[zzzeek]

Bad commits are the absolute most difficult thing to reverse when they get released and the userbase codes to them. When I don't pay super close attention to what's being contributed, it creates three times as much work after I let it through.

[bartread]

Superficially, that sounds more like a release process issue than a bad commit issue.

[nextlevelwizard]

How often is commit bad enough to require revert, yet made it through testing and got released?

[btown]

Code style, inconsistent APIs, a feature that's barely tested by its own contributor, introducing dependencies on innards that were planned to be removed by an ongoing refactoring project, etc. So many ways contributions can go wrong.

[diggan]

All those things should and hopefully would have been caught when the PR was made, not after the commits made it into master.

[btown]

The entire context of this thread is that adding committers can introduce risk in the quality of accepted contributions. Building a trusted network of gatekeepers is hard.

[nextlevelwizard]

Just because stuff gets to master doesn't mean they get to the release.

[snakeanus]

> Code style

How is that an issue in this day and age? Just run astyle or indent.

[gopalv]

> Just run astyle or indent.

"Style" is more than indentation - my favourite example is something I worked with very closely, but always needed a cheat-sheet on my desk - the PHP arrays api.

Just look at the docs for array_search and array_replace.

This is much worse in managed memory languages - do you allocate memory, who frees it, who can use a buffer in zero-copy mode, who can't ... whether the functions are always re-entrant.

All that said - any originating coder who burns out due to support issues, killing the community is measurably worse off than someone who has thousands of users who demand that you support such "legacy" APIs for another six months.

[zzzeek]

I'm talking here about a developer API that is extremely flexible and extensible. Even one-liner contributions can have bad effects, and virtually all contributions I receive "work" at a superficial level and if they provide tests the tests usually pass. "made it through testing" is a low bar, the bigger implications of a patch need to be considered.

Examples of PRs that "work" exactly as intended yet are the wrong way to solve the problem:

https://github.com/zzzeek/sqlalchemy/pull/381

https://github.com/zzzeek/sqlalchemy/pull/382

Here's a large PR:

https://github.com/zzzeek/sqlalchemy/pull/365

that after lots and lots of back and forth I finally merged with lots of edits from me at: https://github.com/zzzeek/sqlalchemy/commit/7d3da6f850dca54b...

...aannnnndddd - after all that work and testing, it was wrong! Fortunately someone caught the mistake within the beta release:

https://bitbucket.org/zzzeek/sqlalchemy/issues/4072/mysqldml...

if I had released that fully and the whole world coded to "mysql.insert.values.bar == 5", that's broken API since it breaks the values() method. Backwards-incompatible fix required.

[ianamartin]

Do you think that's generally true? Or is it a facet of your experience running three of the most important Python packages around?