Nightly Data Builds#

The complete ETL and tests are run each night on a Google Compute Engine (GCE) instance to ensure that any new changes merged into main are fully tested. These complete builds also enable continuous deployment of PUDL’s data outputs. If no changes have been merged into main since the last time the builds ran, the builds are skipped.

The builds are kicked off by the build-deploy-pudl GitHub Action, which builds and pushes a Docker image with PUDL installed to Docker Hub and deploys the image as a container to a GCE instance. The container runs the ETL and tests, then copies the outputs to a public AWS S3 bucket for distribution.

Breaking the Builds#

The nightly data builds based on the main branch are our comprehensive integration tests. When they pass, we consider the results fit for public consumption. The builds are expected to pass. If they don’t then someone needs to take responsibility for getting them working again with some urgency.

Because of how long the full build & tests take, we don’t typically run them individually before merging every PR into main. However, running make nuke (roughly equivalent to the full builds) is recommended when you’ve added a new year of data or made other changes that would be expected to break the data validations, so that the appropriate changes can be made prior to those changes hitting main and the nightly builds.

If your PR causes the build to fail, you are probably the best person to fix the problem, since you already have context on all of the changes that went into it.

Having multiple PRs merged into main simultaneously when the builds are breaking makes it ambiguous where the problem is coming from, makes debugging harder, and diffuses responsibility for the breakage across several people, so it’s important to fix the breakage quickly. In some cases we may delay merging additional PRs into main if the builds are failing to avoid ambiguity and facilitate debugging.

Therefore, we’ve adopted the following etiquette regarding build breakage: On the morning after you merge a PR into main, you should check whether the nightly builds succeeded by looking in the pudl-deployments Slack channel (which all team members should be subscribed to). If the builds failed, look at the logging output (which is included as an attachment to the notification) and figure out what kind of failure occurred:

  • If the failure is due to your changes, then you are responsible for fixing the problem and making a new PR to main that resolves it, and it should be a high priority. If you’re stumped, ask for help!

  • If the failure is due to an infrastructural issue like the build server running out of memory and the build process getting killed, then you need to notify the member who is in charge of managing the builds (Currently @bendnorman), and hand off responsibility for debugging and fixing the issue.

  • If the failure is the result of a transient problem outside of our control like a network connection failing, then wait until the next morning and repeat the above process. If the “transient” problem persists, bring it up with the person managing the builds.

Debugging a Broken Build#

If a build has failed, usually the VM will have shut down. You’ll have to figure out which VM it was running on and then restart it before you can do anything else.

To find the VM name, go into the Github Action listing and click on your run. The GCE_INSTANCE_NAME that gets printed in “Print action vars” is what you’re after.

Then you can go to the Google Compute Engine page and restart it.

Once that’s started, you should be able to SSH to the VM using a command like:

gcloud compute ssh pudl-deployment-tag --zone=us-west1-a

You may run into some permissions issues here, in which case you probably need the Service Account User role on your gcloud user.

Now you want to get some logs about what’s failing.

First, try docker ps - this should show two images, one pudl-ey one and one that’s the Google stackdriver-agent which handles monitoring:

CONTAINER ID   IMAGE                     <snip>  NAMES
d678f709d1f5   catalystcoop/pudl-etl...  <snip>  klt-pudl-deployment-tag-luui
aa3163671da4  <snip>  stackdriver-logging-agent

If the image is running, great! You can get logs via docker logs <container ID or name> (use docker logs -f if the process is still going and you want to stream logs from the container.)

You can also attach a shell to the container and poke around with docker exec -it <ID or name> bash. This is really helpful if something has failed and you want to try to fix the code & re-run, without having to re-run everything before the failed task.


If you use docker attach as recommended by the login message, and then hit Ctrl-C, you will interrupt the running build!

Sometimes you’ll see two containers running, but neither of them are PUDL. That’s because the VM first spins up a “loader” container that downloads the PUDL image, then exits the loader and starts the PUDL image.

If you don’t see two containers running, then there’s probably some issue with the PUDL container startup itself. To find logs about that, run sudo journalctl -u konlet-startup | tail -n 1000 | less. You should be able to see any errors that occurred during container startup, and also the container ID, which you can then boot into via docker run -it <ID> bash.

The GitHub Action#

The build-deploy-pudl GitHub action contains the main coordination logic for the Nightly Data Builds. The action is triggered every night and when new tags are pushed to the PUDL repository. This way, new data outputs are automatically updated on code releases, and PUDL’s code and data are tested every night. The action is modeled after an example from the setup-gcloud GitHub action repository.

The gcloud command in build-deploy-pudl requires certain Google Cloud Platform (GCP) permissions to start and update the GCE instance. The gcloud command authenticates using a service account key for the deploy-pudl-github-action service account stored in PUDL’s GitHub secrets as DEPLOY_PUDL_SA_KEY. The deploy-pudl-github-action service account has the Compute Instance Admin (v1) IAM role on the GCE instances to update the container and start the instance.

Google Compute Engine#

The PUDL image is deployed on a Container Optimized GCE instance, a type of virtual machine (VM) built to run containers. The pudl-deployment-dev and pudl-deployment-tag instances in the catalyst-cooperative-pudl GCP project handle deployments from the main branch and tags or manually initiated workflow_dispatch runs respectively. There are two VMs so a scheduled and a tag build can run at the same time.


If a tag build starts before the previous tag build has finished, the previous build will be interrupted.

The build VMs use the e2-highmem-8 machine type (64 GB of RAM and 8 CPUs) to accommodate the PUDL ETL’s memory-intensive steps. Currently, these VMs do not have swap space enabled, so if they run out of memory, the build will immediately terminate.

Each GCE VM has a service account that gives the VM permissions to GCP resources. The two PUDL deployment VMs share the deploy-pudl-vm-service-account. This service account has permissions to:

  1. Write logs to Cloud Logging.

  2. Start and stop the VM so the container can shut the instance off when the ETL is complete, so Catalyst does not incur unnecessary charges.

  3. Bill the catalyst-cooperative-pudl project for egress fees from accessing the bucket. Note: The catalyst-cooperative-pudl won’t be charged anything because the data stays within Google’s network.

  4. Write logs and outputs to the gs://, gs:// and s3:// buckets. The egress and storage fees of the s3 bucket are covered by Amazon Web Services’s Open Data Sponsorship Program.


The Docker image the VMs pull installs PUDL into a mamba environment. The VMs are configured to run the docker/ script. This script:

  1. Notifies the pudl-deployments Slack channel that a deployment has started. Note: if the container is manually stopped, slack will not be notified.

  2. Runs the ETL and full test suite.

  3. Copies the outputs and logs to a directory in the pudl-etl-logs bucket. The directory is named using the git SHA of the commit that launched the build.

  4. Copies the outputs to the gs:// and s3:// buckets if the ETL and test suite run successfully.

  5. Notifies the pudl-deployments Slack channel with the final build status.

The script is only intended to run on a GCE VM with adequate permissions.

How to access the nightly build outputs from AWS#

To access the nightly build outputs you can download the data directly from the s3:// bucket. To do this, you’ll need to follow the instructions for installing the AWS CLI tool.

To test that the AWS CLI tool and the bucket are working properly, run:

aws s3 ls s3:// --no-sign-request

You should see a list of directories with version names:

PRE nightly/
PRE stable/
PRE v2022.11.30/
PRE v2023.12.01/

The --no-sign-request flag allows you to make requsts to the public bucket without having to load AWS credentials. If you don’t include this flag when interacting with the s3:// bucket, aws will give you an authentication error.


If you download the files directly with aws then you’ll be responsible for updating them, making sure you have the right version, putting them in the right place on your computer, etc.

To copy these files directly to your computer you can use the aws s3 cp command, which behaves very much like the Unix cp command:

aws s3 cp s3:// ./ --no-sign-request


To reduce network transfer times, we gzip the SQLite database files, which can be quite large when uncompressed. To decompress them locally, at the command line on Linux, MacOS, or Windows you can use the gunzip command.

$ gunzip *.sqlite.gz

On Windows you can also use a 3rd party tool like 7zip.

If you wanted to download all of the build outputs (more than 10GB!) you could use cp --recursive flag on the whole directory:

aws s3 cp --recursive s3:// ./ --no-sign-request

For more details on how to use aws in general see the online documentation or run:

aws help

How to access the nightly build outputs and logs (for the Catalyst team only)#

Sometimes it is helpful to download the logs and data outputs of nightly builds when debugging failures. To do this you’ll need to set up the Google Cloud software Development Kit (SDK).

Install the gcloud utilities on your computer. There are several ways to do this. We recommend using conda or its faster sibling mamba. If you’re not using conda environments, there are other ways to install the Google Cloud SDK explained in the link above.

conda install -c conda-forge google-cloud-sdk

Log into the account you used to create your new project above by running:

gcloud auth login

Initialize the gcloud command line interface and select the catalyst-cooperative-pudl project.

If it asks you whether you want to “re-initialize this configuration with new settings” say yes.

gcloud init

Finally, use gcloud to establish application default credentials; this will allow the project to be used for requester pays access through applications:

gcloud auth application-default login

To test whether your GCP account is set up correctly and authenticated you can run the following command to list the contents of the cloud storage bucket containing the PUDL data. This doesn’t actually download any data, but will show you the versions that are available:

gsutil ls -lh gs://

You should see a list of directories with build IDs that have a naming convention: <YYYY-MM-DD-HHMM>-<short git commit SHA>-<git branch>.

To see what the outputs are for a given nightly build, you can use gsutil like this:

gsutil ls -lh gs://

804.57 MiB  2024-01-03T11:19:15Z  gs://
  5.01 GiB  2024-01-03T11:20:02Z  gs://
759.32 MiB  2024-01-03T11:19:17Z  gs://
813.52 MiB  2024-01-03T11:19:18Z  gs://
  1.65 MiB  2024-01-03T11:18:18Z  gs://
  6.94 MiB  2024-01-03T11:18:19Z  gs://
282.71 MiB  2024-01-03T11:19:02Z  gs://
 89.55 MiB  2024-01-03T11:18:40Z  gs://
  1.88 MiB  2024-01-03T11:18:18Z  gs://
  6.78 MiB  2024-01-03T11:18:18Z  gs://
  8.25 MiB  2024-01-03T11:18:20Z  gs://
 20.02 MiB  2024-01-03T11:18:22Z  gs://
731.31 KiB  2024-01-03T11:18:18Z  gs://
  1.77 MiB  2024-01-03T11:18:19Z  gs://
153.72 MiB  2024-01-03T11:18:54Z  gs://
 62.01 MiB  2024-01-03T11:18:28Z  gs://
  1.02 MiB  2024-01-03T11:18:18Z  gs://
  2.74 MiB  2024-01-03T11:18:18Z  gs://
905.31 MiB  2024-01-03T11:19:17Z  gs://
 58.41 KiB  2024-01-03T11:18:18Z  gs://
187.86 KiB  2024-01-03T11:18:18Z  gs://
  4.05 MiB  2024-01-03T11:18:19Z  gs://
     4 MiB  2024-01-03T12:09:34Z  gs://
  13.1 GiB  2024-01-03T11:21:41Z  gs://
       0 B  2024-01-03T11:18:18Z  gs://
TOTAL: 25 objects, 23557650395 bytes (21.94 GiB)

If you want to copy these files down directly to your computer, you can use the gsutil cp command, which behaves very much like the Unix cp command:

gsutil cp gs://<build ID>/pudl.sqlite ./

If you wanted to download all of the build outputs (more than 10GB!) you could use cp -r on the whole directory:

gsutil cp -r gs://<build ID>/ ./

For more details on how to use gsutil in general see the online documentation or run:

gsutil --help