ac5ac81c84
The top_signatures_correlations task is getting killed off for using too much memory. This drops the top signatures we're looking at from 300 to 250 as a temporary attempt to get the job succeeding again. |
||
---|---|---|
.circleci | ||
bin | ||
docs | ||
mozetl | ||
scheduling | ||
tests | ||
.gitignore | ||
CODE_OF_CONDUCT.md | ||
LICENSE | ||
MANIFEST.in | ||
README.md | ||
setup.py | ||
tox.ini |
README.md
Firefox Telemetry Python ETL
This repository is a collection of ETL jobs for Firefox Telemetry.
Benefits
Jobs committed to python_mozetl can be scheduled via airflow or ATMO. We provide a testing suite and code review, which makes your job more maintainable. Centralizing our jobs in one repository allows for code reuse and easier collaboration.
There are a host of benefits to moving your analysis out of a Jupyter notebook and into a python package. For more on this see the writeup at cookiecutter-python-etl.
Tests
Dependencies
First install the necessary runtime dependencies -- snappy and the java runtime
environment. These are used for the pyspark
package. In ubuntu:
$ sudo apt-get install libsnappy-dev openjdk-8-jre-headless
Calling the test runner
Run tests by calling tox
in the root directory.
Arguments to pytest
can be passed through tox using --
.
tox -- -k test_main.py # runs tests only in the test_main module
Tests are configured in tox.ini
Manual Execution
ATMO
The first method of manual execution is the mozetl-submit.sh
script located in bin
.
This script is used with the EMRSparkOperator
in telemetry-airflow
to schedule execution of mozetl
jobs.
It may be used with ATMO to manually test jobs.
In an SSH session with an ATMO cluster, grab a copy of the script:
$ wget https://raw.githubusercontent.com/mozilla/python_mozetl/main/bin/mozetl-submit.sh
Push your code to your own fork, where the job has been added to mozetl.cli
. Then run it.
$ ./mozetl-submit.sh \
-p https://github.com/<USERNAME>/python_mozetl.git \
-b <BRANCHNAME> \
<COMMAND> \
--first-argument foo \
--second-argument bar
See comments in bin/mozetl-submit.sh
for more details.
Databricks
Jobs may also be executed on Databricks.
They are scheduled via the MozDatabricksSubmitRunOperator
in telemetry-airflow
.
This script runs on your local machine and submits the job to a remote spark executor. First, generate an API token in the User Settings page in Databricks. Then run the script.
python bin/mozetl-databricks.py \
--git-path https://github.com/<USERNAME>/python_mozetl.git \
--git-branch <BRANCHNAME> \
--token <TOKEN> \
<COMMAND> \
--first-argument foo \
--second-argument bar
Run python bin/mozetl-databricks.py --help
for more options, including increasing the number of workers and using python 3.
Refer to this pull request for more examples.
It is also possible to use this script for external mozetl-compatible modules by setting the --git-path
and --module-name
options appropriately.
See this pull request for more information about building a mozetl-compatible repository that can be scheduled on Databricks.
Scheduling
You can schedule your job on either ATMO or airflow.
Scheduling a job on ATMO is easy and does not require review, but is less maintainable. Use ATMO to schedule jobs you are still prototyping or jobs that have a limited lifespan.
Jobs scheduled on Airflow will be more robust.
- Airflow will automatically retry your job in the event of a failure.
- You can also alert other members of your team when jobs fail, while ATMO will only send an email to the job owner.
- If your job depends on other datasets, you can identify these dependencies in Airflow. This is useful if an upstream job fails.
ATMO
To schedule a job on ATMO, take a look at the load_and_run notebook. This notebook clones and installs the python_mozetl package. You can then run your job from the notebook.
Airflow
To schedule a job on Airflow, you'll need to add a new Operator to the DAGs and provide a shell script for running your job. Take a look at this example shell script. and this example Operator for templates.
Early Stage ETL Jobs
We usually require tests before accepting new ETL jobs. If you're still prototyping your job, but you'd like to move your code out of a Jupyter notebook take a look at cookiecutter-python-etl.
This tool will initialize a new repository with all of the necessary boilerplate for testing and packaging. In fact, this project was created with cookiecutter-python-etl.