing more on activation, personalization and AI. The other on creating insights, identity resolution and a 360° customer view. However, almost all of the systems require you to to collect and ingest source data. Often (Javascript) trackers have to be implemented on-site and (mostly) CSVs uploaded to join online data with order or CRM data. 

Lately, when talking about a CDP when a cloud data warehouse environment is already present (or will be), generally the following comes up:

1. The data sources that already flow into your data warehouse, also have to be integrated with the CDP.
2. Capturing and collecting real-time (on-site) behavioral data most of the times is already done by systems such as Google Analytics or Snowplow (and more recently leveraging *server-side* tag management solutions). So why is it necessary to create another online behavioral database (in which measurement methodology certainly will differ)?
3. Duplicating business rules already present in your warehouse environment. Think of identity resolution, creation of a 360 customer view or audience selections.

If the above applies to your organization depends on your specific situation, but in most cases there will be overlap. And most CDP and data management tools do not (yet) work together in such a way that they can "piggyback" on everything that is already collected and calculated in your data warehouse.

## Disadvantages of a headless CDP

Of course, stand-alone CDPs have their advantages. Mainly this is time and cost. First off all if you don't have a data warehouse and/or resources are limited, a stand-alone CDP suite could fit your needs and get you started quickly, seen the out-of-the box functionality. The cost factor heavily depends type of system, volume and integrations. When zooming in, the headless CDP has the following disadvantages: 

1. No central user interface & ease of use: a headless system does not have an integrated user interface where settings can be easily adjusted or insight can be given into all processes. It is certainly possible to gain insight into the processes of a headless system, but those insights have to be created yourself or is available in multiple systems. So there is some technical / SQL knowledge required to set up audiences or journeys for example.
2. (Real-time) decision engine / flow builder: making decisions based on real-time data flows (e.g. web or app data). Creating such a real-time decision engine (e.g. online personalization use cases) within the cloud is possible, but is much more complex than clicking some rules together in a user interface. Also, creating flows / journeys is easily done within most CDP solutions.
3. Connectivity to channels and tools is maintained by the platform. In a headless CDP, you have to build the integrations yourself (mostly trough API connections), although there are currently several (open-source) frameworks available, such as [Google Tentacles](https://stacktonic.com/article/build-your-own-cdp-activating-audiences-in-google-ads-customer-match-using-google-tentacles). Interesting development is the rise of "reverse ETL" tools.

## Reverse ETL

This last one deserves some some explanation. There are already a lot of tools that ingest data *into* your warehouse, but less tools that can activate that data to your marketing channels or other platforms. *Reverse ETL* tries to solve this problem. Ultimately, a tool like this should enable you to focus on building use cases and not to worry about connector development and maintenance (e.g. the Facebook API). Next to that, the tools is providing insights in the process and data that flows out of your cloud platform. It also seems that some of these tools are developing a user interface where non-technical users can create audiences or flow on top of your data warehouse data.

*An example setup*

![headless-cdp-architecture_white_example_setup.PNG](https://static.stacktonic.com/medium_headless_cdp_architecture_white_example_setup_0c4b10c778.PNG)

## When to choose for a headless approach?

When your organization already has a cloud data warehouse or planning to set one up, the headless approach can be interesting if you:

1. Don't want to purchase and implement an (expensive) CDP tool, for example if you just want to share audiences, triggers or other data directly from the data warehouse with your (marketing) channels or (external) platforms;
2. Want to realize advanced and more complex use cases based on different sources and combinations, which are not possible or hard to achieve in stand-alone tools (by clicking stuff together) or without moving the data out and in the CDP platform to perform the necessary actions.
3. Do not want to duplicate data and recreate business rules in different environments (single source of thruth).
4. Unlock the potential and activate non-customer data, such as product or campaign data.

Using various cloud services and dedicated / specialized tools (such as reverse ETL), it's possible to connect data and tools relatively easily. Keep in mind that certain skills must be present in the team that will work with the headless CDP on a daily basis and at scale. Think of data engineers, tech-savvy marketers and analysts. This is especially important as use cases become more complex. 

## Start with your use cases

But most importantly, when choosing between a headless or a stand-alone CDP (or whatever solution), start with **defining clear use cases** and determine the roles / persons that will work with such a system. *“The ability to create audiences in a drag-and-drop interface and share these with our marketing channels”* is not a solid use case. Creating 100 audiences, in which the same customer is in 80 of these 100 audiences, and is targeted 80 times with a different message, does not work. Try to work out the details.

Insight and ease of use are certainly advantages of a stand-alone CDP, although the promise that every marketer can work with a CDP is often not fulfilled. Whether that depends on the marketer or the tool will differ per situation and person. The fact is that in reality, the more technical marketers and/or data engineers become the main users of such systems when the "low-hanging-fruit" use cases are live and CDP needs to be build out.

So, the best tip I can give is to involve the right people, work out clear requirements, detailed use cases and then pick a solution.

There is some debate about the effectiveness and future potential of Customer Data Platforms (CDPs). One aspect is the increasing competition from cloud data (warehouse) platforms. Overlap seems to be growing as it became much easier to build (some) of the functionality that existing CDPs are offering. Referencing this approach, a term I hear more often is the *Headless* (Customer) Data Platform. What is this *headless* approach all about and how does it compare to existing, stand-alone tools and suites?

large_headless-cdp-architecture.PNG

medium_headless-cdp-architecture.PNG

small_headless-cdp-architecture.PNG

thumbnail_headless-cdp-architecture.PNG

headless-cdp-architecture.PNG

The Future of CDPs – The Headless approach

**Use Google Analytics 4 data to enrich a 360° Customer View with online behavioral data, such as visit frequency, most used device, channel and/or engagement metrics. This tutorial provides an SQL example you can use directly.**

- Enrich a **single customer view** with online behavioral data collected by Google Analytics 4
- Use the GA4 BigQuery exports to aggregate Google Analytics 4 data to **user level**
- Think of most frequently used marketing channel, device, visit frequency & recency and most viewed content.
- Go [directly to the SQL code](#query-the-customer-view)

## What is a Single Customer View?

A (single) customer view is also called a Unified Customer View, a 360° Customer View or Profile. Basically it aggregates data on user level from different sources to a single profile / representation of a customer (or user). 

Most Customer Data Platforms (CDPs) offer this functionality, but you can also build a single customer view within your data warehouse. Common sources for these customer views are:

- CRM databases
- Customer service systems
- Order databases
- Offline behavior (e.g. in-store purchases using a loyalty card)
- Online behavior

In this article, we will add the online behavioral data the the customer view. Example:

![google-analytics-4-single-customer-view-table.PNG](https://static.stacktonic.com/google_analytics_4_single_customer_view_table_fc79c6859e.PNG)

By creating such a customer view, data can be used to (but certainly not limited to):

- Get a better and more complete understanding of your customer.
- Input for (customer) segmentation (e.g. clustering).
- Input for modeling / predictions (e.g. conversion or churn prediction)
- Target ads (think of email or social/display campaigns, but also on-site personalization)
- Channel preference / optimization
- Content recommendations
- Providing additional information for customer service (e.g. by connecting the view to customer service / CRM software)

## Use Google Analytics 4 data to extract online behavior

As mentioned, Google Analytics 4 (GA4) can be used to add online behavior to the customer view. Leveraging the GA4 BigQuery exports, we can aggregate data to user level, thus providing all sorts of interesting user features.

> Caution! Make sure you have the appropriate user consent before collecting and using the data.

![single-customer-view-google-analytics-4.PNG](https://static.stacktonic.com/small_single_customer_view_google_analytics_4_15b0e1e90b.PNG)

## Customer identification & mapping table

To match GA4 data with other data sources, you will need some sort of a common identifier, like a customer-ID. 

> You could also build user profiles of **anonymous users**. So not for every use case a customer-ID is required. In this article, we will assume that you want to connect the GA4 data to other systems (but provided comments in the SQL code below to build profiles for anonymous users). 

So if you're planning to connect the data to other systems, you will need to configure the [collection and matching of customer-IDs / user-IDs within Google Analytics 4](https://stacktonic.com/article/create-a-user-mapping-table-based-on-the-google-analytics-4-big-query-dataset). This article will show you how to create a user mapping table (private Identity Graph), which you can use to:

1. Match device-IDs (GA4 cookie-ID: `user_pseudo_id`) to customer-IDs or user-IDs
2. Use this match / mapping table to easily add a customer-ID to sessions and unify sessions before aggregating the data to users, thus creating cross-device profiles, depending on the identification methods (e.g. login).

![identity-graph-resolution-bigquery-sql.png](https://static.stacktonic.com/small_identity_graph_resolution_bigquery_sql_09741780f7.png)

*Private Identity Graph*

## Extracting data from the BigQuery GA4 dataset

We assume you've already set up the GA4 exports to Google BigQuery. 

> If you've decided to group on customer-ID, we assume you've set up the mapping table described in the previous chapter. 

Now we're going to transform al the GA4 event level data to user level data. The different steps explained:

1. Aggregate the event (hit) level data to session level. You have to set the **lookback_window** variable for which the user views will be constructed. We've set this to 12 months, but you should adjust this to your specific situation.
2. Match sessions on device-ID (`user_pseudo_id`) to customer-IDs. We will use the mapping table that you can find in [this article](https://stacktonic.com/article/create-a-user-mapping-table-based-on-the-google-analytics-4-big-query-dataset), as described in the previous chapter. Based on this table, we will create an intermediate match table, where the ga_client_id (= `user_pseudo_id`) is unique. We only select device-IDs that have a maximum of 3 (or less) customer-IDs attached. If multiple customer-IDs, the most recent one is selected. You can skip this step if you want to create a view for anonymous users.
3. Aggregate the session level data to user level, and creating the "customer view". 

> The query will provide you with a bunch of user features (described in the table below). You're certainly not limited to these features, as you can come up with a lot more, specific for your business. The query will show you some concepts how to get different kinds of data and aggegrations.

*All features are calculated within the configured lookback period (except `first_touch` and `traffic_source_first`, because these are persistent parameters within the GA4 dataset)*

| Feature | Description |
| --- | --- |
| first_touch | Timestamp of the very first session (can be outside the lookback window) |
| first_session | Timestamp of first session within lookback period |
| last_session | Timestamp of last session |
| last_transaction | Timestamp of last transaction |
| sessions | Total number of sessions (complete lookback window)  |
| sessions_1d_7d | Sessions between last 1 and 7 days |
| sessions_7d_30d | Sessions between last 7 and 30 days |
| sessions_30d_90d | Sessions between last 30 and 90 days |
| sessions_engaged | Total number of engaged sessions ([see definition](https://support.google.com/analytics/answer/11109416?hl=en)) |
| engagement_time_avg | Average engagement time of engaged sessions (complete lookback window) |
| logins | Sessions with a login |
| total_transactions | Total number of transactions |
| total_transaction_value | Sum of transaction values |
| device_types | Number of unique device categories (mobile / tablet / desktop) used |
| channels | Number of unique channel (groupings) used |
| days_since_last_session | Days between current date and last session timestamp |
| days_since_last_transaction | Days between current date and last transaction timestamp | 
| device_first | First device (within lookback period) |
| device_last | Last device |
| device_most_freq | Most frequently used device |
| os_last | Last used operating system |
| os_most_freq | Most frequently used operating system |
| os_browser_last | Last used browser |
| os_browser_most_freq | Most frequently used browser 
| channel_acquisition | Acquisition channel (channel of first touch session, can be outside the lookback window).  |
| channel_first | First used channel (within lookback period ) |
| channel_last | Last used channel. A session can have multiple channels,the first channel within a session is used. |
| channel_most_freq | Most frequently used channel. A session can have multiple channels,the first channel within a session is used. |
| last_viewed_items| Last 5 items viewed. Deduplicated. |
| most_viewed_items | Most viewed items |
| mosted_viewed_categories | Most viewed categories, based on items views |

## Query: The customer view

> Keep in mind, this is a rather resource intensive query, so when used on large datasets make sure you select a small amount of data tot test. I'm also aware that it isn't he most efficient query (tips welcome!). In production, you could split up the query (for example create a separate sessions table that is updated daily, so you don't have to query the complete raw GA4 dataset every day).

*I've included a temporary function for defining channel groupings (based on traffic source and medium), but you could also use a [permanent function (see this article)](https://stacktonic.com/article/google-analytics-4-and-big-query-create-custom-channel-groupings-in-a-reusable-sql-function).*

[View the code on GitHub](https://gist.github.com/krisjan-oldekamp/b3b98ecc579fa40a8fdc48a84256a2de)

```sql
declare lookback_window int64 default 365; -- how many days to lookback into the ga4 dataset to calculate profiles

-- udf: channel grouping (you could put this in a permanent function)
-- also see https://stacktonic.com/article/google-analytics-4-and-big-query-create-custom-channel-groupings-in-a-reusable-sql-function
create temporary function channel_grouping(tsource string, medium string, campaign string) as (
    case
        when (tsource = '(direct)' or tsource is null) 
            and (regexp_contains(medium, r'^($not set$|$none$)$') or medium is null) 
            then 'direct'
        when regexp_contains(medium, r'^(social|social_advertising|social-advertising|social_network|social-network|social_media|social-media)$') 
            then 'social'
        when regexp_contains(medium, r'^(email|mail)$') 
            then 'email'
        when regexp_contains(medium, r'^(affiliate|affiliates)$') 
            then 'affiliate'
        when regexp_contains(medium, r'^(cpc|ppc|paidsearch)$') 
            then 'search_paid'
        when regexp_contains(medium, r'^(display|cpm|banner)$') 
            then 'display'
        when medium = 'organic'
            then 'search_organic'
        when medium = 'referral'
            then 'referral'
        else '(other)'
    end
);

-- udf: most frequent value from array
create temp function most_freq(arr any type) as ((
    select x from unnest(arr) x group by x order by count(1) desc limit 1
));

-- udf: get first channel within session
create temp function first_channel_session(arr any type) as ((
    select 
        channel_grouping(x.source, x.medium, x.campaign) 
    from unnest(arr) x 
    where x.ignore_referrer is null 
    order by x.event_timestamp asc limit 1
));

with 
-- this in-between table is based mapping table described in this article (you can leave out this part if you want to group on device-id only)
-- https://stacktonic.com/article/create-a-user-mapping-table-based-on-the-google-analytics-4-big-query-dataset)
-- only select ga_client_id (`user_pseudo_id`) with 3 of less customer-ids attached. 
-- when multiple customer-ids, select most recent one.
customer_id_mapping as (
    select
        ga_client_id,
        customer_ids,
        (select x.customer_id from unnest(customer_ids) x order by x.timestamp desc limit 1) as customer_id
    from (
        select
            ga_id.id as ga_client_id,
            count(customer_id) as customer_ids_count,
            array_agg(struct(customer_id, ga_id.timestamp)) as customer_ids,
        from 
            `<your-project>.<your-dataset>.<your-mapping-table>`, 
            unnest(ga_client_id) as ga_id
        group by 
            ga_client_id
    ) 
    where
        customer_ids_count <= 3
),
-- aggegrate event level data to session level data and attach customer-id from mapping table
sessions as (
    select
        user_pseudo_id as ga_client_id,
        max(m.customer_id) as customer_id, -- logic to select only one customer-id per user_pseudo_id can be found in `customer_id_mapping`
        concat(user_pseudo_id,'.',(select cast(value.int_value as string) from unnest(event_params) where key = 'ga_session_id')) as session_id, -- combine user_pseudo_id and session_id for a unique session-id
        timestamp_micros(min(event_timestamp)) as session_start,
        max((select value.string_value from unnest(event_params) where key = 'session_engaged')) as session_engaged,
        max((select value.int_value from unnest(event_params) where key = 'engagement_time_msec'))/1000 as engagement_time,
        max(lower(traffic_source.source)) as traffic_source_first,
        max(lower(traffic_source.medium)) as traffic_medium_first,
        min(timestamp_micros(user_first_touch_timestamp)) as session_start_first_touch,
        max(device.category) as device,
        max(device.operating_system) as os,
        max(device.web_info.browser) as browser,
        max(geo.country) as country,
        max(geo.city) as city,
        ifnull(first_channel_session(
            array_agg(
                if(event_name in('page_view','user_engagement','scroll'), struct(
                    event_timestamp,
                    lower((select value.string_value from unnest(event_params) where key = 'source')) as source,
                    lower((select value.string_value from unnest(event_params) where key = 'medium')) as medium,
                    lower((select value.string_value from unnest(event_params) where key = 'campaign')) as campaign,
                    (select value.int_value from unnest(event_params) where key = 'entrances') as is_entrance,
                    (select value.int_value from unnest(event_params) where key = 'ignore_referrer') as ignore_referrer
                ), null) 
            ignore nulls order by event_timestamp asc)
        ), 'direct') as first_channel_in_session,
        --countif(event_name = '<name-of-some-other-conversion-event>') as some_conversion,
        if(countif(event_name = 'login') > 0,1,0) as session_login,
        countif(event_name = 'purchase') as transactions,
        sum(ecommerce.purchase_revenue) as transaction_value,
        array_agg(
            if(event_name = 'view_item' or event_name = 'add_to_cart', struct(
                event_timestamp,
                event_name,
                (select item_id from unnest(items) limit 1) as item_id,
                (select item_category from unnest(items) limit 1) as item_category
            ), null) 
        ignore nulls order by event_timestamp desc limit 100) as item_interactions
    from
        `<your-project>.analytics_<your-ga4-property-id>.events_*` as s
    left join customer_id_mapping as m
        on m.ga_client_id = s.user_pseudo_id
    where
        _table_suffix between format_date('%Y%m%d',date_sub(current_date(), interval lookback_window day))
        and format_date('%Y%m%d',date_sub(current_date(), interval 1 day))
    group by
        user_pseudo_id,
        session_id
)

-- aggegrate to user level
select
    * except (item_interactions),
    array_to_string(
        array(select x.item_id from unnest(item_interactions) x where x.event_name = 'view_item' group by x.item_id order by max(x.event_timestamp) desc limit 5),
        '|'
    ) as last_viewed_items,
    array_to_string(
        array(select x.item_id from unnest(item_interactions) x where x.event_name = 'view_item' group by x.item_id order by count(1) desc limit 5),
        '|'
    ) as most_viewed_items,
    array_to_string(
        array(select x.item_category from unnest(item_interactions) x where x.event_name = 'view_item' group by x.item_category order by count(1) desc limit 5),
        '|'
    ) as most_viewed_categories
from (
    select
        --ga_client_id, -- uncomment when grouping on device-id instead of customer-id
        customer_id, -- remove when grouping on device-id
        min(datetime(session_start_first_touch)) as first_touch,
        min(datetime(session_start)) as first_session,
        max(datetime(session_start)) as last_session,
        max(datetime(if(transactions > 0, session_start, null))) as last_transaction,
        count(session_id) as sessions,
        countif(date_diff(current_date, datetime(session_start), day) between 0 and 7) as sessions_1d_7d,
        countif(date_diff(current_date, datetime(session_start), day) between 8 and 30) as sessions_7d_30d,
        countif(date_diff(current_date, datetime(session_start), day) between 31 and 90) as sessions_30d_90d,
        countif(session_engaged = '1') as sessions_engaged,
        safe_divide(sum(engagement_time), countif(session_engaged = '1')) as engagement_time_avg,
        sum(session_login) as logins,
        sum(transactions) as total_transactions,
        ifnull(sum(transaction_value),0) as total_transaction_value,
        count(distinct device) as devices,
        count(distinct first_channel_in_session) as channels,
        date_diff(current_date(),cast((max(session_start)) as date),day) as days_since_last_session,
        date_diff(current_date(),cast((max(if(transactions > 0, session_start, null))) as date),day) as days_since_last_transaction,
        string_agg(device order by session_start asc limit 1) as device_first,
        string_agg(device order by session_start desc limit 1) as device_last,
        most_freq(array_agg(device)) as device_most_freq,
        string_agg(os order by session_start desc limit 1) as os_last,
        most_freq(array_agg(os)) as os_most_freq,
        string_agg(browser order by session_start desc limit 1) as browser_last,
        most_freq(array_agg(browser)) as browser_most_freq,
        channel_grouping(min(traffic_source_first), min(traffic_medium_first), null) as channel_acquisition,
        string_agg(first_channel_in_session order by session_start asc limit 1) as channel_first,
        string_agg(first_channel_in_session order by session_start desc limit 1) as channel_last,
        most_freq(array_agg(first_channel_in_session ignore nulls)) as channel_most_freq,
        array_concat_agg(item_interactions limit 200) as item_interactions
    from
        sessions
    where -- remove when grouping on device-id
        customer_id is not null -- remove when grouping on device-id
    group by
        customer_id -- remove when grouping on device-id
        --ga_client_id --uncomment when grouping on device-id instead of customer-id
)
```

That's it!

## Closing thoughts & next steps

- Building a customer view enables a lot of marketing and customer service use cases. But, make sure you have the **appropriate consent** and be transparent to your users about how you process and use the data collected.
- You can run the query daily or multiple times per day. Based on the amount of data you have in GA4, keep in mind that this is probably a rather resource intensive query. You could split up the query and create an intermediate table for the session, an append data to this table instead of generating for the complete date range every time.
- If your looking to **activate audiences** (based on the customer view) in your (marketing) channels, you could check out [this article](https://stacktonic.com/article/build-your-own-cdp-activating-audiences-in-google-ads-customer-match-using-google-tentacles) where I describe how to send an audience from BigQuery to Google Ads.
- Lastly, if you want to **personalize** your website or app using (customer) data, you will need some sort of mechanism that can provide this data to your website fast. Using BigQuery directly is not a solid solution. In [this article I describe how to set this up, using Redis as a fast-read database and Google Tag Manager Server to serve the data](https://stacktonic.com/article/create-a-user-profile-and-recommender-service-using-big-query-redis-and-gtm-server).


Use Google Analytics 4 data to enrich a 360° Customer View with online behavioral data, such as visit frequency, most used device, channel and/or engagement metrics. This tutorial provides an SQL example you can use directly.

large_google-analytics-4-single-customer-view.PNG

medium_google-analytics-4-single-customer-view.PNG

small_google-analytics-4-single-customer-view.PNG

thumbnail_google-analytics-4-single-customer-view.PNG

google-analytics-4-single-customer-view.PNG

Enrich a Single Customer View with Google Analytics 4 BigQuery data

Learn how to serve user level data from BigQuery to your website or app, using Redis (low-latency), Cloud Functions and Google Tag Manager Server. The user-level profile data can be used for on-site personalization or data enrichments within GTM Server.

medium_gtm-server-profile-store.PNG

small_gtm-server-profile-store.PNG

thumbnail_gtm-server-profile-store.PNG

gtm-server-profile-store.PNG

Create a User Profile and Recommender service using BigQuery, Redis and GTM Server

 **When joining data sources on user level, in many occasions the same user has different types of identifiers (or multiple identifiers of the same type). Think of different device (cookie)-IDs, customer-IDs or lead-IDs. Grouping all these identifiers into a matching or mapping table (*identity graph*) can be useful when you want to stitch online behavior to customers and build a 360° customer view. **

This tutorial will show you how to create an **identity graph** containing all default and custom user identifiers that can be collected in Google Analytics 4, using the BigQuery exports.

- Group **all available user identifiers** in Google Analytics 4 into **one BigQuery table**.
- Create sub tables based on this master table for joining other data easily.
- Go [directly to the SQL code](#query-the-mapping-table)

## Identity Resolution: What is an Identity Graph?

An identity graph is a a unified collection of user-IDs (like customer- or device-IDs) that can be associated to a user, based on their interactions with a website, app or platform. 

Platforms like Facebook or Google have their own identity graphs, but **the context of this article is to build an identity graph based on your first-party sources**, so the identity graph will be specific for your organization. 

![identity-graph-resolution-bigquery-sql.png](https://static.stacktonic.com/small_identity_graph_resolution_bigquery_sql_09741780f7.png)

*Identity Graph of first-party identifiers found in the GA4 dataset*

## Different types of identifiers

The mapping table we're creating in this article, will relate all the (configured) default and custom user-IDs to a single user. In most cases, these identifiers will be of the following types (also see the figure in the previous chapter):

| ID type | GA4 parameter | Description |
| --- | --- |
| Device-ID / Cookie-IDs | `user_pseudo_id` | Identifies the same user over different visits (sessions). A cookie-ID (GA client-ID) for web of instance-ID for apps. |
| Custom User-IDs | `user_id` / `custom_param` | First-party user-IDs such as customer-IDs or lead-IDs. Should be configured manually (customer-ID has to be made available in your web/app data layer and send to GA4). |
| Click-IDs | `gclid` | Google Click-ID (GCLID/DCLID) or Facebook Click-ID (FBCLID). Can be used to send conversion data to these platforms for example. In most cases this value can be retrieved on the landingpage (URL parameters). |
| Transaction-IDs | `ecommerce.transaction_id` | Identifying transactions belonging to a user / customer. |

Since we're basically grouping all online behavior based on cookie-IDs (device/browser) to (deterministic) user identifiers such as a customer-ID, online behavior can be **joined with other data sets** as well, like CRM data. 

> I can't emphasize enough, but make sure you have the appropriate user consent based on what you're doing with the data (e.g. statistics vs marketing purposes). 

There are some things you should keep in mind when working with online behavioral data in GA4:

- Devices and browsers can be shared between multiple users. You can't be 100% sure that you're dealing with the same user for every session in Google Analytics 4. Devices / browsers can be shared for example. 
- Practical example: when customer #1 logs in on browser #1, a customer-ID is tied to the cookie-ID of browser #1. If that same browser is used by another user, this behavior is tied to customer #1 as well. 
- IDs can be shared (unintentionally). For example sharing links with with click-IDs (?gclid=) or some other form of user specific identifiers (commonly used in email links)
- Cookies (device-IDs) can be deleted. Both by tracking prevention mechanisms such as Apple's ITP or manually.

## Use cases & business value

The identity graph can be used to;

- Join historical (browsing) behavior, thus giving insights in the path leading to a conversion
- Provide insights in cross-device user behavior (and attribution)
- Input for modelling like purchase or churn predictions
- Joining CRM or other data sources with online behavior (thus creating the base for a [360° customer view](https://stacktonic.com/article/enrich-a-single-customer-view-with-google-analytics-4-big-query-data))

## Google Analytics 4 - Prerequisites & definitions

*This tutorial assumes you have an Google Analytics 4 account and enabled the data exports to BigQuery.*

Regarding the user identifiers that are collected in GA4:

- The `user_pseudo_id` is the device/browser-ID identifier and collected automatically.
- You need to 'manually' populate the `user_id` variable with a business specific user identifier like a customer-ID. This identifier can tie together cross-device and cross-browser behavior. Not collected automatically.
- Additionally you could add other identifiers in custom parameters such as secondary customer-IDs, lead-IDs or click-IDs (and make sure you have the appropriate user consent).
- The `ecommerce.transaction_id` parameter is populated when you use the ecommerce parameters of GA4 within the purchase event.

##  The mapping table explained

The mapping table consists of two parts:

1. Collecting all identifiers > Collecting all the configured user identifiers hidden away within the events of the GA4 dataset. The query has a variable called `lookback_window` to set the window in which to look for identifiers. You could query the complete dataset every time (resource intensive) or daily update a table with this data.
2. Grouping the identifiers > Make one identifier the unique key (in this tutorial, the customer-ID). Other identifiers will be aggregated into an array using the `ARRAY_AGG` function, including the timestamp of last recognition. 

![google-analytics-4-user-mapping-table-bigquery.png](https://static.stacktonic.com/google_analytics_4_user_mapping_table_bigquery_bb3a7120d3.png)

*Table: Example result*

We've set a limit of 100 unique alternative identifiers per customer-ID / row, since "snowballing" of user-identifiers can occur (one identifiers is mapped to too many device identifiers for example) or due to some tracking issues (device-ID is reset on every page for example). 

As mentioned,  for each identifier an *array of structs* is created, containing all the unique identifiers (using the custom `DEDUP()` function), including the timestamp of the last event in which the identifier was recognized. This can be used to performs checks on "freshness" or expiration. GCLIDs can expire, for example.

Now, let's create the table!

## Query: The mapping table

This query will collect all the (configured) user-IDs within the GA4 dataset for all the configured user-ID and groups the results on the `customer_id`. 

> Keep in mind, this is a rather **resource intensive query**, so when used on large datasets make sure you select a small amount of data tot test. In production you could split up the query, one process to update the recognized identifiers (daily) and a second query that groups the data on the customer-ID (or other identifiers).

An example result (when parsed into JSON format):

``` json
{
  "customer_id": "123", 
  "ga_client_id": [
     { "id": "789", "timestamp": "123123" },
     { "id": "189", "timestamp": "123123" }
  ],
  "customer_id_secondary": [
      { "id": "897", "timestamp": "123123" }
  ],
  "gclid": [
      { "id": "89134", "timestamp": "123123" },
      { "id": "72347", "timestamp": "123123" }
  ],
  "...": []
}
```

The mapping table SQL:

*[Also available on GitHub](https://gist.github.com/krisjan-oldekamp/0a1a1c0085e6e6f68306f184ed1af4a1)*

``` sql
declare lookback_window int64 default 90; -- how many days to lookback into the dataset to search for ids (compared to today)

-- udf: deduplicate array of struct 
create temp function dedup(arr any type) as ((
    select 
        array_agg(t)
    from (
        select max(a.timestamp) as timestamp, a.id from unnest(arr) a group by a.id order by timestamp desc limit 100
    ) t
));

with ga_user_ids as (
  -- select all user-ids from the ga4 dataset in the specified time period
    select
        *
    from (
        select
            user_pseudo_id as ga_client_id, -- device-id
            user_id as customer_id, -- custom defined user-id (e.g. customer-id)
            (select value.string_value from unnest(user_properties) where key = 'relay_id') as customer_id_secondary, -- secondary custom user-id
            (select value.string_value from unnest(event_params) where event_name = 'page_view' and key = 'gclid') as gclid, -- click-id
            ecommerce.transaction_id as transaction_id, -- order-ids
            max(event_timestamp) as event_timestamp -- timestamps
        from `<your-project>.analytics_<your-dataset>.events_*`
        where
            _table_suffix between format_date('%Y%m%d',date_sub(current_date(), interval lookback_window day))
            and format_date('%Y%m%d',date_sub(current_date(), interval 1 day))
        group by
           1,2,3,4,5
    )
    -- filter out users without any additional user-ids besides ga_client_id
    where
        customer_id is not null
        or customer_id_secondary is not null
        or gclid is not null
        or transaction_id is not null
)

-- select all identifiers and group on customer-id.
select
    customer_id, 
    max(timestamp_micros(event_timestamp)) as timestamp_last_visit,
    -- aggegrate customer-ids to an array of structs
    dedup(array_agg(
        if(ga_client_id is not null,    
            struct(timestamp_micros(event_timestamp) as timestamp, ga_client_id as id),
            null
        ) ignore nulls order by event_timestamp desc
    )) as ga_client_id,
    -- aggegrate secondary customer-ids to an array of structs
    dedup(array_agg(
        if(customer_id_secondary is not null,
            struct(timestamp_micros(event_timestamp) as timestamp, customer_id_secondary as id),
            null
        )ignore nulls order by event_timestamp desc
    )) as customer_id_secondary,
    -- aggegrate gclids to an array of structs
    dedup(array_agg(
        if(gclid is not null, 
            struct(timestamp_micros(event_timestamp) as timestamp, gclid as id), 
            null
        ) ignore nulls order by event_timestamp desc
    )) as gclid,
    -- aggegrate transaction-ids to an array of structs
    dedup(array_agg(
        if(transaction_id is not null, 
            struct(timestamp_micros(event_timestamp) as timestamp, transaction_id as id), 
            null
        ) ignore nulls order by event_timestamp desc
    )) as transaction_id,
from ga_user_ids
where
    customer_id is not null
group by 1
```
## Query: Create an ID specific match table

Based on the result of the previous table, you could create a match table that can quickly provide a customer-ID for a device-ID (`user_pseudo_id`).

![google-analytics-user-matching-table.png](https://static.stacktonic.com/google_analytics_user_matching_table_572ebb79f0.png)

``` sql
select
    ga_id.id AS ga_client_id,
    customer_id
from `<your-project>.<your-dataset>.<your-mapping-table>`, 
    unnest(ga_client_id) as ga_id
group by 1, 2
```

## Summary & next steps

The examples provided in this tutorial can serve as input for a lot of interesting use cases as mentioned earlier. Again, make sure you have the appropriate consent. Some next steps:

- Group mapping table on other user identifiers (for example on device-ID)
- Automate the generation of the mapping table. As mentioned, my advise would be to add new identification events to a master table (daily) and from there build your mapping table
- Use the mapping table to create a *single customer view* or *single user view*, aggregating data into a user profile, which can be used for targeting or personalization.

Learn how to create a user mapping table (or Identity Graph) based on all the available user identifiers in the Google Analytics 4 BigQuery exports (like device-IDs or customer-IDs). The Identity Graph can be useful when you want to stitch online behavior to customers and build a 360° customer view.

large_identity-mapping-graph.png

medium_identity-mapping-graph.png

small_identity-mapping-graph.png

thumbnail_identity-mapping-graph.png

identity-mapping-graph.png

Create a User Mapping table based on the Google Analytics 4 BigQuery dataset

**Making data backups is important. But what about your valuable SQL code, hidden away in views and scheduled queries? This tutorial will show you how to backup all the SQL code into a Git repository using Python.**

When a lot users are actively participating in the development of your Google Cloud Platform (GCP) environment, it's hard to track changes to views and scheduled queries. This solution will;

- **Backup** your BigQuery **views** and **scheduled queries** to `.sql` files on your local filesystem. 
- Ability to commit changes to a **Git** repository. When running the script regularly, you can track **change history**.
- Supports **multiple** GCP projects.
- Go directly [to the code](#the-python-script)

## Preparation

We will use a Service Account to authenticate with your Google Cloud project(s). 

> You can use the same service account for all the projects by adding this service account with the correct role(s) to all the GCP projects you would like to backup.

If you're not familiar with setting up a Service Account, read the [Google documentation](https://cloud.google.com/iam/docs/creating-managing-service-accounts) first.

- Create a Service Account and enable the `BigQuery User` role for the account.
- Generate and download the JSON keyfile for the Service Account.
- Enable the `Data Transfer API` within your GCP project

## How the script works

- The Python script will fetch the SQL code of all the views and the scheduled queries within the GCP projects configured in the script.
- Files are downloaded to your local filesystem. 
- When you enable the `REPO_COMMIT` setting, the script will clone a (remote) repository first, fetches the views and scheduled query SQL and commits the changes to the repo. Before it will clone the remote repository, it will delete an old version of the repo first

> When using a Git repository to backup your SQL and track change history, you can use a SSH key for authentication. See instructions for [GitHub](https://docs.github.com/en/github/authenticating-to-github/connecting-to-github-with-ssh/adding-a-new-ssh-key-to-your-github-account) or [Bitbucket](https://support.atlassian.com/bitbucket-cloud/docs/set-up-an-ssh-key/)

** File structure **

| Entity | Folder |
| --- | --- |
| Views | `<backup-folder>/<project-id>/bq_views/<dataset-name>/<table-name>.sql` |
| Scheduled queries | `<backup-folder>/<project-id>/bq_scheduled_queries/<scheduled-query-name>.sql` |

![bigquery-backup_file_structure.PNG](https://static.stacktonic.com/bigquery_backup_file_structure_b7a80f79e4.PNG)

## Installation and running the code

** Installing packages **

The script requires some dependencies. Use `pip` to install these packages:

- `pip install google-auth`
- `pip install google-cloud-bigquery`
- `pip install GitPython`

Last step is to adjust the settings in the script: 

| Setting | Description |
| --- | --- |
| LOCAL_PATH_BACKUP | Local path to your backup (or local repository) folder. Use a trailing slash |
| REPO_COMMIT | When set to True, the script will clone the Git repository specified in the REPO_LINK setting. Then it will delete all the folders in the repo, download the latest version of the views and scheduled queries to the repo and commits the changes. |
| REPO_KEY_PATH | Path to your SSH key. Leave blank when not using an SSH key. |
| GCP_JSON_KEYPATH | Path to the Service Account JSON keyfile |
| GCP_PROJECTS_BACKUP | Specify the GCP project-IDs (and locations) that the script has to check for views and queries. For the `location` value see the *region* column in the GCP Scheduled Query interface |

Now you're all set to run the code:

- `python backup_views_scheduled_queries.py`

## The Python script

Save the code to a file manually or download the code [directly from our GitHub repository](https://gist.github.com/krisjan-oldekamp/78869851421af2f75325c32302fa2137).

```python
############################################################
# Author    Krisjan Oldekamp / Stacktonic.com              
# Email     krisjan@stacktonic.com                         
# Article   https://stacktonic.com/article/backup-your-valuable-big-query-views-and-scheduled-queries-using-python
############################################################

import os
import git
import google.oauth2.service_account
from google.cloud import bigquery
from google.cloud import bigquery_datatransfer # Also enable the Data Transfer API in the GCP console

############################################################
# Settings

LOCAL_PATH_BACKUP = "./gcp-bigquery-sql-backup/" # Backup directory (end with a trailing /). When REPO_COMMIT is set to True, this folder will also be used to clone and commit to the Git repository

REPO_COMMIT = True # When set to True, the script will clone the Git repository specified in the REPO_LINK setting. Then it will delete all the folders in the repo, download the latest version of the views and scheduled queries to the repo and commits the changes.
REPO_LINK = "git@bitbucket.org..." # Repository link
REPO_KEY_PATH = "../.ssh/..." # Path to SSH private key used for authentication with the repository. Leave empty when not using an SSH key.

GCP_JSON_KEYPATH = "your-keyfile.json" # Path to JSON keyfile for Service Account

# You can specify multiple projects. The location / region is required for downloading scheduled query SQL.
GCP_PROJECTS_BACKUP = [
    {
        "project_id": "your-gcp-project-id",
        "location": "europe" # See the "region" column in the GCP scheduled query interface
    }
]

# End of settings
############################################################

# Authenticate use a Service account / JSON keyfile
def get_credentials(keyfile_json):
    scopes=['https://www.googleapis.com/auth/bigquery']

    return (
        google.oauth2.service_account.Credentials.from_service_account_file(keyfile_json, scopes=scopes)
    )

# Format scheduled query name so it can be used in a filename
def format_name(name):
    return name.lower().replace(" ", "_")

# Push commit to repository
def git_push_commit(repo, repo_key_path, repo_local_path):
    try:
        if repo_key_path == "":
            repo.git.add(all=True)
            repo.index.commit("Automated commit containing changes to BigQuery views and scheduled queries.")
            repo.remotes.origin.push()
        else:
            with git.Git().custom_environment(GIT_SSH_COMMAND="ssh -i " + repo_key_path):
                repo.git.add(all=True)
                repo.index.commit("Automated commit containing changes to BigQuery views and scheduled queries.")
                repo.remotes.origin.push()
    except:
        print("Some error occured while pushing the commit")    

# Save Google BigQuery views to local filesystem
def save_bigquery_views(credentials, project_id, path):
    
    client_bq = bigquery.Client(credentials=credentials, project=project_id)
    datasets = list(client_bq.list_datasets())
    cnt_views = 0

    # Loop datasets
    if datasets: 
        for dataset in datasets: 
            dataset_name = dataset.dataset_id
            dataset_ref = dataset.reference
            tables = list(client_bq.list_tables(dataset_ref))

            # Loop tables in dataset
            for table in tables:
                # Only select VIEW
                if table.table_type == "VIEW":
                    table_name = table.table_id
                    table_ref = dataset_ref.table(table_name)
                    table = client_bq.get_table(table_ref)
                    
                    backup_directory = path + project_id + "/bq_views/" + dataset_name
                    if not os.path.exists(backup_directory):
                        os.makedirs(backup_directory) # Create directory when it doesn't exist
                    
                    # Save view SQL to file
                    f = open(backup_directory + "/" + table_name + ".sql", "w+") # Save view SQL to file
                    f.write(table.view_query)
                    cnt_views+=1

    return cnt_views

# Save Google BigQuery scheduled queries to local filesystem   
def save_bigquery_scheduled_queries(credentials, project_id, location, path):
    
    client_bq_df = bigquery_datatransfer.DataTransferServiceClient(credentials=credentials)
    parent = "projects/{}/locations/{}".format(project_id, location)
    cnt_scheduled_queries = 0

    # Loop all Data Transfer elements in project
    for element in client_bq_df.list_transfer_configs(parent):
        
        scheduled_query_name = format_name(element.display_name)
        scheduled_query_sql = ""
        
        # Loop over Data Transfer parameters, select scheduled queries and get query SQL
        params = element.params.items()
        for key, value in params: 
            if key == "query":
                scheduled_query_sql = value

        backup_directory = path + project_id + "/bq_scheduled_queries/"
        if not os.path.exists(backup_directory): # Create directory when it doesn't exist
            os.makedirs(backup_directory)

        # Write to file
        f = open(backup_directory + "/" + scheduled_query_name + ".sql", "w+")
        f.write(scheduled_query_sql)
        cnt_scheduled_queries+=1

    return cnt_scheduled_queries

def execute():
    
    credentials = get_credentials(GCP_JSON_KEYPATH)
    
    print ("Found {} GCP projects in settings".format(len(GCP_PROJECTS_BACKUP)))

    if REPO_COMMIT:
        # Clone repository if Git commits are enabled
        print("Git commits enabled. Cloning repository {} to {}".format(REPO_LINK, LOCAL_PATH_BACKUP))
        
        # Delete folder / repository when existing
        if os.path.exists(LOCAL_PATH_BACKUP):
            git.rmtree(LOCAL_PATH_BACKUP)
        
        if REPO_KEY_PATH == "":
            repo = git.Repo.clone_from(REPO_LINK, LOCAL_PATH_BACKUP)     
        else:
            repo = git.Repo.clone_from(REPO_LINK, LOCAL_PATH_BACKUP, env={"GIT_SSH_COMMAND": "ssh -i " + REPO_KEY_PATH}) # Use SSH key
    
        # Remove old backups in repository
        repo_dirs = next(os.walk(LOCAL_PATH_BACKUP))[1]
        for dir in repo_dirs:
            if not dir.startswith("."):
                git.rmtree(os.path.join(LOCAL_PATH_BACKUP, dir))
    else:
        # Only download to local filesystem. Create directory when it doesn't exist
        if not os.path.exists(LOCAL_PATH_BACKUP):
            os.makedirs(LOCAL_PATH_BACKUP)

    # Loop through GCP project and save views and scheduled queries
    for i, project in enumerate(GCP_PROJECTS_BACKUP):

        print ("-- Starting backup for project: {}".format(project["project_id"]))
        
        views = save_bigquery_views(
            credentials, 
            project["project_id"], 
            LOCAL_PATH_BACKUP
        )
        print("# {} views saved..".format(views))

        scheduled_queries = save_bigquery_scheduled_queries(
            credentials, project["project_id"], 
            project["location"], 
            LOCAL_PATH_BACKUP
        )
        print("# {} scheduled queries saved..".format(scheduled_queries))

    # Push code to remote repository
    if REPO_COMMIT:
        git_push_commit(repo, REPO_KEY_PATH, LOCAL_PATH_BACKUP)
        print("Pushed code to repository..")

    print("Done.")
       
execute()
```

## Some last tips

- You probably want to schedule this script daily. You can use your favorite scheduling tool like [Apache Airflow](https://stacktonic.com/article/apache-airflow-2-0-examples-a-basic-dag-template-for-any-project) or good ol' cronjobs.
- On windows, configuring SSH keys can give you some headaches. See the [GitHub](https://docs.github.com/en/github/authenticating-to-github/connecting-to-github-with-ssh/generating-a-new-ssh-key-and-adding-it-to-the-ssh-agent) and the [Bitbucket](https://support.atlassian.com/bitbucket-cloud/docs/troubleshoot-ssh-issues/) documentation for troubleshooting.

Making data backups is important. But what about your valuable SQL code, hidden away in views and scheduled queries? This tutorial will show you how to backup your SQL code into a Git repository using Python.

large_google-bigquery-backup-views-scheduled-queries.jpg

medium_google-bigquery-backup-views-scheduled-queries.jpg

small_google-bigquery-backup-views-scheduled-queries.jpg

thumbnail_google-bigquery-backup-views-scheduled-queries.jpg

google-bigquery-backup-views-scheduled-queries.jpg

Backup your valuable BigQuery Views and Scheduled Queries using Python

**In this tutorial we will show you how to activate audiences in Google Ads Customer Match lists using the Google Tentacles framework. The framework can be used activate your (first-party) data in marketing and advertising platforms. Basically building a headless Customer Data Platform (CDP) within an (existing) cloud / data environment without an expensive CDP tool or suite**

- Automate the audience push to Google Ads (Customer Match lists), based on (hashed) email address or other available identifiers.
- Google Tentacles runs "serverless" within the Google Cloud Platform and starts running when a new file is uploaded in a specific Storage bucket.
- These files are new line delimited JSONs, containing the create and remove to audience operations. These files can be generated by your data warehouse (e.g. Google BigQuery or Dataflow).
- Can be used for a lot of common use cases (pushing audiences or conversion events to external platforms).
- The framework itself is focused on Google products (what a suprise!), but connectors for non-Google products can be developed easily (e.g. Facebook Conversion API or Custom Audience connector).

## What is Tentacles?

[Tentacles](https://github.com/GoogleCloudPlatform/cloud-for-marketing/blob/main/marketing-analytics/activation/gmp-googleads-connector/README.md) is an (unofficial) Google framework, operating **serverless** using different services within the Google Cloud Platform, like Cloud Functions & PubSub. 
The **integration framework** is able to handle large amounts of data to be send to other, mostly marketing related Google products (like Google Analytics data upload, Google Campaign conversions or Google Ads Custom Audiences). 
Next to all the Google related connectors, the framework is handling tasks like `queueing`, `batching`, `rate-limits` and `(error) logging` and therefore can be used perfectly for building *custom connectors*, like a Facebook Custom Audience connector we've build.

More background information on the framework [in this article](https://cloud.google.com/architecture/serverless-integration-solution-for-google-marketing-platform).

![google-tentacles-serverless-framework-headless-cdp.png](https://static.stacktonic.com/google_tentacles_serverless_framework_headless_cdp_1ed8fd795c.png)

*Architectural layout of the serverless integration framework. Source: [Google](https://cloud.google.com/architecture/serverless-integration-solution-for-google-marketing-platform)*

## 1. Preparation

If not present already, create a Google Cloud project ([instructions](https://cloud.google.com/resource-manager/docs/creating-managing-projects)) where you can run Tentacles. Tentacles provides you with a bash install script, that will check and activate all services / APIs needed. It assumes you're an project administrator, so if it's not possible to grant administrator rights, at least you need the following permissions / roles;

- Service Account User
- Cloud Functions Developer
- Pub/Sub Editor
- Storage Admin
- Service Usage Admin
- Cloud Datastore User
- Service Management Administrator

The following services / APIs are used by Tentacles, if not activated already, the install script will enable it for you;

> In new GCP projects, you have to select a Datastore / Firestore operation mode. For Tentacles, **choose the Firestore Native mode**. You can't change this setting later on, so make sure you set it to Firestore Native mode.

- **Google Ads API** (you have to enable this manually in your project)
- Google Cloud Storage
- Firestore (Native mode)
- Google Cloud Functions 
- Google Pub/Sub

**Create an OAuth Client**

1. In GCP navigate to Google APIs & Services > Credentials
2. If not present already, configure the consent screen and fill out the required fields.
3. Create new OAuth 2.0 Client credentials. When creating, select the **Desktop app**
Create )Desktop app
4. Write down the `Client-ID` and `Client-Secret`, you will need it during the installation.

** Google Ads API basic access**

In Google Ads, **create an API developer token** and apply for API basic access. This will take some days. [See documention](https://developers.google.com/google-ads/api/docs/access-levels)

## 2. Installation / deployment

Let's move on the deployment of Tentacles.

1. In your Google Cloud project, open up the Cloud shell (top right).

2. Clone the repository to your Cloud shell instance;

```bash
git clone https://github.com/GoogleCloudPlatform/cloud-for-marketing.git
```

3. Run the install script and follow the instructions.

```bash
cd cloud-for-marketing/marketing-analytics/activation/gmp-googleads-connector
bash deploy.sh
```

- The namespace chosen will be the prefix of all services created by Tentacles
- When asked for the APIs (connectors) to install, choose "Google Ads API for Customer Match Upload" (option 5).
- Enter the Client-ID and Cliend-Secret you've created in the earlier steps, this will generate a verification link that results in a response code you need to paste back into the install script.

The install script will enable all the services if not already, creating the Pub/Sub topics, Cloud Functions, Cloud bucket and triggers. Next step is to configure the Google Ads Customer Match connector.

## 3. Configure the Google Ads connector

The Google Ads connector uses the Google Ads API (not the old Google Adwords API). It expects;

- Two separate JSON files per audience, co (per audience),  with one of the available customer identifiers, like (hashed) email or telephone. See [this page](https://developers.google.com/google-ads/api/reference/rpc/v7/UserIdentifier) for all available identifiers.
- One file containing the `create` (add) to audience operations
- One file containining the `remove` from audience operations

> Be aware that you're uploading customer data to Google, so make sure you have the correct consent / permissions from those users.

**Connector configuration**

Every Google Ads audience needs to be defined in the `config_api.json` file. For the same audience you have to create 2 configurations, one for `create` operators and one for `remove` operators. 

> As I'm told, this is because Tentacles mimics the behavior of the Google Ads API. The Google Ads API separates create and delete operations. Next to that, using multiple identifiers (like hashed email and phone) also results in additional configurations, because the Google Ads API also separates these operations.

| Setting | Description |
|---|---|
| developerToken | Google Ads developer token. Make sure you have Basic Access to the Google Ads API |
| customer_id | Google Ads account-ID where you want the audiences to be created |
| login_customer_id | Google Ads MCC account-ID, when accessing the account through an MCC. |
| list_id| Google Ads Customer Match list ID > You will have to create the audience first in the interface and fetch the ID. Hopefully, automatic audience creation will be added in the future  |
| list_type | Type of identifiers you wish to upload to this audience, for example `hashed_email` or `hashed_phone_number`. See [this page](https://developers.google.com/google-ads/api/reference/rpc/v7/UserIdentifier). |
|operation| `create` or `remove`, so you have to create two separate configurations for the same audience |


**Example `config_api.json` file. **

*ACM: Ads Customer Match configurations (the config_api.json can also contain configurations for the other connectors*

``` json
{
  "ACM": {
    "audience_1_add": {
        "developerToken": "<GOOGLE-ADS-API-DEVELOPER TOKEN>",
        "customerMatchConfig": {
            "customer_id": "<ADS-CUSTOMER-ID>",
            "login_customer_id": "<ADS-MCC-CUSTOMER-ID>",
            "list_id": "<CUSTOM-AUDIENCE-ID>",
            "list_type": "hashed_email",
            "operation": "create"
        }
    },
    "audience_1_remove": {
        "developerToken": "<GOOGLE-ADS-API-DEVELOPER TOKEN>",
        "customerMatchConfig": {
            "customer_id": "<ADS-CUSTOMER-ID>",
            "login_customer_id": "<ADS-MCC-CUSTOMER-ID>",
            "list_id": "<CUSTOM-AUDIENCE-ID>",
            "list_type": "hashed_email",
            "operation": "remove"
        }
    }
  }
}
```

1. Go back to the Google Cloud shell. 
2. You can use the file editor (Open Editor button), navigate to `cloud-for-marketing/marketing-analytics/activation/gmp-googleads-connector` 
3. Open `config_api.json` and edit the contents / paste your Google Ads Custom Audience configuration.

Then, run the following command to upload or update the config into Tentacles (the bash script will load or update the configuration into the Firestore database).

```
bash deploy.sh update_api_config
```

> The install script doesn't remove entries  in the Firestore database, only adds or updates.

Now, Tentacles is all set to receive the JSON files containing the create and remove operations for Google Ads audiences.

## 4. Automate the Google Ads import files

The last step is to automate the generation of the newline delimited JSON files. The naming of these files should like like `API[ACM]_config[audience_1_add]_<YYYYMMDD>.json`, where the connector key (ACM) and config key (audience_1_add) should match the entries in `config_api.json`

The content of the newline delimited JSON files should look like;

``` json
{ "hashed_email": "f660ab912ec121d1b1e928a0bb4bc61b15f5ad44d5efdc4e1c92a25e99b8e44a" }
{ "hashed_email": "bceede42d53175d644efb79514d3967d3f9b92eb810b2c30bcddfae571b8b377" }
{ "hashed_email": "8d1dc4e440a9965c3388c8714d3dd25ce99d855fd6828396f0fab07f57fba820" }

```

Most likely, you will have to execute the following tasks;

- Create a selection (audience) based on available data in your data warehouse (e.g. "all high-value customers")
- Calculate delta's / incremental tables, containing the `create` and `remove` audience operations
- When using BigQuery, you could create **date sharded** (partition) tables (audience_1_YYYMMMDD), holding all the the **SHA256** hashed emails of that date for the specific audience. 
- Based on these date sharded tables, you can calculate the differences between the most recent audience state and audience state of the previous date. 
- With these differences/deltas you can **create two intermediate tables**, one containing the **SHA256** hashed emails for the `create` operations (new users in audience), the other containing the `delete` operations (users need to be deleted).
- Export the 2 intermediate tables to JSON files (newline delimited), using the correct filename > `API[ACM]_config[audience_1_add]_<YYYYMMDD>.json` (can be exported directly from BigQuery)
- Move file to the Google Cloud bucket where the Tentacles Cloud Function is listening to (in the `/outbound` folder)

You could use the following SQL to calculate differences based on two date sharded tables (current date vs. previous date):

``` sql
#standardSQL
SELECT
  operator AS operator,
  JSON_EXTRACT_SCALAR(record, '$.hashed_email') AS hashed_email
FROM (
  -- Get new / adjusted records compared to previous table -> CREATE operations
  SELECT 
    *,
    'create' AS operator, 
  FROM (
    SELECT TO_JSON_STRING(a) AS record FROM `{table_name}_*` a WHERE _TABLE_SUFFIX = '{current_date}' -- New table
    EXCEPT DISTINCT
    SELECT TO_JSON_STRING(b) AS record FROM `{table_name}_*` b WHERE _TABLE_SUFFIX = '{prev_date}' -- Old table
  )
  UNION ALL
  -- Get missing records in current table, compared to previous table -> REMOVE operations
  SELECT
    *,
    'remove' AS operator, 
  FROM (
    SELECT TO_JSON_STRING(b) AS record FROM `{table_name}_*` b WHERE _TABLE_SUFFIX = '{prev_date}' -- Old table
    EXCEPT DISTINCT
    SELECT TO_JSON_STRING(a) AS record FROM `{table_name}_*` a WHERE _TABLE_SUFFIX = '{current_date}' -- New table
  )
)
```

There are a lot of solutions possible automating these tasks, but you could use one of the following setups;

1. **Apache Airflow**

Apache Airflow has all the connectors available to automate these tasks (create audience tables, create intermediate / delta tables) and export it in JSON format to the Tentacles Cloud bucket. Using Apache Airflow, you can create a scalable solution that can handle a lot of audiences.

2. **Scheduled Queries + Cloud Scheduler + Cloud Functions**

- Use BigQuery Scheduled Queries to generate the daily full audience tables and the delta tables (see example SQL above).
- Use Cloud Functions to run a function that starts the BigQuery export job to a JSON file in the Tentacles bucket. See [this code](https://cloud.google.com/bigquery/docs/samples/bigquery-extract-table-json).
- Let Cloud Scheduler run the Cloud Function after the scheduled queries ran.

## Summary

Tentacles will enable you to activate your data directly out of your cloud environment pretty easily. It could save some money compared to expensive CDP tools / suites, but keep in mind

- It doesn't provide you with a nice user interface
- Only has some basic (error) logging
- Heavily Google focused, but customer connectors can added / developed easily

Some last tips;

- Google Ads Connector; when you using multiple identifiers (like email and telephone), you need to split up files according to the number of different identifiers used + add/ remove operations. Resulting in 4 files per audience when using hashed_email and hash_phone_number for example. Not that efficient, but you can adjust the Google Ads connector (nodeJs) without much effort, for example combining create and add operations in 1 file.
- You can debug Tentacles, using the logs provided in the Google Cloud Functions interface or the Cloud Logging / Stackdriver interface.

## Useful links & documentation

- [Tentacles GitHub repository](https://github.com/GoogleCloudPlatform/cloud-for-marketing/blob/main/marketing-analytics/activation/gmp-googleads-connector/README.md)
- [Tentacles overview & architecture by Google](https://cloud.google.com/architecture/serverless-integration-solution-for-google-marketing-platform)
- [Tentacles tutorial by Google](https://cloud.google.com/architecture/deploying-a-serverless-integration-solution-cloud-functions-cloud-pubsub)


In this tutorial we will show you how to activate audiences in Google Ads Customer Match lists, using the Google Tentacles framework. It will enable you to automate the audiences push to Customer Match lists, based on (hashed) email address or other available identifiers.

medium_tentacles.jpg

small_tentacles.jpg

thumbnail_tentacles.jpg

tentacles.jpg

Build your own CDP: Activating audiences in Google Ads Customer Match using Google Tentacles

**Use (real-time) triggers like (sports) events or weather conditions to automate your display and/or video campaigns in Google Display & Video 360 (DV360). This tutorial will show you how to automate these operations using Python and the DV360 API.**

- **Activate** and **pause** DV360 **line items**, making use of the DV360 API and Python
- Activate **manual triggers** to quickly activate a line item, that will automatically pause after a configurable amount of time.
- The script can be scheduled using a serverless function like AWS Lamba / Google Cloud Functions or a tool like Apache Airflow.
- Go directly [to the code](#changing-the-line-item-status)

## Practical use cases & business value

The new DV360 API (mid-2020) unlocked some interesting new features and actions became much more simpler to automate. 

![google-display-video-360-enable-line-item.PNG](https://static.stacktonic.com/google_display_video_360_enable_line_item_324ff943bd.PNG)

To give a few practical examples, using this functionality;

- (De)activate line items based on **weather conditions**
- (De)activate line items based on **events** like company / national events or sports matches (e.g. campaigns based on match outcomes)
- (De)activate line items based on **sales or stock levels**

Another interesting, related, feature are the "[manual triggers](https://support.google.com/displayvideo/answer/9265340?hl=en)"; a line-item can be activated quickly and it will automatically deactivate after a desired amount of time has passed (so no need to deactivate the line item trough the API or interface). 

Think of;

- Activate line items when an event has started (and automatically deactivate after X time)
- Activate line items when a commercial has aired (and automatically deactivate after X time)

Line items with manual triggers will also work using our code example (two use cases provided), but you have to attach a manual trigger on the line item within the [DV360 interface](https://support.google.com/displayvideo/answer/9265340?hl=en) first.

## Preparation & configuration

Before running the script, a few things need to be configured:

- [Create a Google Cloud project](https://cloud.google.com/resource-manager/docs/creating-managing-projects) when you don't have one already
- Activate the Google Display & Video 360 API within the Google Cloud Console

![google-display-video-360-enable-api.PNG](https://static.stacktonic.com/google_display_video_360_enable_api_272a0c8370.PNG)
- [Create a service account](https://cloud.google.com/iam/docs/creating-managing-service-accounts) in the Google Cloud console to programmatically authenticate with the different Google services (without using a user account)
- Add the service account to your Google DV360 account and grant access rights to the correct advertiser accounts (I've used the `"Standard" role` but probably can be configured more accurate).

## Changing the line item status

The Python script below will;
- Authenticate using a service account (change GOOGLE_JSON_KEYFILE to the location of your `keyfile.json`)
- Update a line-item based on `advertiser-id`, `line-item-id` and desired `line-item status`.
- Or activate a manual trigger based on `advertiser-id` and `trigger-id` (can be obtained in the DV360 interface when a trigger is created)

> Keep your JSON key file secure!

*View code [on GitHub](https://gist.github.com/krisjan-oldekamp/044945ef754a766bd7dd7390ac41e3cb)*

``` python
###################################################
# Author    Krisjan Oldekamp / Stacktonic.com
# Email     krisjan@stacktonic.com
# Article   https://stacktonic.com/article/how-to-activate-and-pause-line-items-in-google-dv-360-using-python
####################################################

import os
from urllib.error import HTTPError
from googleapiclient import discovery
from oauth2client.service_account import ServiceAccountCredentials

####################################################
# Settings
####################################################

GOOGLE_JSON_KEYFILE = "<your-keyfile>.json" # Google Cloud Platform Service Account JSON keyfile
GOOGLE_DV360_API_VERSION = "v1"
GOOGLE_DV360_API_SCOPES = ["https://www.googleapis.com/auth/display-video"]

DV360_ADVERTISER_ID = "<your-dv360-advertiser-id>"

####################################################

# Google D&V360 API service
def get_dv360_service():
    credentials = ServiceAccountCredentials.from_json_keyfile_name(
        GOOGLE_JSON_KEYFILE,
        scopes=GOOGLE_DV360_API_SCOPES)

    return discovery.build('displayvideo', GOOGLE_DV360_API_VERSION, credentials=credentials, cache_discovery=False)

# Change DV360 Line Item status
def update_lineitem_status(advertiser_id, line_item_id, status_new):

    service = get_dv360_service()

    status_dv360 = 'ENTITY_STATUS_ACTIVE' if status_new == 'ACTIVE' else 'ENTITY_STATUS_PAUSED'

    line_item_obj = {
        'entityStatus': status_dv360,
    }

    # Update the line item.
    update_line_item = service.advertisers().lineItems().patch(
        advertiserId=advertiser_id,
        lineItemId=line_item_id,
        updateMask="entityStatus",
        body=line_item_obj
    ).execute()

    print("Lineitem " + line_item_id + " updated to " + status_dv360)

# Activate a manual trigger
def manual_trigger(advertiser_id, trigger_id):

    service = get_dv360_service()

    # Activate the trigger
    update_line_item = service.advertisers().manualTriggers().activate(
        advertiserId=advertiser_id,
        triggerId=trigger_id
    ).execute()

    print("TriggerId " + trigger_id + " activated.")


# Usecase 1: Update lineitem status

line_item_id = "123" # DV360 Line-Item-ID
line_item_status_new = "ACTIVE" # New status -> ACTIVE or PAUSED

update_lineitem_status(DV360_ADVERTISER_ID, line_item_id, line_item_status_new)

# Usecase 2: Activate a manual trigger (create and attach a manual trigger in the DV360 interface first)

trigger_id = "123"

manual_trigger(DV360_ADVERTISER_ID, trigger_id)
```

## How to automate

There are endless ways to automate the script. The methods I've used;

- Wrapping the script in an Apache Airflow DAG
- Using a serverless function, like AWS Lambda or Google Cloud Functions. In our case, the Cloud function was triggered using Cloud Scheduler and the function fetched line-items and activation status from a Google Sheet.

Use (real-time) triggers like (sports) events or weather conditions to automate your display and/or video campaigns in Google Display & Video 360 (DV360). This tutorial will show you how to automate these operations using Python and the DV360 API.

large_traffic-light.jpg

medium_traffic-light.jpg

small_traffic-light.jpg

thumbnail_traffic-light.jpg

traffic-light.jpg

How to activate and pause line items in Google DV360 using Python

**Over the years I've written a lot of Apache Airflow pipelines (DAGs). Be it in a custom Apache Airflow setup or a Google Cloud Composer instance. I've created a DAG file structure (boilerplate) so that it improved consistency and collaboration within my team, which I'm sharing in this tutorial.**

- Airflow 2.0 starter template, containing practical Airflow concepts and examples
- Go directly [to the code](#the-dag-starter-template--boilerplate)

## Concepts and examples

![Airflow 2.0 DAG example template](https://static.stacktonic.com/medium_airflow_2_0_example_dag_template_abeccf07c6.PNG)

The example / template contains a lot of concepts that I use regularly. I've provided a short description and reference to most of these concepts in the comments. To summarize:

- It shows how to import default and contributed **Operators** and **Hooks** (through providers, like AWS and Google Cloud Platform)
- Grouped the configuration settings of the DAG in the header section (comes in handy especially within large DAGs)
- **Scheduling** and **backfilling** (using the `catchup` parameter)
- Using **custom Python functions** with the new Airflow 2.0 Taskflow API (instead of the PythonOperator). Also see [the tutorial on the Airflow website](https://airflow.apache.org/docs/apache-airflow/stable/tutorial_taskflow_api.html). 
- Setting **task dependency** using >> operators. Also see [this section](https://airflow.apache.org/docs/apache-airflow/stable/concepts/dags.html#task-dependencies) of the Airflow documentation
- Using [dynamic tasks](https://airflow.apache.org/docs/apache-airflow/stable/concepts/dags.html#dynamic-dags): (a selection) of tasks can be placed in a loop, for example multiple different file imports.
- Using **TaskGroups**: combining a set of Tasks in a Taskgroup, which will also be grouped in the interface. Usefull in large DAGs.
- Using [macros](https://airflow.apache.org/docs/apache-airflow/stable/macros-ref.html) like executing date `{{ ds }}` or a modified one, like execution date minus 5 days `{{ macros.ds_add(ds, -5) }}`

## The DAG starter template / boilerplate

> The starter template was originally written for Apache Airflow versions 1.9.x. We've rewritten the code for Airflow 2.x and added Airflow 2.0 and added new functionality and concepts (like the Taskflow API). The starter template for Apache Airflow version 1.9.x can be found here.

*View code [on GitHub](https://github.com/stacktonic-com/apache-airflow-dag-example-template)*

``` python
############################################################
# Author    Krisjan Oldekamp / Stacktonic.com              #
# Email     krisjan@stacktonic.com                         #
############################################################

# Libraries
import json
import os
from datetime import datetime, timedelta
import pendulum

# Airflow
from airflow.models import DAG, Variable
from airflow.utils.task_group import TaskGroup
from airflow.utils.dates import days_ago
from airflow.decorators import task

# Default Airflow operators 
# https://airflow.apache.org/docs/apache-airflow/stable/concepts/operators.html

from airflow.operators.bash import BashOperator
from airflow.operators.dummy import DummyOperator

# Contributed Airflow operators (providers) 
# https://airflow.apache.org/docs/apache-airflow-providers/operators-and-hooks-ref/index.html

#from airflow.providers.amazon.aws.transfers.gcs_to_s3 import GCSToS3Operator
#from airflow.providers.amazon.aws.transfers.local_to_s3 import LocalFilesystemToS3Operator

from airflow.providers.google.cloud.operators.bigquery import BigQueryExecuteQueryOperator
from airflow.providers.google.cloud.transfers.gcs_to_bigquery import GCSToBigQueryOperator
from airflow.providers.google.cloud.transfers.bigquery_to_gcs import BigQueryToGCSOperator
from airflow.providers.google.cloud.transfers.local_to_gcs import LocalFilesystemToGCSOperator

############################################################
# DAG settings
############################################################

LOCAL_TZ = pendulum.timezone("Europe/Amsterdam")

DAG_NAME = "stacktonic_example_dag" # DAG name (proposed format: lowercase underscore). Should be unique.
DAG_DESCRIPTION = "Example DAG by Krisjan Oldekamp / stacktonic.com"
DAG_START_DATE = datetime(2021, 10, 15, tzinfo=LOCAL_TZ) # Startdate. When setting the "catchup" parameter to True, you can perform a backfill when you insert a specific date here like datetime(2021, 6, 20)
DAG_SCHEDULE_INTERVAL = "@daily" # Cron notation -> see https://airflow.apache.org/scheduler.html#dag-runs
DAG_CATCHUP = False # When set to true, DAG will start running from DAG_START_DATE instead of current date
DAG_PAUSED_UPON_CREATION = True # Defaults to False. When set to True, uploading a DAG for the first time, the DAG doesn't start directly 
DAG_MAX_ACTIVE_RUNS = 5 # Configure efficiency: Max. number of active runs for this DAG. Scheduler will not create new active DAG runs once this limit is hit. 

############################################################
# Default DAG arguments
############################################################

default_args = {
    "owner": "airflow",
    "start_date": DAG_START_DATE,
    "depends_on_past": False,
    "email": Variable.get("email_monitoring", default_var="<FALLBACK-EMAIL>"), # Make sure you create the "email_monitoring" variable in the Airflow interface
    "email_on_failure": True,
    "email_on_retry": False,
    "retries": 2, # Max. number of retries before failing
    "retry_delay": timedelta(minutes=60) # Retry delay
}

############################################################
# DAG configuration (custom)
############################################################

SOME_CUSTOM_CONFIG = "yes"

DYNAMIC_TASKS =  [{
    "account": "123",
    "some_setting": True
},{
    "account": "456",
    "some_setting": False
}]

############################################################
# Python functions (custom) using the Taskflow API decorators (@task)
############################################################

@task
def python_function_with_input(value: str):
    print("Custom Python function, print whatever you want!")
    print(value)

############################################################
# Repeatable Airflow Operator functions (for use in dynamic tasks)
############################################################

def dynamic_operator_task(i):
    return DummyOperator(
        task_id="dynamic_task_" + DYNAMIC_TASKS[i]["account"]
    )

############################################################
# Main DAG
#############################################################

# Create DAG
with DAG(
    DAG_NAME,
    description=DAG_DESCRIPTION,
    schedule_interval=DAG_SCHEDULE_INTERVAL,
    catchup=DAG_CATCHUP,
    max_active_runs=DAG_MAX_ACTIVE_RUNS,
    is_paused_upon_creation=DAG_PAUSED_UPON_CREATION,
    default_args=default_args) as dag:

    # Start
    start = DummyOperator(
        task_id="start")

    # Bash Operator
    print_start_bash = BashOperator(
        task_id="print_start_bash",
        bash_command="echo 'hello {{ ds }}'") # Using a macro for execution date, see other macros https://airflow.apache.org/docs/apache-airflow/stable/macros-ref.html

    # Wait till finished
    wait_till_finished = python_function_with_input("All finished")

    # Complete
    complete = DummyOperator(
        task_id="complete")

    # Define a taskgroup so tasks will be grouped in the interface -> usefull when you have a lot of repeated task sequences for example.
    with TaskGroup(group_id='task_group_with_two_tasks') as task_group_with_two_tasks:
        # Set order of execution within taskgroup (see also https://airflow.apache.org/concepts.html#bitshift-composition)
        python_function_with_input('taskgroup task #1') >> python_function_with_input('taskgroup task #2')

    # Set task depencency using >> (see also https://airflow.apache.org/docs/apache-airflow/stable/concepts/dags.html#task-dependencies)
    start >> print_start_bash

    # Add dynamic (repeating) tasks (https://airflow.apache.org/docs/apache-airflow/stable/concepts/dags.html#dynamic-dags), looping the dicts from DYNAMIC_TASK and defining order of execution
    for i, dct in enumerate(DYNAMIC_TASKS):

        print_start_bash \
        >> python_function_with_input(dct["account"]) \
        >> [python_function_with_input("Execute together. This is the execution date: {{ ds }} #1"), python_function_with_input("Execute together. This is the execution date minus 5 days: {{ macros.ds_add(ds, -5) }} #2")] \
        >> wait_till_finished \
        >> task_group_with_two_tasks \
        >> dynamic_operator_task(i) \
        >> complete
```

## Learnings and considerations

Lastly, some learnings and considerations;

- Use connectors to store sensitive information like keys and passwords
- Split up your DAGs when they become to large or to complex
- DRY: Use dynamic tasks (or DAGs) when performing the same flow but with different settings (e.g. file imports for different countries)
- When using the same (custom) functionality in multiple DAG files, try to write generic, reusable functions
- Setup alert monitoring properly (tip: Slack of Microsoft Teams notifications)

Starting to write DAGs in Apache Airflow 2.0? Some useful examples and our starter template to get you up and running quickly.

large_airflow-2-0-example-dag-template.PNG

medium_airflow-2-0-example-dag-template.PNG

small_airflow-2-0-example-dag-template.PNG

thumbnail_airflow-2-0-example-dag-template.PNG

airflow-2-0-example-dag-template.PNG

Apache Airflow 2.0 Examples - A basic DAG template for any project

Cloud (data) platforms such as AWS, Google Cloud or Azure enable you to build a marketing data platform that doesn't only ingest and join data, but can also produce actionable insights and activate data in other tools and platforms. 

Building a cloud Marketing Data Platform

Google Analytics 4 is more than just a measurement and reporting tool. The event data exports to Google BigQuery (available in the free version) and the GA4 API brings a ton of new possibilities. Think of advanced analysis, joining data with other sources, building prediction models or integrate with other systems.

Google Analytics 4 Advanced

Increased privacy concerns, legislation and browser updates affecting (third-) party cookies is resulting in a new era where users regain control over their data and organizations should rethink their customer and marketing strategy. This will impact a lot of area's such as data collection, advertising, first-party data strategy, legal and UX. 