Skip to content

dfinity/dre-airflow

Repository files navigation

Airflow repository of the DRE team

This repository contains:

  • The rollout dashboard.
  • Airflow content artifacts (DAGs, workflows, sensors and operators).
  • A content syncer container to help deliver these extensions.
  • A customized Airflow container to enable enhanced Airflow behaviors, such as use of CockroachDB in production. See heading In production below.
  • Documentation for working effectively with this distributed Reliability Team Airflow setup.

To effectively contribute code to this repository, you must first set up a local development environment. See below for instructions.

This code is distributed under the Apache 2.0 license, the contents of which can be found here. External contributions are not accepted at this time.

[[TOC]]

Airflow content artifacts distributed in this repository

Airflow is an extensible platform. We can write our own pluggable components for Airflow: sensors (tasks that wait until a condition is true) and operators (tasks that do something), and DAGs (workflows that tie up a series of tasks to perform work on our behalf). This repository contains all three kinds of components. Here's the lowdown on what this repo contains:

  • Our DAGs in this repository are distributed under folder dags.
  • Our custom sensors and operators are under folder operators and sensors.
  • Library code is under folder shared.
  • The local runner (bin/run-dag) automatically makes this all of this code available for execution after the local development enviroment is set up. If you write a sensor or an operator, you can run the file locally by running bin/run-dag <path to sensor/operator>. Be sure to have a section at the end of the file that makes execution conditional:
...
...
if __name__ == "__main__":
  <invoke your operator or sensor here>
  <existing operators and sensors have example code already>

Variables and connections required for workflows

The workflows (DAGs) and operators in this repository require a set of variables and connections to be added to Airflow before they are executed. These variables are retrieved at runtime. If a task in a workflow fails to retrieve a variable or a connection, the workflow perishes.

Logged in as administrator through the UI, go to Top menubar -> Admin -> Variables. Then create the following variables:

    • Key: dfinity.ic_admin.mainnet.proposer_key_file
    • Val: the contents of the release automation certificate PEM file used to roll out IC OS (ask DRE team for information, it should be in their vault)
    • Description: The proposal certificate for neuron 80 to be used on mainnet.

Now go to Top menubar -> Admin -> Connections. Then create the following connections (unless specified otherwise):

    • Connection ID: slack.ic_os_rollout
    • Connection type: Slack API
    • Description: The connection used for IC OS rollouts.
    • Slack API token: ask the DRE team for the API token to post to Slack, which should be in the DRE team vault.
    • Connection ID: airflow_logging
    • Connection type: Amazon Web Services
    • Description: Logging storage for Airflow.
    • AWS Access Key ID: the value of AWS_ACCESS_KEY_ID in K8s secret airflow-logging
    • AWS Secret Access Key: the value of AWS_SECRET_ACCESS_KEY in K8s secret airflow-logging
    • Extra: { "endpoint_url": "http://rook-ceph-rgw-ceph-store.rook-ceph.svc.cluster.local" }
    • Only create this connection if running under K8s

Variables and connections are only visible to administrators.

Developing and deploying Airflow content robustly

To ensure robust DAG runs, you must be aware of the lifecycle these artifacts undergo, once they go past quality assurance and they are deployed.

Delivery is discussed in detail under the Continuous delivery heading, but here is a brief summary:

  • You push code changes to this repository.
  • Each scheduler, worker and triggerer pod picks up, every five minutes, the most up-to-date revision of this repository. This is done by the airflow-content-syncer container running on each Airflow component instance in production (more info under the Continuous delivery heading).
  • When a change is detected, the contents are synced in the following order:
    1. the contents of plugins/operators
    2. the contents of plugins/sensors
    3. the contents of shared/*
    4. the contents of dags
  • DAGs are loaded, but runs are not started, when the scheduler notices DAGs change. DAG content is reread periodically by the DAG processor, and reloaded by the workers when tasks are dispatched by the workers. Therefore, changes directly made to DAG files will affect already-running flows. This has been experimentally proven in commit ce5ba7b.
  • The worker and triggerer pods always run the most up to date version of the operator / sensor code you wrote, when a task (sensor or operator) is started.
    • The way the worker / triggerer runs a task is by running an entirely new process that "rehydrates the DAG" with the key parameters of DAG run ID, task ID, possibly a mapped task index ID, and possibly the task data produced by previous tasks already-executed by the DAG run.
    • Concretely, this is what the worker does: Executing command in Celery: ['airflow', 'tasks', 'run', 'dag_name', 'taskgroup_name.task_name', 'run_id', '--local', '--subdir', 'DAGS_FOLDER/file_containing_dag.py', '--map-index', 'mapindex_int']
  • DAG runs are often long-running (could be days or weeks).

From this, the following facts hold:

  • If your DAG requires operator, sensor or shared code, the required code better be present and bug-free, otherwise DAG load will fail in the scheduler, and you will not be able to dispatch new DAG runs or control existing DAG runs (the UI will show you Broken DAG errors onscreen).
  • Any broken operator, sensor, or shared code will induce failures on any scheduled, deferred or future tasks that will execute after you pushed.
  • While any existing DAG runs will not alter their graph shape, any code change you make to sensors, operators and shared code will take effect immediately on currently-executing DAG runs. This means if your operator did thing X in the past, but now does thing Y, any future tasks from already-scheduled DAGs will do thing Y, instead of doing thing X.
  • "Rehydration" implies that what the task does when it "rehydrates" depends on all these parameters discussed above, so if the code changes in ways that these parameters "mean something else" or otherwise become incompatible with currently-scheduled tasks, the currently-scheduled tasks will fail in hard-to-debug ways.
  • Code changes must take into account that operators, sensors and shared code may expect certain parameters and inter-task data to be a certain type or shape. If you have a DAG with tasks A -> B and B expects A to return an int (or B pulls A's result from the XCom result table, and expects the result to be an int), then any pushes that alter B will cause all running DAGs to fail when task B launches.
  • Task IDs are very important. If you change the task ID of a task in a DAG, then any already-dispatched currently- running DAG runs that will attempt to "rehydrate" tasks will fail to "rehydrate" the task, and the DAG will simply be marked as failed.
  • Mapped task indexes are just as important as task IDs. If e.g. your DAG currently "splits into five" parallel flows (usually done with the expand method), each one of these flows is identified by a task index, and will be "rehydrated" upon execution with that task index. Accordingly, if you change a DAG such that the new rendered DAG has four or six parallel flows, or the data each index maps to changes or is reordered, any currently-executing DAG runs will either get the wrong data, or the task will not "rehydrate" successfully and the DAG will fail.

Be especially judicious and careful about the changes you make on code used by DAGs that may be running right now or may be scheduled in the near future.

DAGs

DAGs are workflows expressed in Python, which Airflow loads and enables you to either execute them manually or trigger them under certain conditions.

In production, DAGs are deployed to $AIRFLOW_HOME/dags.

Working with DAGs

Task dependencies within a DAG are typically specified with the operators >> and <<. Sample that demonstrates perfectly what that means:

first_task >> [second_task, third_task]
third_task << fourth_task

# Results in this order (read from left to right):
#
#                  second_task
#                /
# first_task   --
#                \
# fourth_task  --- third_task

The reference documentation has tips on how to specify more complex dependencies.

A DAG won't run, even if manually started, unless it is enabled. In the DAG list, you will see a switch to turn each DAG on or off.

DAGs are automatically reloaded by the standalone Airflow process a few seconds after you save changes to the files under the DAGs folder. To force a reload on the spot:

bin/reload-dags

The web interface will tell you at the top if there was a problem loading a DAG.

If you rename a DAG or its ID, you will have to delete the old DAG through the CLI or the web interface.

Testing DAGs

By convention, DAGs can be tested by running them so:

bin/run-dag dags/<DAG file name> [parameters...]

That command would test-runs your DAG. This works because there's a snippet at the end of each DAG, somewhat similar to this:

if __name__ == "__main__":
    dag.test()

The test run actually executes the operators! This means you are actually testing everything in an integrated fashion.

Test DAG runs are recorded in the local development environment's database as well. You can browse DAG runs using the web interface at http://localhost:8080/dagrun/list/

More on testing DAGs:

TBD:

  • local mock testing story (test DAG dependencies and dry-run)
  • CI/CD testing story

Operators

Operators are Python classes defined in standalone files under the operators folder. They are run by workers after the tasks that require them get dispatched to the workers.

In production, operators are deployed to $AIRFLOW_HOME/plugins/operators.

Working with operators

Your local Airflow instance should reload operator code when you make changes. If it does not, simply restart your Airflow instance.

Testing operators

Unit tests are in directory tests. You can run them directly from Visual Studio Code, or run them via Make using make test.

To run an operator, just write code at the bottom of its file that does something like this:

if __name__ == "__main__":
    import sys

    if sys.argv[1] == "my_operator":
        kn = MyOperator(
            task_id="x",
            arg1=sys.argv[2],
            arg2=sys.argv[3],
        )
        kn.execute({})

Then you can run it under the Airflow environment:

bin/run-dag plugins/operators/<operator file name>

TBD:

  • CI/CD testing story

Sensors

A sensor is a special type of operator which has one job: to wait for something to happen.

Our sensors are defined in standalone files under the sensors folder.

In production, sensors are deployed to $AIRFLOW_HOME/plugins/sensors.

Working with sensors

To have your Airflow development instance reload sensors, use the same procedure as the procedure to have Airflow reload operators.

Testing sensors

Everything under the Testing operators headline applies.

Shared code

Library code used by DAGs, operators and sensors is available under folder shared.

In production, each shared folder is deployed to $AIRFLOW_HOME/plugins.

Quality assurance

TBD: CI/CD testing story

Targets:

  • ruff validation and fixups
  • mypy validation
  • automated unit tests
  • git presubmit implementing all the above
  • integration tests
  • Gitlab pipeline validating all the above

Continuous delivery

The artifacts in this repository are delivered to the relevant Airflow production pods by way of the airflow-content-syncer container built from this repository. Delivery is not done through the container, but rather by using git clone / git fetch within the container periodically, running as a sidecar in all relevant Airflow pods.

In production, the syncer container will check which is the latest revision of the main branch of the repo containing this file, and if it differs from what is deployed in Airflow, it will redeploy from the latest main branch. The source tree can be overridden via an environment variable CONTENT_SYNCER_GIT_REPO_SOURCE on the pod, and the branch can be overridden using variable CONTENT_SYNCER_GIT_REPO_BRANCH.

The container image version is referred to as syncer_image in this K8s file. When the container image is updated, Airflow must be redeployed by updating the reference to the content syncer image in the file linked within this paragraph, then the K8s repository needs to have the update merged.

To determine if / when the artifacts have been synced, look at the log of the container airflow-content-syncer in any of the triggerer, scheduler or worker pods of the Airflow deployment (in namespace airflow). These logs can be retrieved from Loki in production (use the cluster's Grafana instance, Explore tab) by filtering for namespace = airflow and further narrowing the filter to contain the string Updat. Artifacts are not delivered simultaneously to all production pods -- there might be a divergence of up to 5 minutes between pods syncing.

TBD: test these artifacts!

  • Pipeline has to use the exact same version of Python that the Airflow container does -- possibly necessarily the same container itself! The Airflow containers use an ancient version of Python. We have set up the container with tag 2.6.2-python3.10 to be used for our prod setup.

Local development environment setup

To get the right libraries loaded into your IDE, you will need a virtual environment with them installed. Run bin/airflow setup to get yourself set up. You can then tell your IDE to use the specific venv python3 binary under the folder venv in this repository.

To actually run tests or Airflow itself, the setup subcommand will set up an Airflow directory instance. The folder is airflow under this repository.

Once setup, the bin/airflow command can be used anytime you want to invoke the Airflow CLI. It takes care of setting up the environment so everything works as expected. With it, you can run bin/airflow standalone as well as tests such as DAG tests.

The first time you run the standalone server, it will create an admin user and password and start by listening on HTTP port 8080. Note the password that appears on the terminal, and log in. After logging in, you can change the admin user password to a simple password, through the web interface on the top right corner menu.

Note: if you also followed the instructions on how to run Airflow on a VM as indicated below, and the VM is running, Airflow locally will not be able to open TCP port 8080, since the VM will have hogged the port.

Now restart any running bin/airflow standalone instances. The demo DAGs will be gone now.

To actually run the DAGs here, you will almost certainly have to create the required variables on your Airflow instance. See the heading Variables required for workflows for instructions on what to create.

Configuration

The airflow.cfg file under folder $AIRFLOW_HOME controls most aspects of the configuration in Airflow. It is an INI-formatted file.

The complete configuration reference is available here.

CLI

The CLI, named airflow lets you manage Airflow objects and do all sorts of low-level operations.

By default the CLI uses a local client that expects AIRFLOW_HOME to point to the Airflow home directory that contains the database. This can be changed in the Airflow configuration file by specifying a different value for the configuration key cli.api_client; for example, the json_client will make the CLI speak to an Airflow API server at the endpoint specified by the cli.endpoint_url configuration key.

See [https://airflow.apache.org/docs/apache-airflow/stable/howto/usage-cli.html] for general information on the CLI.

Web interface

The CLI command airflow webserver starts a Web interface (by default on port 8080) which lets you do many of the tasks that you can do with the Airflow CLI. You will need an authorized user account to log into Airflow through the Web interface.

The web server is a Flask application. The Flask parameters can be controlled through file webserver_config.py in folder $AIRFLOW_HOME. Authentication and authorization sources can be set up there as well. Documentation on this can be found here.

API server

Airflow has an API server embedded in its web server.

References:

Notes on the operation of Airflow

Changing DAGs, operators, sensors or other shared code on a running system will not affect the the execution of tasks currently executing on any flow currently running. However, they will affect the execution of future tasks instantiated by any currently-running flow (effectively, each task runs an airflow tasks ... command, which independently loads the DAG, instantiates the tasks, locates the task it is supposed to execute, and runs it — all of which is materialized at runtime). Therefore, to avoid failures, changes to long-running flows should be careful to be compatible with any running flows at the time of rollout, or should wait until the currently-running flows which may be impacted have finished.

Deploying a bare-bones Airflow instance to a VM

Some people prefer to use VMs for testing. VMs can also run programs that may not be available or installable in local development environments.

The following shellcode deploys Airflow (at a specific version) in a running Fedora 38 VM as root, using the minimalistic SQLite database.

cd /opt/airflow || { mkdir -p /opt/airflow && cd /opt/airflow ; }
dnf install -y python3-pip
test -f bin/pip3 || python3 -m venv .
bin/pip3 install "apache-airflow[celery]==2.6.1" \
  --constraint https://raw.githubusercontent.com/apache/airflow/constraints-2.6.1/constraints-3.7.txt
export PATH=$PWD/bin:$PATH
export AIRFLOW_HOME="$PWD"/var
airflow db init

Maintained up-to-date instructions are available here.

To run the Airflow web server from a terminal after installation, you must first create a user account if not already created:

cd /opt/airflow
export PATH=$PWD/bin:$PATH
export AIRFLOW_HOME="$PWD"/var
airflow users create \
  --username admin \
  --firstname Admin \
  --lastname Istrator \
  --role Admin \
  --email noreply@dfinity.org
# You will be prompted for a password.  Write it down after typing it.

Now you can run and log in (by default runs on port 8080):

cd /opt/airflow
export PATH=$PWD/bin:$PATH
export AIRFLOW_HOME="$PWD"/var
airflow standalone &

In production

Airflow in Kubernetes uses a customized container which adds functionality we need, such as the ability to connect to CockroachDB (our production database).

This container is built and pushed to this repository's container artifact repo. When a change that needs deployment must be changed, the Airflow on Kubernetes app in the Kubernetes repository must be updated.

See the defaultAirflowTag variable on the Airflow Helm chart used by that Kubernetes app.

Concepts

Architecture of Airflow

The report contains an outline of the architecture of Airflow.

Flows (DAGs)

Flows (DAGs) are directed acyclic graphs written in Python, using Airflow primitives to put them together. Airflow executes these flows for you. Like command executions or processes in a UNIX shell, each DAG can be instantiated multiple times, and each instantiation creates a DAG run.

DAGs can be discovered and loaded from the Python sys.path environment. They are also loaded and discovered from the path specified in the core.dags_folder key on the Airflow configuration file.

You can list all DAGs available using the airflow dags list command. They also appear listed under the DAGs tab of the Airflow web interface.

Tasks

Tasks are the building blocks of DAGs. Each node in a DAG is a task.

The scheduler dispatches to workers tasks to be executed during DAG runs. When the scheduler is running a DAG, each task in the DAG run becomes a task instance in the worker, invoking the corresponding operator.

What the worker runs concretely is what the task directs it to. Therefore, resources, inputs and sinks needed for tasks to run must be available to the worker for the task to succeed.

Operators

An operator is a module that specifies an operation which would normally be a task node in the DAG. Think of an operator as a template for a task.

Troubleshooting

Airflow refuses to start with database is locked

If you see on screen this, when running bin/airflow standalone:

sqlalchemy.exc.OperationalError: (sqlite3.OperationalError) database is locked
[SQL: SELECT name FROM sqlite_master WHERE type='table' ORDER BY name]
(Background on this error at: https://sqlalche.me/e/14/e3q8)

then your Airflow database is locked. After ensuring all of your Airflow processes have been killed, run bin/airflow unlockdb to recreate the database using existing data.