Table of Contents generated with DocToc

Backtracking issues context

The pip tool we are using in Airflow has a long standing problem with backtracking kicking in sometimes randomly. This is something we have very little control over, because the moment when backtracking kicks in depends on how many versions of conflicting packages are released in PyPI and can change completely without any change to Airflow. We have constraint mechanism to protect our users installing Released versions and developers making “regular” PRs, however in main builds and in PRs that change setup.py, this backtracking might lead to extremely long (many hours) image builds and eventually cancelling the image build jobs in CI.

An example of such issue is described here.

Unfortunately the problem is that in such cases, it is not possible to figure out what caused the problem from pip output (state as of pip 22.1.2).

There are a number of issues in pip that describe the issue, and some backtracking reasons have been already tracked down and fixed by pip maintainers, but this is a difficult problem to solve and it is likely it is going to be with us for a while. Some other opened issues:

Some issues here

Also, the PR that might help in a relatively short time is here:

What can we do about it?

Until pip gets an improved way of avoiding or detecting and showing the root cause of the conflict there is unfortunately only a trial-and-error method. We need to track down which dependencies have been changed recently and try to pinpoint the root cause of the backtracking. This is not easy because sometimes the root cause of the problem is not at all obvious and relies on some hidden limitations and design choices of the pip backtracking algorithm, which produce a non-obvious problems.

The issue is a good example of that.

How to detect it

Whenever such situation occurs, The build image workflow of ours from the main repository will start to get cancelled on timeout.

https://github.com/apache/airflow/actions/workflows/build-images.yml?query=event%3Apush+branch%3Amain

You might see various errors:

#32 3883.7 INFO: pip is looking at multiple versions of NNNN to determine which version is compatible with other requirements. This could take a while.
Error: The operation was canceled.

Or you might see errors about various pip installation problems:

#32 664.1 Collecting Flask-OpenID<2,>=1.2.5
  #32 664.2   Downloading Flask-OpenID-1.2.5.tar.gz (43 kB)
  #32 664.2      ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 43.4/43.4 KB 181.6 MB/s eta 0:00:00
  #32 664.2   Preparing metadata (setup.py): started
  #32 664.3   Preparing metadata (setup.py): finished with status 'error'
  #32 664.3   error: subprocess-exited-with-error
  #32 664.3
  #32 664.3   × python setup.py egg_info did not run successfully.
  #32 664.3   │ exit code: 1
  #32 664.3   ╰─> [1 lines of output]
  #32 664.3       error in Flask-OpenID setup command: use_2to3 is invalid.
  #32 664.3       [end of output]

But important thing is that suddenly the main build images stop working without any action from our side.

How to track the root cause

Whenever a conditions occurs which leads to cancelling CI image build, there are steps run in CI. “Candidates for pip resolver backtrack triggers”. Those steps will list the packages that have been updated since the last successful main build in the last day.

You need to find the first such failing job from the list.

And you should find the list of packages with information which versions and when were updated. You will also find a command that you can use for tracking the package, similar to:

pip install ".[devel_all]" --upgrade --upgrade-strategy eager \
        "dill<0.3.3" "certifi<2021.0.0" "google-ads<14.0.1"' "package1==N.N.N" "package2==N.N.N" ...

Example:

Image here

The candidate packages are the ones with ==. The command attempts to install the suspicious packages in the version that was correctly installed before and is stored in the current constraints.

The process of tracking down which package is the “root cause” looks as follows:

  1. Checkout the latest main of Airflow
  2. Build the latest image (with constraints): breeze ci-image build --python 3.7
  3. Enter breeze breeze
  4. Attempt to run the pip install command that was printed in the “Candidates ...” step
  5. The command should succeed (the candidates are pinned to the “working” version)
  6. Attempt to run pip install ".[devel_all]" --upgrade --upgrade-strategy eager "dill<0.3.3" "certifi<2021.0.0" "google-ads<14.0.1"
  7. This one should cause backtracking
  8. Use the original command from “Candidates ...” job the candidates and remove the candidates one-by-one from the command and re-run until you get into backtracking
  9. Even if you enter into backtracking with one candidate - do not stop - bring it back and remove other candidates one-by-one to make sure that the candidate you found is the “REAL” cause. There should usually be only one candidate left and removing this one candidate from the list should cause backtracking.

Example:

This is the original candidate list from https://github.com/pypa/pip/issues/10924. The list was long because when we tracked this one we did not have the “first failing” build and our list of candidates got a bit long after 3 days of failing build.

pip install ".[devel_all]" --upgrade --upgrade-strategy eager "dill<0.3.3" "certifi<2021.0.0" "google-ads<14.0.1" \
   "APScheduler==3.6.3" "boto3==1.21.4" "botocore==1.24.4" "connexion==2.11.2" "github3.py==3.0.0" \
   "google-api-python-client==1.12.10" "google-auth-oauthlib==0.4.6" "google-cloud-automl==2.6.0" \
   "google-cloud-dataproc==3.2.0" "google-cloud-os-login==2.5.1" \
   "google-cloud-redis==2.5.1"  "google-cloud-translate==1.7.0"

This command works correctly without backtracking. Then run the “bare” upgrade command:

pip install ".[devel_all]" --upgrade --upgrade-strategy eager "dill<0.3.3" "certifi<2021.0.0" "google-ads<14.0.1"

This one should enter into backtracking.

After removing all the candidates one-by-one, what is left is:

pip install ".[devel_all]" --upgrade --upgrade-strategy eager "dill<0.3.3" "certifi<2021.0.0" "google-ads<14.0.1" \
    "github3.py==3.0.0"

This command also succeeds.

However removing “github3.py==3.0.0” triggers backtracking.

Now we know what triggers backtracking. You can download and unpack the guilty package from PyPI - in this case github3.py files. By inspecting setup.py and setup.cfg and comparing it with pipdeptree | less output, you should figure out what causes the conflict. In this case github3.py had PyJWT>=2.3.0 and from our pipdeptree PyJWT<2.0 was the limitation of Flask App Builder. In this case the limitation of pip algorithm caused that it was not able to determine that github3.py==3.0.0 is a good candidate. In this case it was misleading because when github3.py did not have a 3.1.2 release, and 3.1.0 and 3.1.1 had the same PyJWT limitation, pip was able to find the right resolution without backtracking.

Finding candidates manually

You can also find the candidates manually. This is especially when you are not sure when the build broke, and you need to extend the time or when you need to run it for another branch. You need to install the breeze:

  • pipx install -e ./dev/breeze if you use pipx install.

Then you can run breeze ci find-newer-dependencies with optional flags. For example if you know that the build was likely broken on a given date and time (in your timezone) and you want to check python 3.8 (because this is the only failing build) you can run:

breeze ci find-newer-dependencies --updated-on-or-after '2022-02-22 10:30:00' --timezone 'CET' --python 3.8

The full list of options for find-newer-dependencies can be seen here

breeze ci find-newer-dependencies