One and a half years ago we decided that it was time to get on Python 3. We’ve talked about it for a long time but now it was time! The journey is now complete, we have switched the last parts of production to Python 3.
- The entire code base is ~240k lines, excluding blank lines and comments
- It’s a web based system with a lot of batch jobs. There’s only one production installation.
- The code base is ~15 years old
- It’s a Django app, but parts predate the release of Django
Some basic statistics on the Python 3 changes based on a very rough filtering of git history:
- 275 commits
- 4080 added lines
- 3432 deleted lines
I found 109 jira issues related to this project.
Py2 → six → py3
Our philosophy was always to go py2 ￫py2/py3 ￫ py3 because we just could not realistically do a big bang in production, an intuition that was proven right in surprising ways. This meant that 2to3 was a non starter which I think is probably common. We tried a while to use 2to3 to detect Python 3 compatibility issues but quickly found that untenable too. Basically it suggests changes that will break your code in Python 2. No good.
The conclusion was to use six, which is a library to make it easy to build a codebase that is valid in both in Python 2 and 3.
The most obvious first step was to update old dependencies. This work started immediately. More on that later.
Python-modernize was the tool we chose to handle the transition. It’s a tool to automatically convert from a py2 codebase to a six-compatible code base. We first introduced a test as part of CI to check that new code was py3 ready based on modernize. The biggest effect of this was to make those who still used py2 idioms aware of the new way to do things, but it obviously didn’t help much in moving the existing 240k lines to six. We all had bad habits of using some old syntax so this was a pedagogical win, even it made little difference counting lines of code. It was also used for our experimental branch:
I started a branch called simply “python3” where I did the following:
- Ran “python-modernize -n -w” on the entire code base. This modifies the code in place. I often did this step and started fixing things without first committing. This was always a mistake that I regretted, more than once forcing me to revert the entire thing and start over. It’s better to commit this stage even if it’s badly broken. Separating things done by a machine vs things done by a human is the important part here.
- Move all imports for dependencies we still hadn’t fixed for py3 into the function bodies that used them.
The idea here is to “run ahead”, i.e. to see what problems we would get if we didn’t have out of date dependencies. This branch allowed me to very quickly start the application in a super broken state and get at least some unit tests to run. The diff for this branch was of course huge, but I used this to find nice low hanging fruit that I could apply in fairly large chunks. I used the excellent GitUp to split up, combine and move around commits. When a commit looked good I cherry picked it to a new branch that I sent to code review.
No one else could work on this branch because it was constantly being rebased, force pushed and generally abused but it did move us along without having to wait for all dependencies to be updated. I highly recommend this approach!
We added pre commit hooks so if you edited a file you got nagged to fix all the python3 changes modernize suggested.
Hand rolled static analysis for
quote_plus: There are some subtleties when dealing with quote_plus and six. We ended up creating our own wrapper and statically enforcing that the code used this wrapper and not the one from the standard lib or the one from six. We also statically checked that you never sent in bytes to quote_plus.
We fixed all python3 issues per django app and enforced this with a whitelist in the CI environment so you couldn’t break an app that was once fixed.
Dependencies were the hardest part for us. We had a lot of them so it took a lot of time, but there were two stand outs:
- splunk-lib. We have a dependency to splunk and they are to this day ignoring all their very angry customers who are begging/screaming/asking them to fix their client library for py3. One person on our team finally just did it himself. Splunk has really handled this badly and even locked the issue for comments! This is unacceptable.
- Cassandra. We use this database for a lot of things across the product, but we used an older driver which used an older API model. This was a huge part of the py3 migration for us because we had to basically rewrite all this code piece by piece.
We have ~65% code coverage on our tests: unit, integration, and UI combined. We did write more tests but the overall number didn’t change much, not surprising when one considers moving coverage from 65% to 66% means writing tests that cover 2000 lines of code.
We had to skip the tests that required Cassandra while we fixed this dependency. I invented a funny little hack to make that work and wrote about that separately.
Some notes on code changes that either weren’t covered well by documentation on how to do a py2 to six transition (or maybe things we just missed):
We had a lot of uses of StringIO.StringIO in our code. The first instinct was to use six.StringIO but this turns out to be the wrong thing in almost all cases (but not all!). We basically had to think very carefully about every place we used StringIO and try to figure out if we should replace it with io.StringIO, io.BytesIO or six.StringIO. Making mistakes here often meant that the code looked like it was py3 ready and worked in py2 but was broken in py3.
from __future__ import unicode_literals
This is a mixed blessing. You find bugs by adding this to a lot of files, but it also introduces bugs in py2 sometimes. It also gets very annoying when logs suddenly write u in front of strings in weird places. Overall not the clear win I was expecting it to be.
This was largely what you’d expect. One surprise to me was the places where you needed str in py2 and py3. If you use the unicode_literals future import some strings need to go from
The implementation of six.moves is a very strange hack so it doesn’t behave like the normal python module it pretends to be. I also disagree with their choice not to include
mock in six.moves. We had to add it outselves with their API which was surprisingly difficult to get to work, and it required us to change
from mock import patch to
from six.moves import mock which also meant that
patch now becomes
CSV parsing is different
If you use the csv module you need to look at csv342. This should be a part of six in my opinion. That it’s not there means you aren’t made aware that there’s a problem. We got away with not using csv342 in many places though, so your milage may vary.
Roll out sequence
First we started with the tests:
- Run unit tests in CI
- Run integration and UI tests (excluding Cassandra) in CI
- Run Cassandra tests in CI (this was much later than the previous step!)
Next it was time to move over the product itself. We built the ability to switch one batch machine at a time to py3 and crucially to switch it back. This was very important since things did break in production when on py3. That’s mostly fine for us since we can just requeue the jobs that broke, but we can’t break too much or anything that is actually critical obviously. Because we use Sentry to collect crash logs it was very easy to review all the problems we hit when turning on py3 and when we had fixed them all, we turned on py3 again until we got a few issues, rinse repeat.
We have these environments:
- Devtest: used internally by dev so mostly this is just used to test database migrations. This environment is very lightly used so problems aren’t found here very often.
- IAT (Internal Acceptance Testing): used to validate changes and perform regression testing before we roll them out to production.
- UAT (User Acceptance Testing): a test environment that customers can access. Used for changes where we need to prepare customer systems or give customers the ability to see the changes before they go live. This environment gets database migrations a few days before they go live.
We rolled out Python 3 to these environments in this order:
- Devtest environment
- IAT environment for short periods
- IAT environment permanently
- One production batch machine for short periods
- One production batch machine on during business hours
- Production SFTP
- Half of all production batch machines
- Production batch
- Production web (after a long manual testing run of the test environment)
- Production load machines. This is a special subset of the batch processing that do the most CPU and memory heavy part of our product.
The load machines exposed configurations for customer data that was incompatible with Python 3 so we had to implement warnings for these cases in Python 2 and make sure we had fixed them all before turning on Python 3 again. This took a few days because we get customer data once a day, so every time there was even one warning we had to wait one more day.
Surprises in production
'ß'.upper()in p2 is
'SS'in py3. This caused a crash in production when the last piece of the product moved to py3!
- Sorting/comparing objects of different types is valid py2 but hides loads of bugs. We got some nasty surprises because this behavior leaked through the stack in some non-obvious ways. Especially that
Noneexisted in some lists that were being sorted. Overall this was a win since we found quite a few bugs.
Noneis sorted first in lists in py2, which might be surprising (you might expect it to be sorted next to zero!), but this was often the behavior we actually wanted. Now we just have to handle this ourselves.
'asd'in Python 2, but
"b'asd'"in Python 3. Almost any other behavior in Python 3 would have been better here: hex output (more obviously different), the old behavior (existing code works), or throwing an exception (would have been best!).
int('1_0')is 10 in py3, but invalid in py2. This even bit us before we switched to py3 because the mismatch caused another team that used py3 before us to send us integer literals that we thought were invalid, but they thought were valid. I personally think this decision was a mistake: very strict parsing is a better default. I fear this will continue to bite us in subtle ways for years to come.
Ultimately we feel that we really had no choice in the matter: Python 2 maintenance will stop at some point, and our dependencies are going py3 only, most notably Django. But we did want to do this transition anyway because we are often bitten by bytes/unicode issues and Python 3 just fixes lots of small annoyances in Python 2. The switch has already found some real bugs/misconfigurations that we’ve had in production for years. We also look forward to using f-strings and ordered dicts everywhere.