[pytest-dev] another "i have had it moment" and a stern request ro review our interaction with backward compatibility

[Sylvain MARIE]

Hi all

I could not agree more with Oliver - I discovered pytest two years ago and it was really a "wow' moment - I quickly moved away from NoseTest :)

Parameters and Fixtures are to me THE features that makes it so powerful and easy to use, and that enable anyone to invent new usage patterns that others did not even think about before. The issue for new users especially the ones coming from traditional unit testing (e.g. junit) is that it is quite difficult to understand how powerful they are at first glance.

It is however very easy to "pre-bundle" usage patterns with more user-friendly names. That's what I did with pytest-cases and pytest-steps (with the help of the great 'decorator' lib to manipulate the signatures a bit). This de-coupling between a very compact and generic core (with the smallest possible amount of API and functionality to ensure maintainability), and multiple, composable, "usage-oriented" plugins, seems to me the best reasonable long-term choice. Of course that does not prevent doing a bit of cleaning and design improvement in the internals, to improve its compactness.

Best;

--
Sylvain 


De : pytest-dev [mailto:pytest-dev-bounces+sylvain.marie=se.com at python.org] De la part de Oliver Bestwalter
Envoyé : mercredi 7 novembre 2018 22:34
À : Bruno Oliveira <nicoddemus at gmail.com>
Cc : pytest-dev at python.org
Objet : Re: [pytest-dev] another "i have had it moment" and a stern request ro review our interaction with backward compatibility

[External email: Use caution with links and attachments]
________________________________________
 
Hi Ronny,

I'd like to add my 2 cents from the perspective of an enthusiastic pytest user, who dabbles in a lot of projects using pytest and who occasionally teaches it:

I use pytest since 2010 and can still remember a time when I was able to trigger the dreaded <INTERNALERROR> quite regularly and it was often really hard for me to figure out what went wrong, leading to arcane workarounds in my test suites. 

Since I am a bit involved in the project and pay more attention, I remember 

* one incident where I triggered an INTERNAL error during normal usage in a fresh release and the problem was easy to spot and fix (https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fpytest-dev%2Fpytest%2Fissues%2F2788&data=02%7C01%7Csylvain.marie%40se.com%7C59e7ab289a2a4698bc9208d644f8fa26%7C6e51e1adc54b4b39b5980ffe9ae68fef%7C0%7C0%7C636772233479207352&sdata=%2BuNmxKZxmKNgLJ9eWFPvzL1C2fZpES9MGxYEiDp11ak%3D&reserved=0).
* one breaking change that was also easy to fix: the change in the logging behaviour and that was also perfectly o.k. because bottom line is that the logging behaviour overall now is far better because of this
* one hard to reproduce race condition due to the introduction of tmp_path (thanks for that btw), which also was fixed faster than I was even able to find some time to have a closer look (https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fpytest-dev%2Fpytest%2Fissues%2F4181&data=02%7C01%7Csylvain.marie%40se.com%7C59e7ab289a2a4698bc9208d644f8fa26%7C6e51e1adc54b4b39b5980ffe9ae68fef%7C0%7C0%7C636772233479217361&sdata=9wKOnAy1azrWVCvHxHWOjN4SETPmZD3XGvQk%2BpbOXhs%3D&reserved=0).

All in all pretty painless and easy to work around and usually fixed really quickly.

In my experience the subset of functionality/plugins that by far the most test suites use are rock solid, have a good API and hardly ever break. Also: the quality of the "user experience" and the documentation has improved vastly in the last few years.

Just today I had the chance to pair program some tests to introduce a colleague to pytest. Seeing the expression of delight in their face, when they realized how easy it is to get started and when they started to comprehend the power of fixtures and parametrization is priceless. 

Long story short: you folks are amazing and pytest is something to be extremely proud of despite all the pain that is always involved when trying to pay off technical debt without breaking the word.

The release automation that is in place now makes more frequent releases easier and from my own experience I agree with Brunos suggestion to make even more frequent releases that contain smaller changesets and also to not shy away from more frequent major releases, if that makes the transition easier.

Thanks that you care and thanks for all your work
Oliver

On Wed, 7 Nov 2018 at 20:56 Bruno Oliveira <mailto:nicoddemus at gmail.com> wrote:
Hi Ronny,
On Wed, Nov 7, 2018 at 7:12 AM Ronny Pfannschmidt <mailto:opensource at ronnypfannschmidt.de> wrote:

...


This is an accumulative cost in many ways - in particular for a project
driven primarily by volunteers - every hour we sacrafice to this altar
of technical debt is an hour lost for improving the project and/or
working on actual features.

I feel your frustration. I agree that the pytest codebase needs some refactorings in order for it to be maintainable and allow us to move forward.

From my perspective this issue directly grows out of driving a large
part of pytests development by examples and tests, but precluding
stepping back and doing actual design.

The history of marks makes a perfect example of such an horror storry.
Each iteration adding a new minimal element while adding a huge factor
to the structural error as it grew.

What I think often happens is that we can't foresee that a minimal increment might lead to a large technical debt in the future. Once we put something in the open, we try very hard to not change that behavior (even if considered incorrect now), which is the main point of your email.
 
I really want to hammer down there that "minimal" is very different
from "minimal viable" - leaving design uncontested for too long is an
ramp up for unsustainable technical debt.

Its aslo critical to note that for **me** there is a inherent
dishonesty in trying to stay backward compatible and not putting
forward a design and development process that deeply supports it -
right now we can observe a state of fragility from pytest where it
feels like every feature release triggers some regression - while at
the same time we keep and keep shipping features that are structurally
and fundamentally broken since years (yield tests, config init, ...).

I agree about the fact that every feature release we end up breaking something unintentionally. Of course that sometimes will happen, but I feel that happens more often in the murky areas of the code which have grown organically over the years. 

This setup, which gives me the impression it is "designed" to make the
project fail its user (aka its broken to begin with and now it will
break some more), is a massive emotional drain. It painfully twists
around the sense of responsibility simply due to the perceived inherent
base-level of failure - and **I** want to do and feel better.

However with the current flow of releases and with our backward
compatibility policies in place i am under a very general impression
that i cant really do anything about it, and even if i try it would
drag on for years to generate even basic positive results - this
impression is a killer to my motivation - as it **feels** like it would
take years to even get a initial iteration out to the users - and the
process of fixing the issues in general would drag on over not only
years - but decades.

A timeflow like that would commpletely disconnect me from perceptions
of reward or archivement  - in turn making it even less likely to
succeed or even start.

About our backward compatibility policies, we don't have any minimal time restriction for major releases, we only state that we will issue deprecation warnings for two feature releases before actually changing/removing a feature.

Having said that, I see that it is possible for us to have major releases more often (a few per year even). As we have learned already, frequent releases which cause few incompatibilities are preferred over infrequent releases which cause a lot of incompatibilities, as it will affect less users and makes it easy to pin point problems.

My gut feeling is that those backward incompatible releases won't be that bad in the end: we only change little used features or porting the code to the new way of doing things is easy to apply to existing code. Of course some friction always happen.

Dragging deprecated features over many releases, specially if they get in the way of new implementations, is a certain way to needlessly increase the burden of us maintainers, as you point out.
 
This is a open source project that volunteer driven - fixing deeper
issues should connect to feeling an archivement - but with out current
setup what seems to be in for **me** is more like dread, fear and
depression - a general defeat.

Thats not something i want to allow anymore. 

**I** want to feel good about my work on pytest.
**I** want to feel the archivements i have on pytest.

and because of that

**I** have to make sure the project can support that.

So **I** want to invite all of you,


Lets make pytest a project again where

* we claim suport for backward compatibility and our actions and
results show for it
* we can enjoy the act of adding new features and enhancing the
ecosystem
* we can have better and quicker feedback, not just from the fabulous
people that work on it - but also the users.


**I** strongly beleive that this is doable, but it will require cutting
some bad ends.

We need to enhance layering, it doesnt have to be perfect but we do
have to be able to keep things appart sanely and we need to have
objects in valid states per default (invalid states should only be
possible deliberately, not accidentially)

We need to talk/write about design more - the difference between
minimal and viable after all only shows itself in the heat and fruction
of different perspectives and oppinions clashing.

We need a conversation and interaction with our advanced users, so they
dont end up on old and dead pytest versions (like pypy for example).

We need a pervasive chain reason to bring trough the period of
papercuts where we do shift around layering so the ecosystem can keep
up (saving the structural integrity of pytest at the expense of
destroying the ecosystem would be a dissaster).

I agree with the points above Ronny, let's make it happen. :)

Cheers,
Bruno.
_______________________________________________
pytest-dev mailing list
mailto:pytest-dev at python.org
https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fmail.python.org%2Fmailman%2Flistinfo%2Fpytest-dev&data=02%7C01%7Csylvain.marie%40se.com%7C59e7ab289a2a4698bc9208d644f8fa26%7C6e51e1adc54b4b39b5980ffe9ae68fef%7C0%7C0%7C636772233479217361&sdata=dHBvhBLA60OOzRjx3TI4bHZFtLxmdhwactGyJehyWe0%3D&reserved=0

______________________________________________________________________
This email has been scanned by the Symantec Email Security.cloud service.
______________________________________________________________________