Remove all options from Homebrew/homebrew-core formulae #31510
Comments
MikeMcQuaid
added
help wanted
|
What about a more general approach akin to Debian's alternatives system? That could handle the issue with options as well as other things. One thing that I've always missed in Homebrew was being able to pick an implementation of MPI: Open MPI vs Mpich. Currently when one installs Mpich, packages depending on MPI (and so on Open MPI in Homebrew's logic), such as Scalapack, have to be built manually, else Homebrew complains about missing dependencies. |
Those things do happen with options, but I don't think it follows that eliminating options provides a better user experience. A formula that lacks needed options means the user makes their own tap, in which case...it will be built from source, not tested by CI, and provides new chances for failure. I.e., those three things that an option-using user currently faces will still be faced just about as often once options are eliminated. |
We can't really do this given our limited CI and options aren't a good approach for this. "Picking an implementation" isn't really something Debian supports for compile-time things; as far as I'm away that only applies for things that can be swapped out underneath (i.e. are API and ABI compatible) without an application change. This is theoretically possible as something we could do but mostly people are requesting the ability to swap out things that aren't necessarily compatible.
I think a less flexible Homebrew/homebrew-core that works for every user every time is superior to a more flexible Homebrew/homebrew-core that works for every user 90% of the time. That's a product/design difference of opinion I realise we don't share. Essentially, though, you touch upon a good point: for things that have to be built from source or build using options there's no real difference between Homebrew/homebrew-core and a tap that doesn't provide bottles. As a result it's much, much easier for the community to step up there and provide solutions to these problems without adding the negative perception when Homebrew/homebrew-core doesn't work and the issues that result from hard to reproduce, option, from source builds. |
I personally tend to agree with @chdiza. (Although I'm willing to concede that I may be in the minority here, and I may not have had the extensive experience that led @MikeMcQuaid to those conclusions.) However, if we go this route, I think we need to greatly improve our documentation for setting up dedicated taps, with all the bells and whistles like bottling etc. I was looking into this a few months ago, and there is no straightforward guide on how to accomplish this. I finally stumbled upon this repository, which seemed like the best example for me to follow. Better documentation for users/contributors of the bottling process would make this type of migration much less of a bitter pill for some (like me!) to swallow. |
The current documentation (e.g. https://docs.brew.sh/How-to-Create-and-Maintain-a-Tap and the
This doesn't really make sense; options are never bottled anyway so to remove them doesn't imply we need to provide an ability for others to bottle them.
The issue isn't documentation, really, it's setting up free building of macOS packages (on multiple OS versions) with free uploads to a free CDN. I don't think that problem is ours to solve, really, but again: if you're interested in solving it please feel free. |
|
I’ll review the documentation again. It may have been updated since I last
looked. I think some insight, not in to how to find a free CDN and CI/CD
service, but how we are doing it in our specific case, could go a long way
to teach by example. Yes finding the right free/cost effective combo is non
trivial, but when I was last looking it, I found the documentation of
shipping bottles somewhat opaque. I’ll review it and do some code
spelunking to try to improve it myself.
I realize the bottling issue is somewhat orthogonal to the options, however
an important use case for Homebrew is in CI pipelines. Being able to
install precompiled dependencies (built with appropriate options) can mean
the difference between a job finishing in time, or a build hitting the CI
time limit and failing.
I’m not voicing my opinion here to try to get Homebrew to reverse course on
this decision; I understand the motivations and trust in the process that
led to the decision being made. Anyway, now is the time to “get up or shut
up” I suppose, and go review the documentation and see if I can work up a
decent PR. 🙂
Cheers!
…On Wed, Sep 5, 2018 at 4:25 AM Mike McQuaid ***@***.***> wrote:
I think we need to greatly improve our documentation for setting up
dedicated taps, with all the bells and whistles like bottling etc. I was
looking into this a few months ago, and there is no straightforward guide
on how to accomplish this.
The current documentation (e.g.
https://docs.brew.sh/How-to-Create-and-Maintain-a-Tap and the brew tap-new
command in the manpage) is sufficient for many people to setup taps. If you
would like to extend that documentation: please feel free.
with all the bells and whistles like bottling etc.
This doesn't really make sense; options are never bottled anyway so to
remove them doesn't imply we need to provide an ability for others to
bottle them.
Better documentation for users/contributors of the bottling process would
make this type of migration much less of a bitter pill for some (like me!)
to swallow.
The issue isn't documentation, really, it's setting up free building of
macOS packages (on multiple OS versions) with free uploads to a free CDN. I
don't think that problem is ours to solve, really, but again: if you're
interested in solving it please feel free.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#31510 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AAREPOm8dgPyVzRY-ormPrx99muQhI14ks5uX4qQgaJpZM4WNX12>
.
|
|
For the record: I think the mass removal of options is patently ridiculous & achieves nothing except stymying debate about individual options because those PRs are too large to easily discuss just one or two elements. I also think it is ridiculous to move ahead with this after 10 days of this issue being opened, when precisely two maintainers thus far have weighed in on the subject, and the core team as a whole was essentially rebuilt from the ground up less than one month ago. Some of these options also have effectively zero maintenance burden, or have never been knowingly problematic in terms of either analytic-logged build failures or human reports, and yet get removed anyway. Other options were the product of days or weeks of discussions about how best to resolve other issues, like Perl binding versioning & incompatibility, and have been casually thrown out the window with zero debate. I truly tip my hat to the "Ram the change through as rapidly as possible and hope most people, including maintainers, don't notice or object until it's too late" tactic. |
|
We’re all passionate here and care deeply about Homebrew. As far as I can
tell no action has been taken yet. Let’s stick to the facts and let’s get
others from @Homebrew-core to weigh in. I don’t think bristly accusations
are warranted. (Last paragraph)
There is a degree of obligation on us maintainers to follow these proposals
and issue tickets and voice our opinions. I’m not sure that putting every
change to a vote would be productive, but it would be good to ensure
maintainers feel some ownership and involvement in major policy decisions.
I also realize that I don’t have all the answers.
…On Wed, Sep 5, 2018 at 6:14 PM Dominyk Tiller ***@***.***> wrote:
For the record: I think the mass removal of options is patently ridiculous
& achieves nothing except stymying debate about individual options because
those PRs are too large to easily discuss *just* one or two elements.
I also think it is ridiculous to move ahead with this after 10 days of
this issue being opened, when precisely two maintainers thus far have
weighed in on the subject, and the core team as a whole was essentially
rebuilt from the ground up less than one month ago.
Some of these options also have effectively zero maintenance burden, or
have never been knowingly problematic in terms of either analytic-logged
build failures or human reports, and yet get removed anyway. Other options
were the process of days or weeks of discussions about how best to resolve
other issues, like Perl binding versioning & incompatibility, and have been
casually thrown out the window with zero debate.
I truly tip my hat to the *"Ram the change through as rapidly as possible
and hope most people, including maintainers, don't notice or object until
it's too late"* tactic.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#31510 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AAREPMf5nBpcV_6MEJVW4Jc4SWioWezJks5uYEzagaJpZM4WNX12>
.
|
|
Fair. I'm on the train and my internet connection is terrible, so I didn't do due diligence before I made that statement. While I personally like options, and find that they add value to Homebrew for power users or devs who need a certain package built a certain way, I do see that there is a trade off and it's not a one sided issue. If it were put to a vote (I'm NOT saying the it necessarily should be...) I would vote to keep --devel and --HEAD options, and want to evaluate other packages on a case by case basis. But anyway, I'd be curious to hear what other from @Homebrew/core think, if they have a strong opinion. There's a fairly good chance that they all agree with Mike which is why they didn't feel the need to comment here... 🤷♂️ |
|
On the whole I'm not a fan of options for the reasons cited by Mike. I do like |
I'll be holding the open PR, and not opening new ones. I apologise if you felt those were rushed: it clearly was not my intention: I just got a little time to do cleaning-type stuff while I was sat at a conference, and took it. Given that: 1. the PR was open for more than a week, 2. it doesn't look much like a call for discussion, but a plan for action, and 3. no maintainer has objected, I figured it was OK to move forward. I've tried to remove options from formulas with very few install counts, i.e. options removed should be lower than 0.5% in usage or lower than 5 in absolute install count (per 30 days). |
|
As to my personal opinion: apart from a few formulas where they are "natural" (aspell, some video or audio players), options run counter to our main goal, i.e. “just working”, for the arguments given by Mike. If we want to keep many options, I'd like us to consider running all option combinations through CI (which limits their number) or officially declare them “semi-supported” (i.e., we remove them when they break, unless someone patches it). |
|
I feel that keeping things lean and simple is a good thing. Also, existing documentation on maintaining a user tap is just fine. What we could do better is advertise the user tap feature more often. As in, whenever options pop up as a topic in a PR. |
@DomT4 I strongly agree with @zbeekman: this was not an acceptable way to refer to the work of other maintainers such as @fxcoudert who received a PR approval before merging based on an open "help wanted" issue.
Again, I agree with @zbeekman: this issue was open for 10 days in which you have been active on this issue tracker and in Slack and you had no comments on it during that time. You've also (re)joined the team more recently than @fxcoudert; he has no obligation to ask you for permission to make a change like this.
You have nothing to apologise for @fxcoudert. It's worth noting we've not been accepting options on new formula for a long time now. When we imported formulae from taps (many of which were in our top 1000 and sometimes top 100) we've removed all options in doing so. This has resulted in dramatically fewer build errors and those formulae being easier to support. To be extra explicit: this has been an undocumented but planned policy for several years.
@claui I agree.
@fxcoudert Is the only person in this thread to even offer a suggestion of how to resolve any of these issues (run them all through CI). If someone points out an engineering problem like this: suggest an alternative solution. |
To be clear on the scope of this PR: I do not consider |
@sjackman on this it's worth noting:
so even if option A and option B may both work fine when both specified together they produce a failure. This produces a dilemma: does option A get removed, option B get removed or what? I'm just personally unconvinced that this is a problem worth us dealing with when taps can maintain optionful formulae as well as we can. |
Each new formula provides a new chance for failures to occur, and I don't see much objection to adding new formulas. Each formula version bump provides a new chance for failures to occur, but HB bumps versions all the time. So it can't just be that options will bring more failures or whatever, but that the cost isn't worth the benefit. I've seen no argument whatsoever for this last part. I've only seen a list of potential costs. Some of these alleged costs really don't seem costly at all to me, and are unlikely to be seen as costly by an option-wanting user. Options won't have CI, but so what? None of Homebrew had CI for a long time, and while it's certainly better with CI, a CI-less Homebrew was much better than no Homebrew at all. Same for options. Options must be built from source, but so what? Having to build it from source is better than not having it at all. But even if they really were costly, still what I'm seeing is: "We won't have options, because that would suck for you: you'd have no CI and you'd have to build from source, and also the build might not work. So our advice is to go make your own tap, where you will build from source and not have CI, and where the build might not work." That is so absurd that I think the CI and source-build stuff are just extras obscuring the real motivation, and the real underlying thing is "We won't have options, because that sucks for us: it will generate tickets. So our advice is to go make your own tap; and if your build fails you shouldn't make a ticket (except maybe when it seems clearly our fault)." This is also absurd, but in a totally different way. |
Each new formula can be bottled. Bottle failures are much rarer than build from source failures (which are rarer than non-CI-tested build-from-source failures e.g. options).
Yes, and it sucked. I was a maintainer before we had CI and no-one else in this conversation was. Homebrew was worse quality, had a poorer user experience and many fewer users and contributors as a result. Relatedly, I want Homebrew to outlive my involvement in the project and part of that is defining the scope of the project in relation to the expectations of the users and, frankly, yes making the project easier for the maintainers to maintain and focusing on the stuff that only we can do in this tap and not the stuff that a copy-paste into another tap will do identically.
And at this point the technical discussion ends and the hyperbole begins. If you're going to speak in this way this conversation is going to get locked very quickly. |
Expectation management is no easy task. I’d never turn down anyone who’d ask for advice over a custom formula. I still believe drawing a line somewhere is a healthy thing to do. Even if it’s only to manage, quote unquote:
How so? |
I can only apologise for having to prioritise where I spend the time I have for Homebrew. The vast majority of the time that I spent on Slack recently has been 1) having a really prolonged discussion with you post-lead maintainer stuff, 2) talking Claudia & Izaak through the I took a look at this issue, decided it would be quite a complex discussion and dared to assume we wouldn't be bombing ahead with it within a fortnight & I'd have time to generate a discussion point that wasn't overly-simplistic. If feedback is that urgent maybe point out in Slack that you'd like opinions sooner rather than later, as I did repeatedly with things like the As for your little personal ding at the end, yes, as someone who at times single-handedly maintained most of core for an extended period in the past & has the most commits to core of any maintainer today, I generally am a little surprised when people don't at least ping me for an opinion on things that represent a major change in core, which isn't to say one has to agree with that given opinion. ====== |
|
The major issue we run into taps is that there is no way to give a tap preference over homebrew-core when it comes to dependencies. Even when the tap is pinned. This makes the situation on our build server very complicated. For example in my case I need a version of However when brew installs poppler as a dependency of a core package (for example gdal in my case) it still installs the core version, not the one from the (pinned) tap. This differs from pinning in Debian, where a package from the pinned repo can be a drop-in replacement for the mainstream package. This may seem small, but it makes the situation unmanageable. I need a few dozen modified formulas in my tap which are built with different options than core. This I can do, however brew keeps running into conflicts because it has already installed the core version of the same formula as a dependency. I think if you expect people to run taps for customized formulae, brew should allow taps that provide replacements for core formula, and which get preferred even when installed as a dependency. |
This is intentional. We do not want to have random dependencies in our tree swapped out with versions that may not work.
To illustrate my example: if a package in Homebrew/homebrew-core relies on the
I've elaborated a bit on this here: Homebrew/brew#5615 (comment). I'm open to improving the docs to make this (i.e. use a different formula name) more obvious and well-documented. |
|
So now, instead of copy-pasting homebrew options from Google, I need to use brew edit all the time when one single option would have solved my problem. Even worse, editing formulas on my own is no better than the pre-homebrew age, when I read docs, carefully set up compiling options, and fail, and try again. Popular Homebrew options are at lease given a chance to be tested and solved by the whole community. Without options, installing packages are only getting harder. Please don't just remove all the options. Rather split the formulas into two taps, one ready-to-use out-of-the box (core), one with options that may not be tested fully, but can be improved together by the community. |
The sites you're Googling will be updated with correct instructions.
They aren't tested before updates: that's the reason for their removal.
This is what's happening except the latter is being maintained by the community outside of the Homebrew organisation. |
@jeroen Does Homebrew still give you conflicts if you use non-conflicting formula names in your tap (such as |
|
@jeroen I feel your pain on this. A solution I use is to put everything you need (even "mere" dependencies) in your own tap. Then just copy the HB formulas---even the ones you don't actually have to adjust the flags for---into your own tap, modifying the "depends_on" part as necessary to point explicitly to your own tap. E.g., put a formula for This means you won't get bottles, but that just goes to show why build-from-source package managers are Good and Necessary. I have about 60 kegs installed, and now none of them are from the core tap. They are from my tap, which carries several duplicates of core formulas (except for their dependency references are changed to point at my tap). This way, I will never have the rug pulled out from under me by a sudden HB decision to mess with deps or flags.
It should, but the above describes a way to ape this. Having to do |
@bumfo Would it help if Homebrew allowed you to write custom formula extensions, or some kind of formula patch, as @saagarjha suggested in a recent forum thread? Here’s the relevant quote from this thread that outlines the suggested feature:
|
@chdiza Such determination! Given that you’re already going that route: have you looked at https://github.com/portage-brew yet? (Not being snarky or something, I 100% mean it; I’ve only glanced at portage-brew briefly but it looks like a promising fork to me.)
That is actually a really cool idea. |
This is what I've debated doing as well. I've resisted for the obvious reason that it places an undue amount of maintenance on our personal user taps and why a centralized tap can help. One formula may have many - 20, 30, 40, ... - dependencies that I need to also manage and update. It's a burden I'm not stoked on taking on. Hopefully if portage-brew gains traction amongst the community then this should not be as much of an issue. |
I agree that centralising effort may be useful but disagree that this needs to be in the Homebrew organisation to be successful. |
|
@MikeMcQuaid Don't disagree with you on this. I think it may be easier and radiate a more inclusive feel for the end users, but it can work outside of Homebrew as well assuming enough power users the changes impact are willing to join and build up a critical mass. We're good :) |
|
My ignorance is showing, but is it possible to subscribe to a formula in core so that we can receive an alert if/when it undergoes an edit? |
|
Nope, sorry. Someone could build it on to of GitHub's APIs, though. |
I think it is the responsibility of the user at that point if I specifically decide to swap out core formulas from a pinned tap. We just need some way that makes it possible to install core formulas with another configuration than the default one. The current situation make this very difficult. On the one hand you want core to be low-mainenance with little flexility (no options), more like Debian, which is OK. But then at the same time you prevent users from swapping out formulas with a custom build, as you would do in with a custom deb/PPA. So now there is no solution anymore for users that need a non-default configuration of a formula.
Most libraries won't change their API or ABI based on the configuration. Usually only certain features of the software will be enabled/disabled. Configuring a library with or without support for some driver/lib rarely affects how the program is called or linked, usually they can be swapped out safely. Even if it would break a reverse dependency, at that point it is my own responsibility, not yours. But I do feel strongly that if you decide to remove all flexibility in core formulas to lower mainenance, you need to relax the requirement that core take precedence over all.
A different formula name is really not a solution, then I would also need to add all reverse dependencies of the formula to my tap with a dependency on the changed formula name. This makes the problem an order of magnitude larger. I think almost all users in this thread simply need some way to configure a core formula with features they need.. If options are no longer supported and you present taps as the solution, you should allow core formula to be swapped out with a custom one from a tap. |
|
I've been thinking about this for awhile now. My first interaction with the Homebrew community has been pretty negative, but I'm going to put that aside and give this another try. @MikeMcQuaid, would you be open to the idea of an "override" tap? This would functionally be the same as a custom yum or apt repo. The way I'm thinking this could work is that a user would specify a tap (or taps) that would be checked prior to checking the core formulae. Formulae installed from the "override" tap would satisfy the dependencies of core formulae. This would allow a user to substitute in a custom formulae, but also allow Homebrew to offload the support requests to tap maintainers. This would also help a third party tap aggregate common custom formulae into one tap. |
@voldemortensen We have custom repositories (taps): https://docs.brew.sh/Taps. These can be pinned with
This will not work because these formulae will end up breaking core formulae in ways that are hard to debug (e.g. missing symbols/libraries) and we will have to be the ones fielding those support requests. You can create your own taps with your own formulae that are installed however you like. You can't inject random formulae as dependencies for the ones we test and build and expect them to work. |
@jeroen Regardless of whose responsibility it is: we will be the ones receiving the bug reports so we'd rather not add functionality to increase our support burden, sorry.
Debian doesn't provide the same degree of user accessible support on GitHub if you have a package that breaks. We definitely do support swapping out formulae but we do not support swapping out the dependencies of formulae because, as I mentioned in the last comment: they will break.
Yes, there is: maintain your own version of the formula in a tap. As noted in Homebrew/brew#5622 the best way to do this is to use a different name so they can live side-by-side. This has been successfully working with e.g. the
"Most", "usually" and "rarely" add up to a non-trivial number of support requests at Homebrew's scale. We're not willing to support that.
Again: regardless of whose responsibility it is: we field the support requests so it ends up adding extra work for us. Core formulae remain the primary option because they are what is used by the vast majority of users. The majority of users configurations are now handled by the defaults we chose for the software in Homebrew/homebrew-core now.
Again: we do allow these formulae to be swapped out but not as dependencies of existing formulae. I don't think we can really continue discussion indefinitely on this closed issue. Where's there's specific requests for things that make life easier for tap authors without making things harder for the Homebrew maintainers: we have and will help. Where PRs are submitted that make life easier for tap authors: we will review them and may merge them. Finally, it's worth remembering that the Homebrew maintainers do not need to do anything. In fact, open source maintainers owe you nothing. If you want some functionality we are not willing to support: you need to be willing to do the work yourself (which may be more work than you wish given our constraints) or you need to accept what we're willing to provide you (for free, in our spare time). |
That's exactly how PPA's work in Ubuntu/Debian and it's fine. It's the only way to customize a formula if you don't want to support options in core formulae. Users that choose to swap out core libraries with a custom tap probably know what they are doing. It's not your responsibility to debug this. However your current policy of hardcoding a single config for all core formulae and at the same time not providing any way to swap them out with a custom built, is obviously leading to frustration by many users for which the default configuration is inappropriate. |
It's not exactly how they work because they both ship binaries and Homebrew/homebrew-core ship binaries but taps/options ship source builds. Even if it were literally identical: It's perhaps fine to them/you but t's not fine to us. It will create bugs that are not immediately obvious, it will create user confusion and it will create extra work for us (even if we could instantly detect these issues and close them). I'm relocking this thread in favour of our Discourse forum for more general discussion and specific issues (e.g. |
|
I’ve reopened the Discourse thread Maintaining a tap to provide options for a formula to discuss possible PRs that help tap authors do their work with less friction. @bumfo Feel free to follow up there if you wish. |
Closes #36066. Signed-off-by: Chongyu Zhu <i@lembacon.com>
Options in formulae don't produce a good user experience because they have to be built from source, we don't test them in CI and each combination of options provides a new chance for new failures to occur. We should seek to (eventually!) remove all options from formulae in Homebrew/homebrew-core in favour of enabling as much non-exclusive functionality as possible in a given formula for widely used options or encouraging the community to maintain their own custom options in a tap (e.g. https://github.com/denji/homebrew-nginx/blob/master/Formula/nginx-full.rb). As an absolute last resort if we need to depend on the same formula multiple times with different options (e.g.
#13133) we can consider vendoring formulae using
resourceblocks or even duplicating formulae.This should be done in multiple phases:
I'm open to thoughts on the above and different approaches to solve the problem as well as additional tooling required in Homebrew/brew to accomplish this.
The text was updated successfully, but these errors were encountered: