Ideally, of course, everything should be hidden behind leak-tight abstractions.
Unfortunately, that is rarely the case.
I recently had a developer reuse a function that generates IDs for some other type of ID.
That function that generates IDs happened to use OS entropy pool directly. It was used before, once every startup of the application so it was deemed OK.
But now with high demand for IDs that is no longer acceptable.
I noticed it because I followed the code to see how the function actually generates IDs. That would not happen in web based tool because there simply isn't convenient way to follow the references from the new code.
That is just a simple example, I get things like that every other day.
And I often get frustrated because I see you call foo(), but I need to see how foo works with the given arguments.
Though to be honest, most of the time I catch my self playing grammar checker, if the style isn't violated and you didn't make any of the dozens of checklist items you pass (I work in C++, so I often catch people writing c++98 code when there is a c++14 construct that is better)
I've seen more bugs than I like OK'd because the single line change looked innocuous in the web UI review tool.
Tools like Pull Panda are worse still, for game-ification of PRs, with leader boards for most PR reviews and fastest reviews. :(
Just as rewarding developers for lines of code written, rewarding them to mark off as many lines of PR as quickly as possible is just so totally moronic I can't begin to comprehend who and for what reason would think that may be a good idea.
Did the author manually modify 30k LoC or is this result of some automated process?
If it is for example a refactoring run from IDE (which is completely fine in itself), then the better way to handle it is to separate that single huge change into its own commit that explains how the change was generated.
For example, I insist that there are three categories of changes: reformats, refactors and functional changes and that commits can only contain single type of change and that it must be immediately apparent what type of change it is. So if you see 2k LoC reformat you skim over it and don't expect functional changes to be hiding.
On the other hand if the code is actual manual modifications it means somebody has spent a huge amount of work on it and it is not unreasonable that Code Review will take proportionally huge amount of work.
My longest Code Review took me a week and was for much smaller change in LoC than that.
There are probably better ways to handle a change like that. My preferred way is to pair up and assume that if two people actually worked on it, it is as good as a code review.
And for every one you notice, you miss three. Because the system is too complex. This is why we need to program defensively.
It was used before, once every startup of the application so it was deemed OK.
And that's where someone went wrong. There should have been a test, or runtime failure, or even better (but more difficult) a compilation failure if someone misuses something in that way.
> And that's where someone went wrong. There should have been a test, or runtime failure, or even better (but more difficult) a compilation failure if someone misuses something in that way.
I don't subscribe to the idea of creating so much additional code for a relatively simple constraint.
Code should be dense. Dense means a lot of important constraints implemented in small amount of code.
Of course this is not yet guarantee that the code is readable and easy to maintain, but as a general rule, a codebase that has 5 times the amount of code to implement same functionality will more likely than not be less maintainable. That is because it is just more difficult to work with that much code.
> There should have been (...) or even better (but more difficult) a compilation failure if someone misuses (...)
In this particular case what I suggested was to move the original function with generic name from a generic utility package to a specific module package with specific name.
See? Compilation failure, not difficult at all.
Code should be dense. Dense means a lot of important constraints implemented in small amount of code.
Sorry, but I can't believe how wrong this is. "Dense" is a really poor metric for code quality. "Simple" is a much better metric. And one aspect of "simple" is "not interwoven".
So the question here is, could the new code I'm advocating be implemented in a way that is simple and not interwoven? That is, not adding complexity to any of the rest of the code base. It is nothing more than a safer leaf.
And the answer is, in this case, yes. This does not increase the complexity of the code base. The amount of code added is commensurate with the need to make sure it's not a foot gun.
Clever naming, or visibility restriction, is a poor solution. You may prevent the function from being called explicitly more than once. But you don't prevent future refactors from leading to the function that calls that thing being called more than once.
The crucial bit of humility that is necessary is to realize that there's no way you can foresee how these problems will come up in the future. The only thing you can do is be defensive.
I agree wholeheartedly and I wish more developers felt this way.
I hate opening new projects and seeing the potential for null reference exceptions everywhere when you could instead just check the constraints directly and give a more meaningful response.
My review flow now is to check out the branch under review, read it in IDE with git annotate on, and do all my actual review work there so I can see the full context. Then add line-by-line comments on Github where I can. Comments relating to older parts of the system that should change in tandem with the current PR have to go in as "general" comments, with file/line number included manually.
You can't see that the code they changed also lives in another function that they missed, or than in fact the whole thing should be factored out into a function and fixed once.