Wait, does urlib not use semvar? Don't remove APIs on minor releases people. A major release doesn't have to be a problem or a major redesign, you can do major release 400 for all I care, just don't break things on minor releases.
Lots of things not using semvar that I always just assumed did.
As an example, I always knew urllib3 as one of the foundational packages that Requests uses. And I was curious, what versions of urllib3 does Requests pull in?
That is exactly the kind of dependency specification I would expect to see for a package that is using semver: The current version of urllib3 is 2.x, so with semver, you set up your dependencies to avoid the next major-version number (in this case, 3).
So, it seems to me that even the Requests folks assumed urllib3 was using semver.
I would almost expect the 3 in urllib3 to be the major version and if something needed to break it would become urllib4. Which, I know, is terribly naive of me. But that is how psycopg does it.
That was how psycopg2 did it, but now the package is psycopg (again) version 3, as it should be. Python package management has come a long way since psycopg 1 was created.
urllib2/3’s etymology is different: urllib2’s name comes from urllib in the standard library.
Because everyone is afraid of a v4, after the 2-3 debacle. And there are things which need to be culled every once in a while to keep the stdlib fresh.
Python is culling stuff all the time, but that doesn't warrant a major version jump.
You are probably right about Pythons careful approach of when to ship v4, but for the wrong reasons. Python 3 was necessary not for the removal of functions, but because of the syntax changes e.g., moving print from a statement to a method.
Semver works fine for SDL and has worked fine since the start of the century, despite the library's complexity and scale. A few simple rules can go a long way if you're disciplined about enforcing them.
Making you distrust updates is absolutely the correct versioning method. Pin your versions in software you care about and establish a maintenance schedule. Trusting that people don't break things unintentionally all the time is extremely naive.
It was dumb and user-hostile to remove an interface for no good reason that just makes it more work for people to update, but everyone not pinning versions needs to acknowledge that they're choosing to live dangerously.
In practice, semver is very helpful. Its major benefit is allowing packages to declare compatibility with versions of their own dependencies that don’t exist yet. (Distrusting updates and pinning versions is important and correct, but it’s not a “versioning method” that stands in contrast to semver or anything. That’s what lockfiles are for.) The pre-semver Python package ecosystem is a good example of what happens without it: fresh installs of packages break all the time because they have open-ended or overly permissive upper bounds on their dependencies. If they were to specify exact preexisting upper bounds, they’d slow down bugfixes (and in Python, where you can only have one version of a package in a given environment, new features) and add maintenance busywork; I’m not aware of any packages that choose this option in practice.
> You have released version 1.0.0 of something. Then you add a feature and fix a bug unrelated to that feature. Are you at version 1.1.0 or 1.1.1? Well, it depends on the order you added your changes, doesn't it? If you fixed the bug first you'll go from 1.0.0 to 1.0.1 to 1.1.0, and if you add the feature first you'll go from 1.0.0 to 1.1.0 to 1.1.1. And if that difference doesn't matter, then the last digit doesn't matter.
It depends on the order you released your changes, yes. If you have the option, the most useful order is to release 1.0.1 with the bugfix and 1.1.0 with both changes, but you can also choose to release 1.1.0 without the bugfix (why intentionally release a buggy version?) and then 1.1.1 (with or without 1.0.1), or just 1.1.0 with both changes. You’re correct that the starting point of the patch version within a particular minor version doesn’t matter – you could pick 0, 1, or 31415. You can also increment it by whatever you want in practice. All this flexibility is a total non-problem (let alone a problem with the versioning scheme, considering it’s flexibility that comes from which releases you even choose to cut – semver just makes the relationship between them clear), and doesn’t indicate that the patch field is meaningless in general. (Obviously, you should start at 0 and increment by 1, since that’s boring and normal.)
Sure, it’s impossible to classify breaking changes and new features with perfect precision, and maintainers can make mistakes, but semver is pretty clearly a net positive. (It takes almost no effort and has no superior competitors, so it would be hard for it not to be.)
They article does validly point out that deprecation warnings don't work. Turns out in this day and age that the only thing you can reliably inform about changes is the package manager and its dependency solver, and pip requires semver or similar for that.
Fixing deprecations is unfortunately the lowest prio of any kind of work for majority of the projects. Part of the problem is probably lack of pressure to do so it if the timeline is unclear. What if this is actually never removed? Why going through the pain?
IMO telling "we deprecate now and let's see when we remove it" is counterproductive.
A better way: deprecate now and tell "in 12 (or 24?) months this WILL be removed".
After 12/24 months, cut a new semver-major release. People notice the semver-major through the dependency management tools at some point, an maybe they have a look at changelog.
If they don't, at some point they may want to use a new feature, and finally be incentivised to update.
If there's no incentive other than "do the right thing", it never gets done.
Having said that, I think LLMs are really going to help with chores like this, if e.g. deprecations and migration steps are well documented.
Alternative option: create a codemod CLI that fixes deprecations for the users, doing the right thing automatically. If migration is painless and quick, it's more likely people will do it.
> Fixing deprecations is unfortunately the lowest prio of any kind of work for majority of the projects.
... and the right answer to that is to make it entirely their problem.
> Part of the problem is probably lack of pressure to do so it if the timeline is unclear. What if this is actually never removed?
In this case, the warnings said exactly what release would remove the API. Didn't help.
> Why going through the pain?
Because you're not a feckless irresponsible idiot? I don't think it's an accident that the projects they said didn't react were an overcomplicated and ill-designed management layer for an overcomplicated and ill-designed container system, a move-fast-and-break-things techbro company, and what looks to be a consolation project for the not-too-bright.
You probably get an extra measure of that if you're operating in the Python ecosystem, which is culturally all about half-assed, 80-percent-right-we-hope approaches.
The right answer is to remove it when you say you're going to remove it, and let them pick up the pieces.
It also helps if you design your API right to begin with, of course. But this is Python we're talking about again.
> After 12/24 months, cut a new semver-major release. People notice the semver-major through the dependency management tools at some point, an maybe they have a look at changelog.
Deprecations in all forms are always a shitshow. There isn’t a particular pattern that “just works”. Anybody that tells you about one, best case scenario, it just worked for them because of their consumer/user not because of the method itself.
The best I have seen is a heavy handed in-editor strike through with warnings (assuming the code is actively being worked on) and even then it’s at best a 50/50 thing.
50% of the developers would feel that using an API with a strike through in the editor is wrong. And the other 50% will just say “I dunno, I copied it from there. What’s wrong with it??”
Deprecations via warnings don't reliably work anywhere, in general.
If you are a good developer, you'll have extensive unit test coverage and CI. You never see the unit test output (unless they fail) - so warnings go unnoticed.
If you are a bad developer, you have no idea what you are doing and you ignore all warnings unless program crashes.
You can turn warnings into errors with the `-Werror` option. I personally use that in CI runs, along with the `-X dev` option to enable additional runtime checks. Though that wont solve the author's problem, since most Python devs don't use either of those options
In PHP I don’t think there is a native way to convert E_DEPRECATED into E_ERROR, but the most common testing framework has a quick way of doing the same.
When I update python version, python packages, container image, etc for a service, I take a quick look at CI output, in addition to the all the other checks I do (like a couple basic real-world-usage end-to-end usage tests), to "smoke test" whether something not caught by outright CI failure caused some subtle problem.
So, I do often see deprecation warnings in CI output, and fix them. Am I a bad developer?
I think the mistake here is making some warnings default-hidden. The developer who cares about the user running their the app in a terminal can add a line of code to suppress them for users, and be more aware of this whole topic as a result (and have it more evident near the entrypoint of the program, for later devs to see also).
I think that making warnings error or hidden removes warnings as a useful tool.
Author here! Agreed that are different levels of "engaged" from users, which is okay. The concerning part of this finding is that even dependent users that I know to be highly engaged didn't respond to the deprecation warnings, so they're not working for even the most engaged users.
There was this one library we depended on, it was sort of in limbo during the Python 2 -> 3 migration. During that period is was maintained by this one person who'd just delete older versions when never ones became available. In one year I think we had three or four instances where our CI and unit tests just broke randomly one day, because the APIs had changed and the old version of the library had been yanked.
In hindsight it actually helped us, because in frustrations we ended up setting up our own Python package repo and started to pay more attention to our dependencies.
> If you are a good developer, you'll have extensive unit test coverage and CI. You never see the unit test output (unless they fail) - so warnings go unnoticed.
In my opinion test suites should treat any output other than the reporter saying that a test passed as a test failure. In JavaScript I usually have part of my test harness record calls to the various console methods. At the end of each test it checks to see if any calls to those methods were made, and if they were it fails the tests and logs the output. Within tests if I expect or want some code to produce a message, I wrap the invocation of that code in a helper which requires two arguments: a function to call and an expected output. If the code doesn't output a matching message, doesn't output anything, or outputs something else then the helper throws and explains what went wrong. Otherwise it just returns the result of the called function:
let result = silenceWarning(() => user.getV1ProfileId(), /getV1ProfileId has been deprecated/);
expect(result).toBe('foo');
This is dead simple code in most testing frameworks. It makes maintaining and working with the test suite becomes much easier as when something starts behaving differently it's immediately obvious rather than being hidden in a sea of noise. It makes working with dependencies easier because it forces you to acknowledge things like deprecation warnings when they get introduced and either solve them there or create an upgrade plan.
That's why you, very early on, release code that slows the API down once it has been deprecated. Every place you issue a deprecation warning, you also sleep 60. Problem solved.
Why is it that CI tools don't make warnings visible? Why are they ignored by default in the first place? Seems like that should be a rather high priority.
> Why is it that CI tools don't make warnings visible?
A developer setting up CI decides to start an ubuntu 24.04 container and run 'apt-get install npm'
This produces 3,600 lines of logging (5.4 log lines per package, 668 packages) and 22 warnings (all warnings about man page creation being skipped)
Then they decide "Nobody's going to read all that, and the large volume might bury important information. I think I'll hide console output for processes that don't fail."
It isn't that easy. If you have a new warning on upgrade you probably want to work on it "next week", but that means you need to ignore it for a bit. Or you might still want to support a really old version without the new API and so you can't fix it now.
> If you have a new warning on upgrade you probably want to work on it "next week", but that means you need to ignore it for a bit.
So you create a bug report or an issue or a story or whatever you happen to call it, and you make sure it gets tracked, and you schedule it with the rest of your work. That's not the same thing as "ignoring" it.
Why does your codebase generate hundreds of warnings, given that every time one initially appeared, you should have stamped it out (or specifically marked that one warning to be ignored)? Start with one line of code that doesn't generate a warning. Add a second line of code that doesn't generate a warning...
> Why does your codebase generate hundreds of warnings
Well, it wasn't my codebase yesterday, because I didn't work here.
Today I do. When I build, I get reports of "pkg_resources is deprecated as an API" and "Tesla T4 does not support bfloat16 compilation natively" and "warning: skip creation of /usr/share/man/man1/open.1.gz because associated file /usr/share/man/man1/xdg-open.1.gz (of link group open) doesn't exist" and "datetime.utcnow() is deprecated and scheduled for removal in a future version"
The person onboarding me tells me those warnings are because of "dependencies" and that I should ignore them.
It's rare that I work on a project I myself started. If I start working on an existing codebase, the warnings might be there already. Then what do I do?
I'm also referring to all the warnings you might get if you use an existing library. If the requirements entail that I use this library, should I just silence them all?
But I'm guessing you might be talking about more specific warnings. Yes I do fix lints specific to my new code before I commit it, but a lot of warnings might still be logged at runtime, and I may have no control over them.
If this code crashed all the time there'd be a business need to fix it and I could justify spending time on this.
But that's not what we're discussing here, we're discussing warnings that have been ignored in the past, and all of a sudden I'm supposed to take the political risk to fix them all somehow, even though there's no new crash, no new information.
I don't know how much freedom you have at your job; but I definitely can't just go to my manager and say: "I'm spending the next few weeks working on warnings nobody else cared about but that for some reason I care about".
Because most people are working at Failure/Feature factories where they might work on something and at last minute, they find out something is now warning. If they work on fixing it, the PM will screaming about time slippage and be like "I want you to work on X, not Y which can wait".
You found that out at the last minute. So then you did a release. It's no longer the last minute. Now what's your excuse for the next release?
If your management won't resource your project to the point where you can assure that the software is correct, you might want to see if you can find the free time to look for another job. You'll have to do that anyway when they either tank the company, or lay you off next time they feel they need to cut more costs.
>We ended up adding the APIs back and creating a hurried release to fix the issue.
So it was entirely possible to keep the software working with these. Why change/remove them in the first place? Is the benefit of of the new abstraction greater than the downside of requiring everyone using the software to re-write theirs?
OS software maintainers don't like maintaining legacy ugly APIs forever and want to refactor/remove legacy code to keep themselves sane and the project maintainable.
Every public API change is a cost on the user: for an extreme example, if every library I ever used renamed half its APIs every year to align with the latest ontology, then there would hardly be any point in saving my scripts, since I'd have to be constantly rewriting them all.
Of course, the reality is hardly ever as bad as that, but I'd say having to deal with trivial API changes is a reasonable basis for a user to dislike a given project or try to avoid using it. It's up to the maintainers how friendly they want to be toward existing user code, and whether they pursue mitigating options like bundling migrations into less-common bigger updates.
Yep. Or you can see it as, "This software doesn't really care about the users and their use cases. It prioritizes making things look pretty and easier on the dev side over maintaining functionality." Or in the worse but fairly common OSS case, CADT, but that doesn't seem to apply in this context.
I think this is a valid question for this specific case, but may not always be possible. That said, I think as a user I would probably prefer it if under the hood the old function called the new so they can deprecate the behavior without breaking the API. In that way you can still emit the deprecation warning while also only having one actual code path to maintain.
Wild (and I guess most of the time bad) idea: on top of the warnings, introduce a `sleep` in the deprecated functions. At every version, increase the sleep.
Has this ever been considered?
The problem with warnings is that they're not really observable: few people actually read these logs, most of the time. Making the deprecation observable means annoying the library users. The question is then: what's the smallest annoyance we can come up with, so that they still have a look?
I prefer Go's solution to this problem. Just don't deprecate stuff. And to make that possible, slow down and design stuff you will be willing to support forever.
Deprecations don't work. Don't deprecate stuff without a really really good reason. The new API being cleaner is not a good reason. There are very few good reasons.
If you deprecate something in a popular library, you're forcing millions of people to do work. Waste time that could be used for something better, possibly at a time of your choice, not theirs. It was emitting warnings for 3 years... so you think everyone should have to rewrite their software every 3 years?
Especially for something like this. Only document it in a footnote, mark it as deprecated, etc - but don't remove the alias.
Don't break stuff, unless, to quote a famous work, you think your users are scum. Do you think your users are scum? Why do you hate your users?
Open source developers have no such obligation. They’re doing things on their own time. But hey, I’m sure they’d be willing to give you a full refund for the amount you paid them to use their work
I'm very confused on the debate of semver here - the fundamental principle seems very simple, and important.
"give me all updates to my core version that's still compatible"
Semver simply puts a 'protocol' to this - define your major version and off you go.
While in practice you could go and search for each and every library you use to check when/how they do breaking versions, but semver just allows matching a single number across the board - it makes it more consistent and error proof.
Because it's not "error proof". It's not even "more error proof". Relying on updates not breaking if the maintainer doesn’t intend it means relying on software being bug-free, and it should be immediately obvious that that would be a very very stupid thing to believe. Letting software you rely on change on its own is always dangerous no matter how much someone pinky promises that you can trust them to have thought of everything. Semver numbers are determined by humans who aren't oracles and therefore make mistakes all the time about what changes will break something for someone else. It may get people to think a little bit about whether they _think_ they're breaking something, but it's never safe to rely on that judgement because there's a good chance that they're wrong. And given that it's never safe to rely on that judgment, the process of pretending is worse than worthless; the pretense is inherently dangerous, because people start to unthinkingly trust it. If you care about things working, you pin and vendor (npm left-pad) your dependencies and only update a package if you _need_ to and only after explicit evaluation.
And if after reading that you think to yourself, "But BugsJustFindMe, I have unit tests that will catch semver mistakes!" I think you need to ask yourself what your unit tests tell you about semver code that they don't also tell you about non-semver code.
The devil is in the details. It seems `getHeaders` v. `headers` is non-security, non-performance related issue. Why people should spend time fixing these?
Even if getHeaders() has security/performance concerns, the better solution is to make it an alias to the newer headers.get() in this case. Keeping the old API is a small hassle to a handful of developers but breaking existing code puts a much bigger burden on a lot more users.
Instead of reverting the updates to appease those who've been ignoring the warnings, a way forward could be to gate the old API behind an environment variable, with a name like "TEMPORARILY_REENABLE_DEPRECATED_APIS". And maybe put an expiry on the variable so it resets occasionally. Otherwise the project just becomes mired and stagnant.
A deprecation warning is not actionable for typical end users. Why don't more warnings include calls to action?
Instead of a warning that says, "The get_widget method in libfoo is deprecated and will be removed by November 30", the warning could say:
"This app uses the deprecated get_widget method. Please report this bug to the app developer. Ask them to fix this issue so you can continue using this app after November 30."
This is exactly the sort of breaking change that I really struggle to see the value of — maintaining the deprecated method seems incredibly unlikely to be a notable maintenance burden when it is literally just:
Like sure — deprecate it, which might have _some_ downstream cost, rather than having two non-deprecated ways to do the same thing, just to make it clear which one people should be using; but removing it has a much more significant cost on every downstream user, and the cost of maintenance of the old API seems like it should be almost nothing.
(I also don't hate the thought of having a `DeprecationWarning` subclass like `IndefiniteDeprecationWarning` to make it clear that there's no plan to remove the deprecated function, which can thus be ignored/be non-fatal in CI etc.)
There is value for the person maintaining this library cause they want it that way.
If you develop a useful library and give it away for free then all power to you if you want to rearrange the furniture every 6 months. I'll roll with it.
The secret trick I've used on rare occasion, but when necessary, is the "ten second rule."
Users don't notice a deprecation warning. But they might notice adding a "time.sleep(10)" immediately at the top of the function. And that gives them one last grace period to change out their software before it breaks-breaks.
A breaking change causes a full-stop to a service.
An intentional slowdown lets the service continue to operate at degraded performance.
I concur that it's less clear for debugging purposes (although any reasonable debugging infrastructure should allow you to break and see what function you're in when the program hangs; definitely not as clear as the program crashing because the called function is gone, however).
A breaking change in a dependency doesn’t cause a full-stop to a service at all. The old version continues to work. Making subtly harmful changes so that new broken versions sneak in is just a bad idea and totally unnecessary.
> A breaking change in a dependency doesn’t cause a full-stop to a service at all
From the article:
"We still received feedback from users that this removal was unexpected and was breaking dependent libraries."
I think we may be assuming different floors on service maintainer competency; with so many users pulling in dependencies across an arbitrarily-wide version window with no testing, such changes do break services.
It’s not necessary to cater to the absolute least competent end user to begin with, but inserting slowdown bugs does not even achieve that. (Note that the bit about the breaking of dependent libraries you’re quoting is still not actually a service being affected.)
Just break, then revert when anyone complains, on every single release. eventually you will get a release where nobody complains as they move off the depreciated api due to breakage annoyance.
It's worked in the past. But it does require someone at your org to care that CI times are spiking, which is not always a thing you can rely upon.
In addition: if CI is the only place the issue shows up, and never in a user interaction... Why does that software exist in the first place? In that context, the slowdown may be serving as a useful signal to the project to drop the entire dependency.
ETA: To be clear, I don't do this as a substitute for a regular deprecation cycle (clear documentation, clear language-supported warnings / annotations, clear timeline to deprecate); I do it in addition before the final yank that actually breaks end-users.
Honestly, I think that the author already found (and rejected the solution):
> I could ask for more Python developers to run with warnings enabled, but solutions in the form of “if only we could all just” are a folly.
I get where he's coming from. But the facts are: the language provides a tool to warn users, the library is using that tool, and users are choosing to turn that tool off. That's fine, but then they don't get to complain when stuff breaks without warning. It is the user's responsibility at that point, not the library maintainers'.
Wait, does urlib not use semvar? Don't remove APIs on minor releases people. A major release doesn't have to be a problem or a major redesign, you can do major release 400 for all I care, just don't break things on minor releases.
Lots of things not using semvar that I always just assumed did.
This, I think, is the crux of the matter.
As an example, I always knew urllib3 as one of the foundational packages that Requests uses. And I was curious, what versions of urllib3 does Requests pull in?
Well, according to https://github.com/psf/requests/blob/main/setup.cfg, it's this:
That is exactly the kind of dependency specification I would expect to see for a package that is using semver: The current version of urllib3 is 2.x, so with semver, you set up your dependencies to avoid the next major-version number (in this case, 3).So, it seems to me that even the Requests folks assumed urllib3 was using semver.
I would almost expect the 3 in urllib3 to be the major version and if something needed to break it would become urllib4. Which, I know, is terribly naive of me. But that is how psycopg does it.
That was how psycopg2 did it, but now the package is psycopg (again) version 3, as it should be. Python package management has come a long way since psycopg 1 was created.
urllib2/3’s etymology is different: urllib2’s name comes from urllib in the standard library.
er, *urllib3’s name comes from urllib2 in the standard library.
Python itself doesn't use semver.
Because everyone is afraid of a v4, after the 2-3 debacle. And there are things which need to be culled every once in a while to keep the stdlib fresh.
Python is culling stuff all the time, but that doesn't warrant a major version jump.
You are probably right about Pythons careful approach of when to ship v4, but for the wrong reasons. Python 3 was necessary not for the removal of functions, but because of the syntax changes e.g., moving print from a statement to a method.
semver is funny because it gives the illusion of working but does not work at all
Glory to 0ver: https://0ver.org/
Semver works fine for SDL and has worked fine since the start of the century, despite the library's complexity and scale. A few simple rules can go a long way if you're disciplined about enforcing them.
+1
The industry has a solution for the exact problem the urllib is having (semver). Urllib just actively refuses to use it.
The value of semver has always been a lie. https://news.ycombinator.com/item?id=37426532
Making you distrust updates is absolutely the correct versioning method. Pin your versions in software you care about and establish a maintenance schedule. Trusting that people don't break things unintentionally all the time is extremely naive.
It was dumb and user-hostile to remove an interface for no good reason that just makes it more work for people to update, but everyone not pinning versions needs to acknowledge that they're choosing to live dangerously.
In practice, semver is very helpful. Its major benefit is allowing packages to declare compatibility with versions of their own dependencies that don’t exist yet. (Distrusting updates and pinning versions is important and correct, but it’s not a “versioning method” that stands in contrast to semver or anything. That’s what lockfiles are for.) The pre-semver Python package ecosystem is a good example of what happens without it: fresh installs of packages break all the time because they have open-ended or overly permissive upper bounds on their dependencies. If they were to specify exact preexisting upper bounds, they’d slow down bugfixes (and in Python, where you can only have one version of a package in a given environment, new features) and add maintenance busywork; I’m not aware of any packages that choose this option in practice.
> You have released version 1.0.0 of something. Then you add a feature and fix a bug unrelated to that feature. Are you at version 1.1.0 or 1.1.1? Well, it depends on the order you added your changes, doesn't it? If you fixed the bug first you'll go from 1.0.0 to 1.0.1 to 1.1.0, and if you add the feature first you'll go from 1.0.0 to 1.1.0 to 1.1.1. And if that difference doesn't matter, then the last digit doesn't matter.
It depends on the order you released your changes, yes. If you have the option, the most useful order is to release 1.0.1 with the bugfix and 1.1.0 with both changes, but you can also choose to release 1.1.0 without the bugfix (why intentionally release a buggy version?) and then 1.1.1 (with or without 1.0.1), or just 1.1.0 with both changes. You’re correct that the starting point of the patch version within a particular minor version doesn’t matter – you could pick 0, 1, or 31415. You can also increment it by whatever you want in practice. All this flexibility is a total non-problem (let alone a problem with the versioning scheme, considering it’s flexibility that comes from which releases you even choose to cut – semver just makes the relationship between them clear), and doesn’t indicate that the patch field is meaningless in general. (Obviously, you should start at 0 and increment by 1, since that’s boring and normal.)
Sure, it’s impossible to classify breaking changes and new features with perfect precision, and maintainers can make mistakes, but semver is pretty clearly a net positive. (It takes almost no effort and has no superior competitors, so it would be hard for it not to be.)
They article does validly point out that deprecation warnings don't work. Turns out in this day and age that the only thing you can reliably inform about changes is the package manager and its dependency solver, and pip requires semver or similar for that.
Fixing deprecations is unfortunately the lowest prio of any kind of work for majority of the projects. Part of the problem is probably lack of pressure to do so it if the timeline is unclear. What if this is actually never removed? Why going through the pain?
IMO telling "we deprecate now and let's see when we remove it" is counterproductive.
A better way: deprecate now and tell "in 12 (or 24?) months this WILL be removed".
After 12/24 months, cut a new semver-major release. People notice the semver-major through the dependency management tools at some point, an maybe they have a look at changelog.
If they don't, at some point they may want to use a new feature, and finally be incentivised to update.
If there's no incentive other than "do the right thing", it never gets done.
Having said that, I think LLMs are really going to help with chores like this, if e.g. deprecations and migration steps are well documented.
Alternative option: create a codemod CLI that fixes deprecations for the users, doing the right thing automatically. If migration is painless and quick, it's more likely people will do it.
> Fixing deprecations is unfortunately the lowest prio of any kind of work for majority of the projects.
... and the right answer to that is to make it entirely their problem.
> Part of the problem is probably lack of pressure to do so it if the timeline is unclear. What if this is actually never removed?
In this case, the warnings said exactly what release would remove the API. Didn't help.
> Why going through the pain?
Because you're not a feckless irresponsible idiot? I don't think it's an accident that the projects they said didn't react were an overcomplicated and ill-designed management layer for an overcomplicated and ill-designed container system, a move-fast-and-break-things techbro company, and what looks to be a consolation project for the not-too-bright.
You probably get an extra measure of that if you're operating in the Python ecosystem, which is culturally all about half-assed, 80-percent-right-we-hope approaches.
The right answer is to remove it when you say you're going to remove it, and let them pick up the pieces.
It also helps if you design your API right to begin with, of course. But this is Python we're talking about again.
> After 12/24 months, cut a new semver-major release. People notice the semver-major through the dependency management tools at some point, an maybe they have a look at changelog.
The urllib3 package doesn't use SemVer.
Deprecations in all forms are always a shitshow. There isn’t a particular pattern that “just works”. Anybody that tells you about one, best case scenario, it just worked for them because of their consumer/user not because of the method itself.
The best I have seen is a heavy handed in-editor strike through with warnings (assuming the code is actively being worked on) and even then it’s at best a 50/50 thing.
50% of the developers would feel that using an API with a strike through in the editor is wrong. And the other 50% will just say “I dunno, I copied it from there. What’s wrong with it??”
Deprecations via warnings don't reliably work anywhere, in general.
If you are a good developer, you'll have extensive unit test coverage and CI. You never see the unit test output (unless they fail) - so warnings go unnoticed.
If you are a bad developer, you have no idea what you are doing and you ignore all warnings unless program crashes.
You can turn warnings into errors with the `-Werror` option. I personally use that in CI runs, along with the `-X dev` option to enable additional runtime checks. Though that wont solve the author's problem, since most Python devs don't use either of those options
In PHP I don’t think there is a native way to convert E_DEPRECATED into E_ERROR, but the most common testing framework has a quick way of doing the same.
https://docs.phpunit.de/en/12.5/configuration.html#the-failo...
When I update python version, python packages, container image, etc for a service, I take a quick look at CI output, in addition to the all the other checks I do (like a couple basic real-world-usage end-to-end usage tests), to "smoke test" whether something not caught by outright CI failure caused some subtle problem.
So, I do often see deprecation warnings in CI output, and fix them. Am I a bad developer?
I think the mistake here is making some warnings default-hidden. The developer who cares about the user running their the app in a terminal can add a line of code to suppress them for users, and be more aware of this whole topic as a result (and have it more evident near the entrypoint of the program, for later devs to see also).
I think that making warnings error or hidden removes warnings as a useful tool.
But this is an old argument: Who should see Python warnings? (2017) https://lwn.net/Articles/740804/
Author here! Agreed that are different levels of "engaged" from users, which is okay. The concerning part of this finding is that even dependent users that I know to be highly engaged didn't respond to the deprecation warnings, so they're not working for even the most engaged users.
There was this one library we depended on, it was sort of in limbo during the Python 2 -> 3 migration. During that period is was maintained by this one person who'd just delete older versions when never ones became available. In one year I think we had three or four instances where our CI and unit tests just broke randomly one day, because the APIs had changed and the old version of the library had been yanked.
In hindsight it actually helped us, because in frustrations we ended up setting up our own Python package repo and started to pay more attention to our dependencies.
> If you are a good developer, you'll have extensive unit test coverage and CI. You never see the unit test output (unless they fail) - so warnings go unnoticed.
In my opinion test suites should treat any output other than the reporter saying that a test passed as a test failure. In JavaScript I usually have part of my test harness record calls to the various console methods. At the end of each test it checks to see if any calls to those methods were made, and if they were it fails the tests and logs the output. Within tests if I expect or want some code to produce a message, I wrap the invocation of that code in a helper which requires two arguments: a function to call and an expected output. If the code doesn't output a matching message, doesn't output anything, or outputs something else then the helper throws and explains what went wrong. Otherwise it just returns the result of the called function:
This is dead simple code in most testing frameworks. It makes maintaining and working with the test suite becomes much easier as when something starts behaving differently it's immediately obvious rather than being hidden in a sea of noise. It makes working with dependencies easier because it forces you to acknowledge things like deprecation warnings when they get introduced and either solve them there or create an upgrade plan.That's why you, very early on, release code that slows the API down once it has been deprecated. Every place you issue a deprecation warning, you also sleep 60. Problem solved.
A solid idea. There should be a package for this.
Why is it that CI tools don't make warnings visible? Why are they ignored by default in the first place? Seems like that should be a rather high priority.
> Why is it that CI tools don't make warnings visible?
A developer setting up CI decides to start an ubuntu 24.04 container and run 'apt-get install npm'
This produces 3,600 lines of logging (5.4 log lines per package, 668 packages) and 22 warnings (all warnings about man page creation being skipped)
Then they decide "Nobody's going to read all that, and the large volume might bury important information. I think I'll hide console output for processes that don't fail."
Now your CI doesn't show warnings.
It isn't that easy. If you have a new warning on upgrade you probably want to work on it "next week", but that means you need to ignore it for a bit. Or you might still want to support a really old version without the new API and so you can't fix it now.
> If you have a new warning on upgrade you probably want to work on it "next week", but that means you need to ignore it for a bit.
So you create a bug report or an issue or a story or whatever you happen to call it, and you make sure it gets tracked, and you schedule it with the rest of your work. That's not the same thing as "ignoring" it.
And you always have something more important/interesting to do and so never get around to it.
... which means that when the axe falls, the results are 100 percent your fault.
Most of the time they don't matter and aren't an immediate problem.
The Business doesn't care about warnings, they want working software NOW.
If you are a good developer, you consider warnings to be errors until proven otherwise.
What does a good developer do when working in a codebase with hundreds of warnings?
Or are you only considering a certain warnings?
Why does your codebase generate hundreds of warnings, given that every time one initially appeared, you should have stamped it out (or specifically marked that one warning to be ignored)? Start with one line of code that doesn't generate a warning. Add a second line of code that doesn't generate a warning...
> Why does your codebase generate hundreds of warnings
Well, it wasn't my codebase yesterday, because I didn't work here.
Today I do. When I build, I get reports of "pkg_resources is deprecated as an API" and "Tesla T4 does not support bfloat16 compilation natively" and "warning: skip creation of /usr/share/man/man1/open.1.gz because associated file /usr/share/man/man1/xdg-open.1.gz (of link group open) doesn't exist" and "datetime.utcnow() is deprecated and scheduled for removal in a future version"
The person onboarding me tells me those warnings are because of "dependencies" and that I should ignore them.
It's rare that I work on a project I myself started. If I start working on an existing codebase, the warnings might be there already. Then what do I do?
I'm also referring to all the warnings you might get if you use an existing library. If the requirements entail that I use this library, should I just silence them all?
But I'm guessing you might be talking about more specific warnings. Yes I do fix lints specific to my new code before I commit it, but a lot of warnings might still be logged at runtime, and I may have no control over them.
> If I start working on an existing codebase, the warnings might be there already. Then what do I do?
What would you do if the code you inherited crashed all the time?
Come up with a strategy for fixing them steadily until they're gone.
If this code crashed all the time there'd be a business need to fix it and I could justify spending time on this.
But that's not what we're discussing here, we're discussing warnings that have been ignored in the past, and all of a sudden I'm supposed to take the political risk to fix them all somehow, even though there's no new crash, no new information.
I don't know how much freedom you have at your job; but I definitely can't just go to my manager and say: "I'm spending the next few weeks working on warnings nobody else cared about but that for some reason I care about".
Because most people are working at Failure/Feature factories where they might work on something and at last minute, they find out something is now warning. If they work on fixing it, the PM will screaming about time slippage and be like "I want you to work on X, not Y which can wait".
2 Years later, you have hundreds of warning.
You found that out at the last minute. So then you did a release. It's no longer the last minute. Now what's your excuse for the next release?
If your management won't resource your project to the point where you can assure that the software is correct, you might want to see if you can find the free time to look for another job. You'll have to do that anyway when they either tank the company, or lay you off next time they feel they need to cut more costs.
Start adding a 500ms delay to every invocation of a deprecated function. Incentives work!
Or even scale it up over time: start with 100ms of delay and crank it up the closer you get to deprecated
>We ended up adding the APIs back and creating a hurried release to fix the issue.
So it was entirely possible to keep the software working with these. Why change/remove them in the first place? Is the benefit of of the new abstraction greater than the downside of requiring everyone using the software to re-write theirs?
OS software maintainers don't like maintaining legacy ugly APIs forever and want to refactor/remove legacy code to keep themselves sane and the project maintainable.
Every public API change is a cost on the user: for an extreme example, if every library I ever used renamed half its APIs every year to align with the latest ontology, then there would hardly be any point in saving my scripts, since I'd have to be constantly rewriting them all.
Of course, the reality is hardly ever as bad as that, but I'd say having to deal with trivial API changes is a reasonable basis for a user to dislike a given project or try to avoid using it. It's up to the maintainers how friendly they want to be toward existing user code, and whether they pursue mitigating options like bundling migrations into less-common bigger updates.
Yep. Or you can see it as, "This software doesn't really care about the users and their use cases. It prioritizes making things look pretty and easier on the dev side over maintaining functionality." Or in the worse but fairly common OSS case, CADT, but that doesn't seem to apply in this context.
It’s like 5 lines of code https://github.com/urllib3/urllib3/pull/3732/files
I think this is a valid question for this specific case, but may not always be possible. That said, I think as a user I would probably prefer it if under the hood the old function called the new so they can deprecate the behavior without breaking the API. In that way you can still emit the deprecation warning while also only having one actual code path to maintain.
Funny enough, in Python, a sufficiently-dedicated client can also do this on their end by just monkey-patching the class definition.
Wild (and I guess most of the time bad) idea: on top of the warnings, introduce a `sleep` in the deprecated functions. At every version, increase the sleep.
Has this ever been considered?
The problem with warnings is that they're not really observable: few people actually read these logs, most of the time. Making the deprecation observable means annoying the library users. The question is then: what's the smallest annoyance we can come up with, so that they still have a look?
Yes, people do notice sleep. But it has to be on the scale of minutes or it will be ignored especially if it happens during a CI run.
I prefer Go's solution to this problem. Just don't deprecate stuff. And to make that possible, slow down and design stuff you will be willing to support forever.
Go does deprecate stuff. It just never removes it.
> Maybe the answer is to do away with advance notice and adopt SemVer with many major versions
Yes.
Deprecations don't work. Don't deprecate stuff without a really really good reason. The new API being cleaner is not a good reason. There are very few good reasons.
If you deprecate something in a popular library, you're forcing millions of people to do work. Waste time that could be used for something better, possibly at a time of your choice, not theirs. It was emitting warnings for 3 years... so you think everyone should have to rewrite their software every 3 years?
Especially for something like this. Only document it in a footnote, mark it as deprecated, etc - but don't remove the alias.
Don't break stuff, unless, to quote a famous work, you think your users are scum. Do you think your users are scum? Why do you hate your users?
Open source developers have no such obligation. They’re doing things on their own time. But hey, I’m sure they’d be willing to give you a full refund for the amount you paid them to use their work
I'm very confused on the debate of semver here - the fundamental principle seems very simple, and important.
"give me all updates to my core version that's still compatible"
Semver simply puts a 'protocol' to this - define your major version and off you go.
While in practice you could go and search for each and every library you use to check when/how they do breaking versions, but semver just allows matching a single number across the board - it makes it more consistent and error proof.
Because it's not "error proof". It's not even "more error proof". Relying on updates not breaking if the maintainer doesn’t intend it means relying on software being bug-free, and it should be immediately obvious that that would be a very very stupid thing to believe. Letting software you rely on change on its own is always dangerous no matter how much someone pinky promises that you can trust them to have thought of everything. Semver numbers are determined by humans who aren't oracles and therefore make mistakes all the time about what changes will break something for someone else. It may get people to think a little bit about whether they _think_ they're breaking something, but it's never safe to rely on that judgement because there's a good chance that they're wrong. And given that it's never safe to rely on that judgment, the process of pretending is worse than worthless; the pretense is inherently dangerous, because people start to unthinkingly trust it. If you care about things working, you pin and vendor (npm left-pad) your dependencies and only update a package if you _need_ to and only after explicit evaluation.
And if after reading that you think to yourself, "But BugsJustFindMe, I have unit tests that will catch semver mistakes!" I think you need to ask yourself what your unit tests tell you about semver code that they don't also tell you about non-semver code.
I think they are misreading the situation.
The devil is in the details. It seems `getHeaders` v. `headers` is non-security, non-performance related issue. Why people should spend time fixing these?
Even if getHeaders() has security/performance concerns, the better solution is to make it an alias to the newer headers.get() in this case. Keeping the old API is a small hassle to a handful of developers but breaking existing code puts a much bigger burden on a lot more users.
Ya, why not just alias old api calls to the new if implementation details changed?
If you (either directly or from SerpApi) are supporting the urllib3 folks (through a Tidelift subscription), then yes, that is a valid point.
Otherwise, I'd say that's a very brave comment you are making.
"Brave" in what sense? It's a legitimate question why an API is having a pattern deprecated and removed for what appears to be pattern reasons.
If he gave money valid point if not he must be brave?
Instead of reverting the updates to appease those who've been ignoring the warnings, a way forward could be to gate the old API behind an environment variable, with a name like "TEMPORARILY_REENABLE_DEPRECATED_APIS". And maybe put an expiry on the variable so it resets occasionally. Otherwise the project just becomes mired and stagnant.
A deprecation warning is not actionable for typical end users. Why don't more warnings include calls to action?
Instead of a warning that says, "The get_widget method in libfoo is deprecated and will be removed by November 30", the warning could say:
"This app uses the deprecated get_widget method. Please report this bug to the app developer. Ask them to fix this issue so you can continue using this app after November 30."
# Deprecated APIs resp.getheader("Content-Length")
# Recommended APIs resp.headers.get("Content-Length")
Why inflict this change on tens of millions of users? It's such a nonsense tiny busywork change. Let sleeping dogs lie.
This is exactly the sort of breaking change that I really struggle to see the value of — maintaining the deprecated method seems incredibly unlikely to be a notable maintenance burden when it is literally just:
Like sure — deprecate it, which might have _some_ downstream cost, rather than having two non-deprecated ways to do the same thing, just to make it clear which one people should be using; but removing it has a much more significant cost on every downstream user, and the cost of maintenance of the old API seems like it should be almost nothing.(I also don't hate the thought of having a `DeprecationWarning` subclass like `IndefiniteDeprecationWarning` to make it clear that there's no plan to remove the deprecated function, which can thus be ignored/be non-fatal in CI etc.)
There is value for the person maintaining this library cause they want it that way. If you develop a useful library and give it away for free then all power to you if you want to rearrange the furniture every 6 months. I'll roll with it.
I definitely agree with the sentiment that people working for free can do whatever the heck they want.
But if you're trying to help your users and grow your project, I think GP's advice is sound.
> If you develop a useful library and give it away for free then all power to you if you want to rearrange the furniture every 6 months.
That would make it no longer a useful library
The secret trick I've used on rare occasion, but when necessary, is the "ten second rule."
Users don't notice a deprecation warning. But they might notice adding a "time.sleep(10)" immediately at the top of the function. And that gives them one last grace period to change out their software before it breaks-breaks.
This is so much worse than just making the breaking change.
It depends on how we define "worse."
A breaking change causes a full-stop to a service.
An intentional slowdown lets the service continue to operate at degraded performance.
I concur that it's less clear for debugging purposes (although any reasonable debugging infrastructure should allow you to break and see what function you're in when the program hangs; definitely not as clear as the program crashing because the called function is gone, however).
A breaking change in a dependency doesn’t cause a full-stop to a service at all. The old version continues to work. Making subtly harmful changes so that new broken versions sneak in is just a bad idea and totally unnecessary.
> A breaking change in a dependency doesn’t cause a full-stop to a service at all
From the article:
"We still received feedback from users that this removal was unexpected and was breaking dependent libraries."
I think we may be assuming different floors on service maintainer competency; with so many users pulling in dependencies across an arbitrarily-wide version window with no testing, such changes do break services.
It’s not necessary to cater to the absolute least competent end user to begin with, but inserting slowdown bugs does not even achieve that. (Note that the bit about the breaking of dependent libraries you’re quoting is still not actually a service being affected.)
Just break, then revert when anyone complains, on every single release. eventually you will get a release where nobody complains as they move off the depreciated api due to breakage annoyance.
I think it's more likely that they move off the broken library due to breakage annoyance.
This will just waste CI compute and not solve anything.
It's worked in the past. But it does require someone at your org to care that CI times are spiking, which is not always a thing you can rely upon.
In addition: if CI is the only place the issue shows up, and never in a user interaction... Why does that software exist in the first place? In that context, the slowdown may be serving as a useful signal to the project to drop the entire dependency.
ETA: To be clear, I don't do this as a substitute for a regular deprecation cycle (clear documentation, clear language-supported warnings / annotations, clear timeline to deprecate); I do it in addition before the final yank that actually breaks end-users.
Most CI runs I see have more than a 10s variance.
Are you saying you wouldn't notice if your CI suddenly started taking twice as long, ten times as long, a hundred times as long to run?
Honestly, I think that the author already found (and rejected the solution):
> I could ask for more Python developers to run with warnings enabled, but solutions in the form of “if only we could all just” are a folly.
I get where he's coming from. But the facts are: the language provides a tool to warn users, the library is using that tool, and users are choosing to turn that tool off. That's fine, but then they don't get to complain when stuff breaks without warning. It is the user's responsibility at that point, not the library maintainers'.