Today I saw a short YouTube video about “cozy games” and started writing a comment, then realized that this was somehow prompting me to write the most succinct summary of my own personal views on politics and economics that I have ever managed. So, here goes.

Apparently all I needed to trim down 50,000 words on my annoyance at how the term “capitalism” is frustratingly both a nexus for useful critque and also reductive thought-terminating clichés was to realize that Animal Crossing: New Horizons is closer to my views on political economy than anything Adam Smith or Karl Marx ever wrote.

Cozy games illustrate that the core mechanics of capitalism are fun and motivating, in a laboratory environment. It’s fun to gather resources, to improve one’s skills, to engage in mutually beneficial exchanges, to collect things, to decorate. It’s tremendously motivating. Even merely pretending to do those things can captivate huge amounts of our time and attention.

In real life, people need to be motivated to do stuff. Not because of some moral deficiency, but because in a large complex civilization it’s hard to tell what needs doing. By the time it’s widely visible to a population-level democratic consensus of non-experts that there is an unmet need — for example, trash piling up on the street everywhere indicating a need for garbage collection — that doesn’t mean “time to pick up some trash”, it means “the sanitation system has collapsed, you’re probably going to get cholera”. We need a system that can identify utility signals more granularly and quickly, towards the edges of the social graph. To allow person A to earn “value credits” of some kind for doing work that others find valuable, then trade those in to person B for labor which they find valuable, even if it is not clearly obvious to anyone else why person A wants that thing. Hence: money.

So, a market can provide an incentive structure that productively steers people towards needs, by aggregating small price signals in a distributed way, via the communication technology of “money”. Authoritarian communist states are famously bad at this, overproducing “necessary” goods in ways that can hold their own with the worst excesses of capitalists, while under-producing “luxury” goods that are politically seen as frivolous.

This is the kernel of truth around which the hardcore capitalist bootstrap grindset ideologues build their fabulist cinematic universe of cruelty. Markets are motivating, they reason, therefore we must worship the market as a god and obey its every whim. Markets can optimize some targets, therefore we must allow markets to optimize every target. Markets efficiently allocate resources, and people need resources to live, therefore anyone unable to secure resources in a market is undeserving of life. Thus we begin at “market economies provide some beneficial efficiencies” and after just a bit of hand-waving over some inconvenient details, we get to “thus, we must make the poor into a blood-sacrifice to Moloch, otherwise nobody will ever work, and we will all die, drowning in our own laziness”. “The cruelty is the point” is a convenient phrase, but among those with this worldview, the prosperity is the point; they just think the cruelty is the only engine that can possibly drive it.

Cozy games are therefore a centrist¹ critique of capitalism. They present a world with the prosperity, but without the cruelty. More importantly though, by virtue of the fact that people actually play them in large numbers, they demonstrate that the cruelty is actually unnecessary.

You don’t need to play a cozy game. Tom Nook is not going to evict you from your real-life house if you don’t give him enough bells when it’s time to make rent. In fact, quite the opposite: you have to take time away from your real-life responsibilities and work, in order to make time for such a game. That is how motivating it is to engage with a market system in the abstract, with almost exclusively positive reinforcement.

What cozy games are showing us is that a world with tons of “free stuff” — universal basic income, universal health care, free education, free housing — will not result in a breakdown of our society because “no one wants to work”. People love to work.

If we can turn the market into a cozy game, with low stakes and a generous safety net, more people will engage with it, not fewer. People are not lazy; laziness does not exist. The motivation that people need from a market economy is not a constant looming threat of homelessness, starvation and death for themselves and their children, but a fun opportunity to get a five-star island rating.

Acknowledgments

Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor!

PEP 0593 added the ability to add arbitrary user-defined metadata to type annotations in Python.

At type-check time, such annotations are… inert. They don’t do anything. Annotated[int, X] just means int to the type-checker, regardless of the value of X. So the entire purpose of Annotated is to provide a run-time API to consume metadata, which integrates with the type checker syntactically, but does not otherwise disturb it.

Yet, the documentation for this central purpose seems, while not exactly absent, oddly incomplete.

The PEP itself simply says:

A tool or library encountering an Annotated type can scan through the annotations to determine if they are of interest (e.g., using isinstance()).

But it’s not clear where “the annotations” are, given that the PEP’s entire “consuming annotations” section does not even mention the __metadata__ attribute where the annotation’s arguments go, which was only even added to CPython’s documentation. Its list of examples just show the repr() of the relevant type.

There’s also a bit of an open question of what, exactly, we are supposed to isinstance()-ing here. If we want to find arguments to Annotated, presumably we need to be able to detect if an annotation is an Annotated. But isinstance(Annotated[int, "hello"], Annotated) is both False at runtime, and also a type-checking error, that looks like this:

1

Argument 2 to "isinstance" has incompatible type "<typing special form>"; expected "_ClassInfo"

The actual type of these objects, typing._AnnotatedAlias, does not seem to have a publicly available or documented alias, so that seems like the wrong route too.

Now, it certainly works to escape-hatch your way out of all of this with an Any, build some version-specific special-case hacks to dig around in the relevant namespaces, access __metadata__ and call it a day. But this solution is … unsatisfying.

What are you looking for?

Upon encountering these quirks, it is understandable to want to simply ask the question “is this annotation that I’m looking at an Annotated?” and to be frustrated that it seems so obscure to straightforwardly get an answer to that question without disabling all type-checking in your meta-programming code.

However, I think that this is a slight misframing of the problem. Code that is inspecting parameters for an annotation is going to do something with that annotation, which means that it must necessarily be looking for a specific set of annotations. Therefore the thing we want to pass to isinstance is not some obscure part of the annotations’ internals, but the actual interesting annotation type from your framework or application.

When consuming an Annotated parameter, there are 3 things you probably want to know:

1. What was the parameter itself? (type: The type you passed in.)
2. What was the name of the annotated object (i.e.: the parameter name, the attribute name) being passed the parameter? (type: str)
3. What was the actual type being annotated? (type: type)

And the things that we have are the type of the Annotated we’re querying for, and the object with annotations we are interrogating. So that gives us this function signature:

1
2
3
4
5

def annotated_by(
    annotated: object,
    kind: type[T],
) -> Iterable[tuple[str, T, type]]:
    ...

To extract this information, all we need are get_args and get_type_hints; no need for __metadata__ or get_origin or any other metaprogramming. Here’s a recipe:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

def annotated_by(
    annotated: object,
    kind: type[T],
) -> Iterable[tuple[str, T, type]]:
    for k, v in get_type_hints(annotated, include_extras=True).items():
        all_args = get_args(v)
        if not all_args:
            continue
        actual, *rest = all_args
        for arg in rest:
            if isinstance(arg, kind):
                yield k, arg, actual

It might seem a little odd to be blindly assuming that get_args(...)[0] will always be the relevant type, when that is not true of unions or generics. Note, however, that we are only yielding results when we have found the instance type in the argument list; our arbitrary user-defined instance isn’t valid as a type annotation argument in any other context. It can’t be part of a Union or a Generic, so we can rely on it to be an Annotated argument, and from there, we can make that assumption about the format of get_args(...).

This can give us back the annotations that we’re looking for in a handy format that’s easy to consume. Here’s a quick example of how you might use it:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

@dataclass
class AnAnnotation:
    name: str

def a_function(
    a: str,
    b: Annotated[int, AnAnnotation("b")],
    c: Annotated[float, AnAnnotation("c")],
) -> None:
    ...

print(list(annotated_by(a_function, AnAnnotation)))

# [('b', AnAnnotation(name='b'), <class 'int'>),
#  ('c', AnAnnotation(name='c'), <class 'float'>)]

Acknowledgments

Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor! I am also available for consulting work if you think your organization could benefit from expertise on topics like “how do I do Python metaprogramming, but, like, not super janky”.

Facebook — and by extension, most of Silicon Valley — rightly gets a lot of shit for its old motto, “Move Fast and Break Things”.

As a general principle for living your life, it is obviously terrible advice, and it leads to a lot of the horrific outcomes of Facebook’s business.

I don’t want to be an apologist for Facebook. I also do not want to excuse the worldview that leads to those kinds of outcomes. However, I do want to try to help laypeople understand how software engineers—particularly those situated at the point in history where this motto became popular—actually meant by it. I would like more people in the general public to understand why, to engineers, it was supposed to mean roughly the same thing as Facebook’s newer, goofier-sounding “Move fast with stable infrastructure”.

Move Slow

In the bad old days, circa 2005, two worlds within the software industry were colliding.

The old world was the world of integrated hardware/software companies, like IBM and Apple, and shrink-wrapped software companies like Microsoft and WordPerfect. The new world was software-as-a-service companies like Google, and, yes, Facebook.

In the old world, you delivered software in a physical, shrink-wrapped box, on a yearly release cycle. If you were really aggressive you might ship updates as often as quarterly, but faster than that and your physical shipping infrastructure would not be able to keep pace with new versions. As such, development could proceed in long phases based on those schedules.

In practice what this meant was that in the old world, when development began on a new version, programmers would go absolutely wild adding incredibly buggy, experimental code to see what sorts of things might be possible in a new version, then slowly transition to less coding and more testing, eventually settling into a testing and bug-fixing mode in the last few months before the release.

This is where the idea of “alpha” (development testing) and “beta” (user testing) versions came from. Software in that initial surge of unstable development was extremely likely to malfunction or even crash. Everyone understood that. How could it be otherwise? In an alpha test, the engineers hadn’t even started bug-fixing yet!

In the new world, the idea of a 6-month-long “beta test” was incoherent. If your software was a website, you shipped it to users every time they hit “refresh”. The software was running 24/7, on hardware that you controlled. You could be adding features at every minute of every day. And, now that this was possible, you needed to be adding those features, or your users would get bored and leave for your competitors, who would do it.

But this came along with a new attitude towards quality and reliability. If you needed to ship a feature within 24 hours, you couldn’t write a buggy version that crashed all the time, see how your carefully-selected group of users used it, collect crash reports, fix all the bugs, have a feature-freeze and do nothing but fix bugs for a few months. You needed to be able to ship a stable version of your software on Monday and then have another stable version on Tuesday.

To support this novel sort of development workflow, the industry developed new technologies. I am tempted to tell you about them all. Unit testing, continuous integration servers, error telemetry, system monitoring dashboards, feature flags... this is where a lot of my personal expertise lies. I was very much on the front lines of the “new world” in this conflict, trying to move companies to shorter and shorter development cycles, and move away from the legacy worldview of Big Release Day engineering.

Old habits die hard, though. Most engineers at this point were trained in a world where they had months of continuous quality assurance processes after writing their first rough draft. Such engineers feel understandably nervous about being required to ship their probably-buggy code to paying customers every day. So they would try to slow things down.

Of course, when one is deploying all the time, all other things being equal, it’s easy to ship a show-stopping bug to customers. Organizations would do this, and they’d get burned. And when they’d get burned, they would introduce Processes to slow things down. Some of these would look like:

1. Let’s keep a special version of our code set aside for testing, and then we’ll test that for a few weeks before sending it to users.
2. The heads of every department need to sign-off on every deployed version, so everyone needs to spend a day writing up an explanation of their changes.
3. QA should sign off too, so let’s have an extensive sign-off process where each individual tester does a fills out a sign-off form.

Then there’s my favorite version of this pattern, where management decides that deploys are inherently dangerous, and everyone should probably just stop doing them. It typically proceeds in stages:

1. Let’s have a deploy freeze, and not deploy on Fridays; don’t want to mess up the weekend debugging an outage.
2. Actually, let’s extend that freeze for all of December, we don’t want to mess up the holiday shopping season.
3. Actually why not have the freeze extend into the end of November? Don’t want to mess with Thanksgiving and the Black Friday weekend.
4. Some of our customers are in India, and Diwali’s also a big deal. Why not extend the freeze from the end of October?
5. But, come to think of it, we do a fair amount of seasonal sales for Halloween too. How about no deployments from October 10 onward?
6. You know what, sometimes people like to use our shop for Valentine’s day too. Let’s just never deploy again.

This same anti-pattern can repeat itself with an endlessly proliferating list of “environments”, whose main role ends up being to ensure that no code ever makes it to actual users.

… and break things anyway

As you may have begun to suspect, there are a few problems with this style of software development.

Even back in the bad old days of the 90s when you had to ship disks in boxes, this methodology contained within itself the seeds of its own destruction. As Joel Spolsky memorably put it, Microsoft discovered that this idea that you could introduce a ton of bugs and then just fix them later came along with some massive disadvantages:

The very first version of Microsoft Word for Windows was considered a “death march” project. It took forever. It kept slipping. The whole team was working ridiculous hours, the project was delayed again, and again, and again, and the stress was incredible. [...] The story goes that one programmer, who had to write the code to calculate the height of a line of text, simply wrote “return 12;” and waited for the bug report to come in [...]. The schedule was merely a checklist of features waiting to be turned into bugs. In the post-mortem, this was referred to as “infinite defects methodology”.

Which lead them to what is perhaps the most ironclad law of software engineering:

In general, the longer you wait before fixing a bug, the costlier (in time and money) it is to fix.

A corollary to this is that the longer you wait to discover a bug, the costlier it is to fix.

Some bugs can be found by code review. So you should do code review. Some bugs can be found by automated tests. So you should do automated testing. Some bugs will be found by monitoring dashboards, so you should have monitoring dashboards.

So why not move fast?

But here is where Facebook’s old motto comes in to play. All of those principles above are true, but here are two more things that are true:

1. No matter how much code review, automated testing, and monitoring you have some bugs can only be found by users interacting with your software.
2. No bugs can be found merely by slowing down and putting the deploy off another day.

Once you have made the process of releasing software to users sufficiently safe that the potential damage of any given deployment can be reliably limited, it is always best to release your changes to users as quickly as possible.

More importantly, as an engineer, you will naturally have an inherent fear of breaking things. If you make no changes, you cannot be blamed for whatever goes wrong. Particularly if you grew up in the Old World, there is an ever-present temptation to slow down, to avoid shipping, to hold back your changes, just in case.

You will want to move slow, to avoid breaking things. Better to do nothing, to be useless, than to do harm.

For all its faults as an organization, Facebook did, and does, have some excellent infrastructure to avoid breaking their software systems in response to features being deployed to production. In that sense, they’d already done the work to avoid the “harm” of an individual engineer’s changes. If future work needed to be performed to increase safety, then that work should be done by the infrastructure team to make things safer, not by every other engineer slowing down.

The problem is that slowing down is not actually value neutral. To quote myself here:

If you can’t ship a feature, you can’t fix a bug.

When you slow down just for the sake of slowing down, you create more problems.

The first problem that you create is smashing together far too many changes at once.

You’ve got a development team. Every engineer on that team is adding features at some rate. You want them to be doing that work. Necessarily, they’re all integrating them into the codebase to be deployed whenever the next deployment happens.

If a problem occurs with one of those changes, and you want to quickly know which change caused that problem, ideally you want to compare two versions of the software with the smallest number of changes possible between them. Ideally, every individual change would be released on its own, so you can see differences in behavior between versions which contain one change each, not a gigantic avalanche of changes where any one of hundred different features might be the culprit.

If you slow down for the sake of slowing down, you also create a process that cannot respond to failures of the existing code.

I’ve been writing thus far as if a system in a steady state is inherently fine, and each change carries the possibility of benefit but also the risk of failure. This is not always true. Changes don’t just occur in your software. They can happen in the world as well, and your software needs to be able to respond to them.

Back to that holiday shopping season example from earlier: if your deploy freeze prevents all deployments during the holiday season to prevent breakages, what happens when your small but growing e-commerce site encounters a catastrophic bug that has always been there, but only occurs when you have more than 10,000 concurrent users. The breakage is coming from new, never before seen levels of traffic. The breakage is coming from your success, not your code. You’d better be able to ship a fix for that bug real fast, because your only other option to a fast turn-around bug-fix is shutting down the site entirely.

And if you see this failure for the first time on Black Friday, that is not the moment where you want to suddenly develop a new process for deploying on Friday. The only way to ensure that shipping that fix is easy is to ensure that shipping any fix is easy. That it’s a thing your whole team does quickly, all the time.

The motto “Move Fast And Break Things” caught on with a lot of the rest of Silicon Valley because we are all familiar with this toxic, paralyzing fear.

After we have the safety mechanisms in place to make changes as safe as they can be, we just need to push through it, and accept that things might break, but that’s OK.

Some Important Words are Missing

The motto has an implicit preamble, “Once you have done the work to make broken things safe enough, then you should move fast and break things”.

When you are in a conflict about whether to “go fast” or “go slow”, the motto is not supposed to be telling you that the answer is an unqualified “GOTTA GO FAST”. Rather, it is an exhortation to take a beat and to go through a process of interrogating your motivation for slowing down. There are three possible things that a person saying “slow down” could mean about making a change:

1. It is broken in a way you already understand. If this is the problem, then you should not make the change, because you know it’s not ready. If you already know it’s broken, then the change simply isn’t done. Finish the work, and ship it to users when it’s finished.
2. It is risky in a way that you don’t have a way to defend against. As far as you know, the change works, but there’s a risk embedded in it that you don’t have any safety tools to deal with. If this is the issue, then what you should do is pause working on this change, and build the safety first.
3. It is making you nervous in a way you can’t articulate. If you can’t describe an known defect as in point 1, and you can’t outline an improved safety control as in step 2, then this is the time to let go, accept that you might break something, and move fast.

The implied context for “move fast and break things” is only in that third condition. If you’ve already built all the infrastructure that you can think of to build, and you’ve already fixed all the bugs in the change that you need to fix, any further delay will not serve you, do not have any further delays.

Unfortunately, as you probably already know,

This motto did a lot of good in its appropriate context, at its appropriate time. It’s still a useful heuristic for engineers, if the appropriate context is generally understood within the conversation where it is used.

However, it has clearly been taken to mean a lot of significantly more damaging things.

Purely from an engineering perspective, it has been reasonably successful. It’s less and less common to see people in the industry pushing back against tight deployment cycles. It’s also less common to see the basic safety mechanisms (version control, continuous integration, unit testing) get ignored. And many ex-Facebook engineers have used this motto very clearly under the understanding I’ve described here.

Even in the narrow domain of software engineering it is misused. I’ve seen it used to argue a project didn’t need tests; that a deploy could be forced through a safety process; that users did not need to be informed of a change that could potentially impact them personally.

Outside that domain, it’s far worse. It’s generally understood to mean that no safety mechanisms are required at all, that any change a software company wants to make is inherently justified because it’s OK to “move fast”. You can see this interpretation in the way that it has leaked out of Facebook’s engineering culture and suffused its entire management strategy, blundering through market after market and issue after issue, making catastrophic mistakes, making a perfunctory apology and moving on to the next massive harm.

In the decade since it has been retired as Facebook’s official motto, it has been used to defend some truly horrific abuses within the tech industry. You only need to visit the orange website to see it still being used this way.

Even at its best, “move fast and break things” is an engineering heuristic, it is not an ethical principle. Even within the context I’ve described, it’s only okay to move fast and break things. It is never okay to move fast and harm people.

So, while I do think that it is broadly misunderstood by the public, it’s still not a thing I’d ever say again. Instead, I propose this:

Make it safer, don’t make it later.

Acknowledgments

Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor! I am also available for consulting work if you think your organization could benefit from expertise on topics like “how do I make changes to my codebase, but, like, good ones”.

One of my longest-running frustrations with iOS is that the default mail app does not have a “share” action, making it impossible to do the one thing that a mail client needs to be able to do for me, which is to selectively turn messages into tasks. This deficiency has multiple components which makes it difficult to work around:

1. There is no UI to “share” a message from within the mail app, so you can’t share it to OmniFocus or a shortcut.
2. There is no way to query for a mail message in Shortcuts and then get some kind of “message” object, so there’s no way you can start in Shortcuts and manually invoke something.
3. There’s no such thing as MailKit on iOS, so third-party apps can’t query your mail database either.

To work around this, I’ve long subscribed to the “AirMail” app, which has a “message to omnifocus” action but is otherwise kind of a buggy mess.

But today, I read that you can set up an iPad in a split-screen view and drag messages from the built-in Mail app’s message list view into the OmniFocus inbox, and I extrapolated, and discovered how to get Mail-message-to-Omnifocus-task work on an iPhone.

I’m thrilled that this functionality exists, but it is a bit of a masterclass in how to get a terrible UX out of a series of decisions that were probably locally reasonable. So, without any further ado, here’s how you do it:

1. Open up mail.app, and find the message you want to share in the message list. Here, you have two choices:
   1. With one finger, press and hold on a message until you feel a haptic. When you feel the haptic, immediately move your finger a little bit, before you see the preview come up. This lets you operate directly from the message list, but is very fiddly.
   2. Tap the message to open the detail view, then press and hold on the sent date in the top right.
2. Continue holding down with your first finger. With a second finger, swipe up from the bottom to enter the multitasking view, or to go back to your home screen. While holding your first finger in place, either switch to or launch OmniFocus.
3. With your second finger, navigate to the Inbox, drag your first finger to the bottom of the list, and release it. Voila! You should have a task with a brief summary and a link back to the message.
4. Swipe up from the bottom to switch back to Mail, then archive the message.

One of the most unfortunate things about learning Python is that there are so many different ways to get it installed, and you need to choose one before you even begin. The differences can also be subtle and require technical depth to truly understand, which you don’t have yet.¹ Even experts can be missing information about which one to use and why.

There are perhaps more of these on macOS than on any other platform, and that’s the platform I primarily use these days. If you’re using macOS, I’d like to make it simple for you.

The One You Probably Want: Python.org

My recommendation is to use an official build from python.org.

I recommed the official installer for most uses, and if you were just looking for a choice about which one to use, you can stop reading now. Thanks for your time, and have fun with Python.

If you want to get into the nerdy nuances, read on.

For starters, the official builds are compiled in such a way that they will run on a wide range of macs, both new and old. They are universal2 binaries, unlike some other builds, which means you can distribute them as part of a mac application.

The main advantage that the Python.org build has, though, is very subtle, and not any concrete technical detail. It’s a social, structural issue: the Python.org builds are produced by the people who make CPython, who are more likely to know about the nuances of what options it can be built with, and who are more likely to adopt their own improvements as they are released. Third party builders who are focused on a more niche use-case may not realize that there are build options or environment requirements that could make their Pythons better.

I’m being a bit vague deliberately here, because at any particular moment in time, this may not be an advantage at all. Third party integrators generally catch up to changes, and eventually achieve parity. But for a specific upcoming example, PEP 703 will have extensive build-time implications, and I would trust the python.org team to be keeping pace with all those subtle details immediately as releases happen.

(And Auto-Update It)

The one downside of the official build is that you have to return to the website to check for security updates. Unlike other options described below, there’s no built-in auto-updater for security patches. If you follow the normal process, you still have to click around in a GUI installer to update it once you’ve clicked around on the website to get the file.

I have written a micro-tool to address this and you can pip install mopup and then periodically run mopup and it will install any security updates for your current version of Python, with no interaction besides entering your admin password.

(And Always Use Virtual Environments)

Once you have installed Python from python.org, never pip install anything globally into that Python, even using the --user flag. Always, always use a virtual environment of some kind. In fact, I recommend configuring it so that it is not even possible to do so, by putting this in your ~/.pip/pip.conf:

1
2

[global]
require-virtualenv = true

This will avoid damaging your Python installation by polluting it with libraries that you install and then forget about. Any time you need to do something new, you should make a fresh virtual environment, and then you don’t have to worry about library conflicts between different projects that you may work on.

If you need to install tools written in Python, don’t manage those environments directly, install the tools with pipx. By using pipx, you allow each tool to maintain its own set dependencies, which means you don’t need to worry about whether two tools you use have conflicting version requirements, or whether the tools conflict with your own code.²

The Others

There are, of course, several other ways to install Python, which you probably don’t want to use.

The One For Running Other People’s Code, Not Yours: Homebrew

In general, Homebrew Python is not for you.

The purpose of Homebrew’s python is to support applications packaged within Homebrew, which have all been tested against the versions of python libraries also packaged within Homebrew. It may upgrade without warning on just about any brew operation, and you can’t downgrade it without breaking other parts of your install.

Specifically for creating redistributable binaries, Homebrew python is typically compiled only for your specific architecture, and thus will not create binaries that can be used on Intel macs if you have an Apple Silicon machine, or will run slower on Apple Silicon machines if you have an Intel mac. Also, if there are prebuilt wheels which don’t yet exist for Apple Silicon, you cannot easily arch -x86_64 python ... and just install them; you have to install a whole second copy of Homebrew in a different location, which is a headache.

In other words, homebrew is an alternative to pipx, not to Python. For that purpose, it’s fine.

The One For When You Need 20 Different Pythons For Debugging: pyenv

Like Homebrew, pyenv will default to building a single-architecture binary. Even worse, it will not build a Framework build of Python, which means several things related to being a mac app just won’t work properly. Remember those build-time esoterica that the core team is on top of but third parties may not be? “Should I use a Framework build” is an enduring piece of said esoterica.

The purpose of pyenv is to provide a large matrix of different, precise legacy versions of python for library authors to test compatibility against those older Pythons. If you need to do that, particularly if you work on different projects where you may need to install some random super-old version of Python that you would not normally use to test something on, then pyenv is great. But if you only need one version of Python, it’s not a great way to get it.

The Other One That’s Exactly Like pyenv: asdf-python

The issues are exactly the same as with pyenv, as the tool is a straightforward alternative for the exact same purpose. It’s a bit less focused on Python than pyenv, which has pros and cons; it has broader community support, but it’s less specifically tuned for Python. But a comparative exploration of their differences is beyond the scope of this post.

The Built-In One That Isn’t Really Built-In: /usr/bin/python3

There is a binary in /usr/bin/python3 which might seem like an appealing option — it comes from Apple, after all! — but it is provided as a developer tool, for running things like build scripts. It isn’t for building applications with.

That binary is not a “system python”; the thing in the operating system itself is only a shim, which will determine if you have development tools, and shell out to a tool that will download the development tools for you if you don’t. There is unfortunately a lot of folk wisdom among older Python programmers who remember a time when apple did actually package an antedeluvian version of the interpreter that seemed to be supported forever, and might suggest it for things intended to be self-contained or have minimal bundled dependencies, but this is exactly the reason that Apple stopped shipping that.

If you use this option, it means that your Python might come from the Xcode Command Line Tools, or the Xcode application, depending on the state of xcode-select in your current environment and the order in which you installed them.

Upgrading Xcode via the app store or a developer.apple.com manual download — or its command-line tools, which are installed separately, and updated via the “settings” application in a completely different workflow — therefore also upgrades your version of Python without an easy way to downgrade, unless you manage multiple Xcode installs. Which, at 12G per install, is probably not an appealing option.³

The One With The Data And The Science: Conda

As someone with a limited understanding of data science and scientific computing, I’m not really qualified to go into the detailed pros and cons here, but luckily, Itamar Turner-Trauring is, and he did.

My one coda to his detailed exploration here is that while there are good reasons to want to use Anaconda — particularly if you are managing a data-science workload across multiple platforms and you want a consistent, holistic development experience across a large team supporting heterogenous platforms — some people will tell you that you need Conda to get you your libraries if you want to do data science or numerical work with Python at all, because Conda is how you install those libraries, and otherwise things just won’t work.

This is a historical artifact that is no longer true. Over the last decade, Python Wheels have been comprehensively adopted across the Python community, and almost every popular library with an extension module ships pre-built binaries to multiple platforms. There may be some libraries that only have prebuilt binaries for conda, but they are sufficiently specialized that I don’t know what they are.

The One for Being Consistent With Your Cloud Hosting

Another way to run Python on macOS is to not run it on macOS, but to get another computer inside your computer that isn’t running macOS, and instead run Python inside that, usually using Docker.⁴

There are good reasons to want to use a containerized configuration for development, but they start to drift away from the point of this post and into more complicated stuff about how to get your Python into the cloud.

So rather than saying “use Python.org native Python instead of Docker”, I am specifically not covering Docker as a replacement for a native mac Python here because in a lot of cases, it can’t be one. Many tools require native mac facilities like displaying GUIs or scripting applications, or want to be able to take a path name to a file without elaborate pre-work to allow the program to access it.

Summary

If you didn’t want to read all of that, here’s the summary.

If you use a mac:

1. Get your Python interpreter from python.org.
2. Update it with mopup so you don’t fall behind on security updates.
3. Always use venvs for specific projects, never pip install anything directly.
4. Use pipx to manage your Python applications so you don’t have to worry about dependency conflicts.
5. Don’t worry if Homebrew also installs a python executable, but don’t use it for your own stuff.
6. You might need a different Python interpreter if you have any specialized requirements, but you’ll probably know if you do.

Acknowledgements

Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor! I am also available for consulting work if you think your organization could benefit from expertise on topics like “which Python is the really good one”.