How to develop Python Google Cloud Functions

I’ve been using Google’s Cloud Functions for a project recently. My use case was building a small webscraper that periodically downloads excel files from a list of websites, parses them and write the structured data into a BigQuery database. In this blogpost I’ll show you why going serverless with cloud functions is so amazing and cost-effective. And I’ll discuss some practical problems that have solutions you can’t easily find in the documentation or stackoverflow. In this blog:

The basics

Cloud Functions are small pieces of code that execute in an event-driven manner. They do one small thing very efficiently in reaction to a trigger — usually an HTTP request. The neat thing is you manage zero infrastructure and usually only pay for the execution of your function and a few seconds of compute. You may have heard of other Functions-as-a-Service offerings such as AWS Lambda. (source)

This means you can write a python 3.7 function and deploy it without handling infrastructure, OS or networking. Google takes care of scaling, whether you call your function once a month or millions of times a day. You only pay per 100ms of functions running, which means using Cloud Functions is much cheaper than deploying on your own servers (see pricing), with a generous free tier as well.

Let’s try it out. Create a project on Cloud Console and setup billing (don’t worry; all examples below are in free tier). Install the Google Cloud SDK (brew cask install google-cloud-sdk on macOS), login and setup some defaults:

gcloud auth login
gcloud config set project <PROJECT_ID>
gcloud config set compute/region europe-west1
gcloud config set compute/zone europe-west1-b

Cloud Functions have their own independent environment, so we put them in their own folder (tip: use the function name). The folder should have a file and optionally if you need extra packages a requirements.txt file. Here is a simple project structure:

tree myproject
#> myproject
#> ├──
#> └── functions
#>     └── hello_world
#>         └──

Cloud functions are called either by events in your project cloud environment, or by certain triggers (docs). We’ll focus on HTTP trigger functions. Google uses Flask in their python runtime to handle incoming requests. This means the input for our function should be Flask request object, and the output should be a valid Flask response:

def hello_world(request):
  return 'Hello World!'

Deployment is a breeze:

cd functions/hello_world
gcloud beta functions deploy hello_world \
  --runtime python37 \
  --trigger-http \
  --region europe-west1
#> Deploying function (may take a while - up to 2 minutes)...done.
#> availableMemoryMb: 256
#> entryPoint: hello_world
#> httpsTrigger:
#>   url: https://us-central1-<YOUR PROJECT_ID>
#> ...

Visit the httpsTrigger URL to view the output in your browser, or use gcloud functions call hello_world in your terminal.

Integrate with Cloud Storage

As an example we’ll demonstrate how to integrate with a Cloud Storage bucket and create a cloud function that downloads a file for us. I’ll use gsutil to create a bucket and upload an image of the SpaceX starship test vehicle.

gsutil mb -l europe-west1 gs://<your_bucket_name>
gsutil cp ~/Downloads/starship.jpeg gs://<your_bucket_name>
#> / [1 files][  5.9 KiB/  5.9 KiB]
#> Operation completed over 1 objects/5.9 KiB.

We could just make this bucket public, or share a signed URL to download this specific file. But for practice, we’ll write the cloud function that will let us download the file. To get the image from the bucket into the python cloud function environment, we could use tempfile.gettempdir() to download it to the /tmp directory, an in-memory mount of the cloud function (source). Instead, we’ll use io.BytesIO to create an object in memory directly. We’ll use the flask.send_file() to return the image as a file:

from io import BytesIO
from flask import Flask, request, send_file
from import storage
storage_client = storage.Client()

def download_file(request):
    bucket = storage_client.get_bucket('<your bucket name>')
    blob = bucket.get_blob('starship.jpeg')
    file = BytesIO(blob.download_as_string())
    return send_file(file,
        attachment_filename =,

To interface with the bucket from python, I’m using the google-cloud-storage package. This is not installed in the python runtime so we need to add it to a requirements.txt file:

# requirements.txt

You file structure should now look like:

tree myproject
#> myproject
#> ├──
#> └── functions
#>     ├── download_file
#>     │   └──
#>     │   └── requirements.txt
#>     └── hello_world
#>         └──

And then deploy with:

cd functions/download_file
gcloud beta functions deploy download_file \
  --runtime python37 \
  --trigger-http \
  --region europe-west1

Visit the URL to download the image!

Debugging your Cloud Functions

So far it’s easy sailing. But what happens if you cloud functions starts returning Error: could not handle the request ?

To find the mistake, you have some options. In your project’s cloud console, go to cloud functions and click on your function.

  • In the general tab you can see the latest errors.
  • In the testing tab you can run your functions and see some logs.
  • The View Logs button shows python logs of your function.

This workflow becomes annoying very quickly: it can take up to 2 minutes to deploy a new cloud function. And it does not always pinpoint the problem. In my logs I had a finished with status: 'connection error' and a test run returned that Error: cannot communicate with function., both of which did not help me find the error. But there’s a better way!

Testing Cloud Functions locally

Google describes how to use unittest to mock Flask and test a HTTP-trigger python function. This is great for unit testing, but for development I preferred to write a simple Flask app so I could call my functions locally:

# (in root of project)
import os
os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = "serviceaccount.json"

from flask import Flask, request, render_template, send_from_directory
from functions.download_file import main as f_download_file

app = Flask(__name__)

def flask_download_file():
    return f_download_file.download_file(request)

if __name__ == '__main__':

Notice I set an environment variable pointing to a JSON. Our google-cloud-storage package uses these credentials to authenticate. We want to use the same permissons as the cloud function would have at runtime:

At runtime, Cloud Functions uses the service account [email protected], which has the Editor role on the project. You can change the roles of this service account to limit or extend the permissions for your running functions. (source)

Let’s download these credentials to the root of our project (and if you use git, don’t forgot to update .gitignore):

gcloud iam service-accounts keys create serviceaccount.json \
  --iam-account <PROJECT_ID>

To run our flask app:

pip install flask
export FLASK_ENV=development
flask run

You can then visit localhost:5000/download_file to see if you cloud function works, without having to wait 2 minutes to deploy! And with the credentials json already downloaded, you could also opt to develop some functionality in a notebook.


In practice, you can run into all sorts of problems. Some of them were hard to debug and fix, so sharing them here:

Zombie deployments

I had a lot of trouble getting one of my cloud functions to work. It worked perfect locally. I spent a lot of time reading about permissions, but it turns out the function was not overwritten after deployment (!!). Luckily, I’m not alone in this problem (github issue). Even deleting my function wouldn’t stop the URL from working. The solution for me was re-deploying using a different name.

Setting permissions

If you happen to have a permission problem, it’s fairly easy to solve. In Google Cloud, a policy binding is where a member (user or serviceaccount) gets attached to a role (which contains 1 or more permissions). Remember, for cloud functions the member will be the app engine service account. Next, find an appropriate predefined role for a cloud product. Here’s an example of adding a policy binding using gcloud:

gcloud projects add-iam-policy-binding \
        --member serviceAccount:<PROJECT_ID> \
        --role roles/storage.admin

Sometimes, gcloud will ask you to enable a certain API. Here’s an example for BigQuery:

gcloud services enable

Connection reset by peer

You might find an ConnectionResetError: [Errno 104] Connection reset by peer error in your logs, and it’s not helpful at all. In my case, it had to do with creating clients for storage buckets and bigquery. This SO post confirmed my suspicion, that in ~10% of the cases creating the connection throws a connection reset error. The solution is a simple retry with some random wait time:

from import storage
from import bigquery
from retrying import retry

@retry(stop_max_attempt_number=3, wait_random_min=1000, wait_random_max=2000)
def get_google_client(type):
    if type == 'storage':
        return storage.Client()
    if type == 'bigquery':
        return bigquery.Client()

storage_client = get_google_client('storage')

Finished with status ‘timeout’

If you call your HTTP function and get the very generic Error: could not handle the request, dive into the logs. You might find a Function execution took 60003 ms, finished with status: 'timeout'. I had missed it from reading the documentation, but cloud functions are capped to at most 60 seconds execution time. You can increase the timeout to up to 9 minutes. Alternatively, you need to split up your function. In my case, I had to create a separate cloud function to download and process each file on a webpage.

Error with status 500

Another error that took me some time to figure out. I was invocating many functions at the same time using asyncio, and got really vague status 500 errors. Cause: Google Cloud Functions have many different types of limits and quotas, and: A function returns an HTTP 500 error code when one of the resources is over quota and the function cannot execute. For me it was that I was not using global variables to reuse objects in future invocations (see the best practices). Another way to solve it could be to move to background functions listening to pub/sub events, or increasing the quotas for your project.

Bonus: Static sites with serverless backend

In my case I was hosting a static website with app engine and using cloud functions on the backend. I wanted to test and develop the site locally as a Flask app. In order to change the URL locally, you can use some javascript:

  var gcf_download_file = ""
  if (location.hostname === "localhost" || location.hostname === "") {
    var gcf_download_file = "/download_file"
  $('#my-download-link').attr("href", gcf_download_file);


Cloud functions are extremely flexible and offer myriad possibilities. And because each invocation of a function has it’s own environment, you can easily parallelize your code. As an example, let’s say you need to scrape 1000 websites. Write two cloud functions: one that can scrapes a website, and another that calls the former using python’s async.

Good luck!