Amber’s excellent lightning talk on identity yesterday made me feel many feels, and reminded me of this excellent post by Patrick McKenzie about false assumptions regarding names.

While that list is helpful, it’s very light on positively-framed advice, i.e. “you should” rather than “you shouldn’t”. So I feel like I want to give a little bit of specific, prescriptive advice to programmers who might need to deal with names.

First and foremost: stop asking for unnecessary information. If I’m just authenticating to your system to download a comic book, you do not need to know my name. Your payment provider might need a billing address, but you absolutely do not need to store my name.

Okay, okay. I understand that may make your system seem a little impersonal, and you want to be able to greet me, or maybe have a name to show to other users beyond my login ID or email address that has to be unique on the site. Fine. Here’s what a good “name” field looks like:

Name:

You don’t need to break my name down into parts. If you just need a way to refer to me, then let me tell you whatever the heck I want. Honorific? Maybe I have more than one; maybe I don’t want you to use any.

And this brings me to “first name / last name”.

In most cases, you should not use these terms. They are oversimplifications of how names work, appropriate only for children in English-speaking countries who might not understand the subtleties involved and only need to know that one name comes before the other.

The terms you’re looking for are given name and surname, or perhaps family name. (“Middle name” might still be an appropriate term because that fills a more specific role.) But by using these more semantically useful terms, you include orders of magnitude more names in your normalization scheme. More importantly, by acknowledging the roles of the different parts of a name, you’ll come to realize that there are many other types of name, such as:

• Mononym
• Generation Name
• Clan name
• Matronym and Patronym
• Teknonym
• Nickname

If your application does have a legitimate need to normalize names, for example, to interoperate with third-party databases, or to fulfill some regulatory requirement:

• When you refer to a user of the system, always allow them to customize how their name is presented. Give them the benefit of the doubt. If you’re concerned about users abusing this display-name system to insult other users, it's understandable that you may need to moderate that a little. But there's no reason to ever moderate or regulate how a user's name is displayed to themselves. You can start to address offensive names by allowing other users to set nicknames for them. Only as a last resort, allow other users to report their name as not-actually-their-name, abusive or rude; if you do that, you have to investigate those reports. Let users affirm other users’ names, too, and verify reports: if someone attracts a million fake troll accounts, but all their friends affirm that their name is correct, you should be able to detect that. Don’t check government IDs in order to do this; they’re not relevant.

• Allow the user to enter their normalized name as a series of names with classifiers attached to each one. In other words, like this:
  

  given namesurnamematronympatronymfirst namemiddle name... surnamegiven namematronympatronymfirst namemiddle name... +

• Keep in mind that spaces are valid in any of these names. Many people have multi-word first names, middle names, or last names, and it can matter how you classify them. For one example that should resonate with readers of this blog, it’s “Guido” “van Rossum”, not “Guido” “Van” “Rossum”. It is definitely not “Guido” “Vanrossum”.

• So is other punctuation. Even dashes. Even apostrophes. Especially apostrophes, you insensitive clod. Literally ten billion people whose surnames start with “O’” live in Ireland and they do not care about your broken database input security practices.
• Allow for the user to have multiple names with classifiers attached to each one: “legal name in China”, “stage name”, “name on passport”, “maiden name”, etc. Keep in mind that more than one name for a given person may simultaneously accurate for a certain audience and legally valid. They can even be legally valid in the same context: many people have social security cards, birth certificates, driver’s licenses and passports with different names on them; sometimes due to a clerical error, sometimes due to the way different systems work. If your goal is to match up with those systems, especially more than one of them, you need to account for that possibility.

If you’re a programmer and you’re balking at this complexity: good. Remember that for most systems, sticking with the first option - treating users’ names as totally opaque text - is probably your best bet. You probably don’t need to know the structure of the user’s name for most purposes.

attrs is still an excellent library, and it still has much to recommend it over the standard library for many applications. However, as of this update in 2023, much new code — perhaps even most of it — should just be using dataclasses. If you want to see where my thinking has moved on to in the intervening 7 years, you may want to check out this more recent post. To make a long story short though, I was right, and attrs set us on the road to a major upgrade to best practices for class definition in Python.

Do you write programs in Python? You should be using attrs.

Why, you ask? Don’t ask. Just use it.

Okay, fine. Let me back up.

I love Python; it’s been my primary programming language for 10+ years and despite a number of interesting developments in the interim I have no plans to switch to anything else.

But Python is not without its problems. In some cases it encourages you to do the wrong thing. Particularly, there is a deeply unfortunate proliferation of class inheritance and the God-object anti-pattern in many libraries.

One cause for this might be that Python is a highly accessible language, so less experienced programmers make mistakes that they then have to live with forever.

But I think that perhaps a more significant reason is the fact that Python sometimes punishes you for trying to do the right thing.

The “right thing” in the context of object design is to make lots of small, self-contained classes that do one thing and do it well. For example, if you notice your object is starting to accrue a lot of private methods, perhaps you should be making those “public”¹ methods of a private attribute. But if it’s tedious to do that, you probably won’t bother.

Another place you probably should be defining an object is when you have a bag of related data that needs its relationships, invariants, and behavior explained. Python makes it soooo easy to just define a tuple or a list. The first couple of times you type host, port = ... instead of address = ... it doesn’t seem like a big deal, but then soon enough you’re typing [(family, socktype, proto, canonname, sockaddr)] = ... everywhere and your life is filled with regret. That is, if you’re lucky. If you’re not lucky, you’re just maintaining code that does something like values[0][7][4][HOSTNAME][“canonical”] and your life is filled with garden-variety pain rather than the more complex and nuanced emotion of regret.

This raises the question: is it tedious to make a class in Python? Let’s look at a simple data structure: a 3-dimensional cartesian coordinate. It starts off simply enough:

1

class Point3D(object):

So far so good. We’ve got a 3 dimensional point. What next?

1
2

class Point3D(object):
    def __init__(self, x, y, z):

Well, that’s a bit unfortunate. I just want a holder for a little bit of data, and I’ve already had to override a special method from the Python runtime with an internal naming convention? Not too bad, I suppose; all programming is weird symbols after a fashion.

At least I see my attribute names in there, that makes sense.

1
2
3

class Point3D(object):
    def __init__(self, x, y, z):
        self.x

I already said I wanted an x, but now I have to assign it as an attribute...

1
2
3

class Point3D(object):
    def __init__(self, x, y, z):
        self.x = x

... to x? Uh, obviously ...

1
2
3
4
5

class Point3D(object):
    def __init__(self, x, y, z):
        self.x = x
        self.y = y
        self.z = z

... and now I have to do that once for every attribute, so this actually scales poorly? I have to type every attribute name 3 times?!?

Oh well. At least I’m done now.

1
2
3
4
5
6

class Point3D(object):
    def __init__(self, x, y, z):
        self.x = x
        self.y = y
        self.z = z
    def __repr__(self):

Wait what do you mean I’m not done.

1
2
3
4
5
6
7
8

class Point3D(object):
    def __init__(self, x, y, z):
        self.x = x
        self.y = y
        self.z = z
    def __repr__(self):
        return (self.__class__.__name__ +
                ("(x={}, y={}, z={})".format(self.x, self.y, self.z)))

Oh come on. So I have to type every attribute name 5 times, if I want to be able to see what the heck this thing is when I’m debugging, which a tuple would have given me for free?!?!?

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

class Point3D(object):
    def __init__(self, x, y, z):
        self.x = x
        self.y = y
        self.z = z
    def __repr__(self):
        return (self.__class__.__name__ +
                ("(x={}, y={}, z={})".format(self.x, self.y, self.z)))
    def __eq__(self, other):
        if not isinstance(other, self.__class__):
            return NotImplemented
        return (self.x, self.y, self.z) == (other.x, other.y, other.z)

7 times?!?!?!?

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

class Point3D(object):
    def __init__(self, x, y, z):
        self.x = x
        self.y = y
        self.z = z
    def __repr__(self):
        return (self.__class__.__name__ +
                ("(x={}, y={}, z={})".format(self.x, self.y, self.z)))
    def __eq__(self, other):
        if not isinstance(other, self.__class__):
            return NotImplemented
        return (self.x, self.y, self.z) == (other.x, other.y, other.z)
    def __lt__(self, other):
        if not isinstance(other, self.__class__):
            return NotImplemented
        return (self.x, self.y, self.z) < (other.x, other.y, other.z)

9 times?!?!?!?!?

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18

from functools import total_ordering
@total_ordering
class Point3D(object):
    def __init__(self, x, y, z):
        self.x = x
        self.y = y
        self.z = z
    def __repr__(self):
        return (self.__class__.__name__ +
                ("(x={}, y={}, z={})".format(self.x, self.y, self.z)))
    def __eq__(self, other):
        if not isinstance(other, self.__class__):
            return NotImplemented
        return (self.x, self.y, self.z) == (other.x, other.y, other.z)
    def __lt__(self, other):
        if not isinstance(other, self.__class__):
            return NotImplemented
        return (self.x, self.y, self.z) < (other.x, other.y, other.z)

Okay, whew - 2 more lines of code isn’t great, but now at least we don’t have to define all the other comparison methods. But now we’re done, right?

1
2

from unittest import TestCase
class Point3DTests(TestCase):

You know what? I’m done. 20 lines of code so far and we don’t even have a class that does anything; the hard part of this problem was supposed to be the quaternion solver, not “make a data structure which can be printed and compared”. I’m all in on piles of undocumented garbage tuples, lists, and dictionaries it is; defining proper data structures well is way too hard in Python.²

namedtuple to the (not really) rescue

The standard library’s answer to this conundrum is namedtuple. While a valiant first draft (it bears many similarities to my own somewhat embarrassing and antiquated entry in this genre) namedtuple is unfortunately unsalvageable. It exports a huge amount of undesirable public functionality which would be a huge compatibility nightmare to maintain, and it doesn’t address half the problems that one runs into. A full enumeration of its shortcomings would be tedious, but a few of the highlights:

• Its fields are accessable as numbered indexes whether you want them to be or not. Among other things, this means you can’t have private attributes, because they’re exposed via the apparently public __getitem__ interface.
• It compares equal to a raw tuple of the same values, so it’s easy to get into bizarre type confusion, especially if you’re trying to use it to migrate away from using tuples and lists.
• It’s a tuple, so it’s always immutable. Sort of.

As to that last point, either you can use it like this:

1

Point3D = namedtuple('Point3D', ['x', 'y', 'z'])

in which case it doesn’t look like a type in your code; simple syntax-analysis tools without special cases won’t recognize it as one. You can’t give it any other behaviors this way, since there’s nowhere to put a method. Not to mention the fact that you had to type the class’s name twice.

Alternately you can use inheritance and do this:

1
2

class Point3D(namedtuple('_Point3DBase', 'x y z'.split()])):
    pass

This gives you a place you can put methods, and a docstring, and generally have it look like a class, which it is... but in return you now have a weird internal name (which, by the way, is what shows up in the repr, not the class’s actual name). However, you’ve also silently made the attributes not listed here mutable, a strange side-effect of adding the class declaration; that is, unless you add __slots__ = 'x y z'.split() to the class body, and then we’re just back to typing every attribute name twice.

And this doesn’t even mention the fact that science has proven that you shouldn’t use inheritance.

So, namedtuple can be an improvement if it’s all you’ve got, but only in some cases, and it has its own weird baggage.

Enter The attr

So here’s where my favorite mandatory Python library comes in.

Let’s re-examine the problem above. How do I make Point3D with attrs?

1
2

import attr
@attr.s

Since this isn’t built into the language, we do have to have 2 lines of boilerplate to get us started: the import and the decorator saying we’re about to use it.

1
2
3

import attr
@attr.s
class Point3D(object):

Look, no inheritance! By using a class decorator, Point3D remains a Plain Old Python Class (albeit with some helpful double-underscore methods tacked on, as we’ll see momentarily).

1
2
3
4

import attr
@attr.s
class Point3D(object):
    x = attr.ib()

It has an attribute called x.

1
2
3
4
5
6

import attr
@attr.s
class Point3D(object):
    x = attr.ib()
    y = attr.ib()
    z = attr.ib()

And one called y and one called z and we’re done.

We’re done? Wait. What about a nice string representation?

1
2

>>> Point3D(1, 2, 3)
Point3D(x=1, y=2, z=3)

Comparison?

1
2
3
4
5
6

>>> Point3D(1, 2, 3) == Point3D(1, 2, 3)
True
>>> Point3D(3, 2, 1) == Point3D(1, 2, 3)
False
>>> Point3D(3, 2, 3) > Point3D(1, 2, 3)
True

Okay sure but what if I want to extract the data defined in explicit attributes in a format appropriate for JSON serialization?

1
2

>>> attr.asdict(Point3D(1, 2, 3))
{'y': 2, 'x': 1, 'z': 3}

Maybe that last one was a little on the nose. But nevertheless, it’s one of many things that becomes easier because attrs lets you declare the fields on your class, along with lots of potentially interesting metadata about them, and then get that metadata back out.

1
2
3
4
5

>>> import pprint
>>> pprint.pprint(attr.fields(Point3D))
(Attribute(name='x', default=NOTHING, validator=None, repr=True, cmp=True, hash=True, init=True, convert=None),
 Attribute(name='y', default=NOTHING, validator=None, repr=True, cmp=True, hash=True, init=True, convert=None),
 Attribute(name='z', default=NOTHING, validator=None, repr=True, cmp=True, hash=True, init=True, convert=None))

I am not going to dive into every interesting feature of attrs here; you can read the documentation for that. Plus, it’s well-maintained, so new goodies show up every so often and I might miss something important. But attrs does a few key things that, once you have them, you realize that Python was sorely missing before:

1. It lets you define types concisely, as opposed to the normally quite verbose manual def __init__.... Types without typing.
2. It lets you say what you mean directly with a declaration rather than expressing it in a roundabout imperative recipe. Instead of “I have a type, it’s called MyType, it has a constructor, in the constructor I assign the property ‘A’ to the parameter ‘A’ (and so on)”, you say “I have a type, it’s called MyType, it has an attribute called a”, and behavior is derived from that fact, rather than having to later guess about the fact by reverse engineering it from behavior (for example, running dir on an instance, or looking at self.__class__.__dict__).
3. It provides useful default behavior, as opposed to Python’s sometimes-useful but often-backwards defaults.
4. It adds a place for you to put a more rigorous implementation later, while starting out simple.

Let’s explore that last point.

Progressive Enhancement

While I’m not going to talk about every feature, I’d be remiss if I didn’t mention a few of them. As you can see from those mile-long repr()s for Attribute above, there are a number of interesting ones.

For example: you can validate attributes when they are passed into an @attr.s-ified class. Our Point3D, for example, should probably contain numbers. For simplicity’s sake, we could say that that means instances of float, like so:

1
2
3
4
5
6
7

import attr
from attr.validators import instance_of
@attr.s
class Point3D(object):
    x = attr.ib(validator=instance_of(float))
    y = attr.ib(validator=instance_of(float))
    z = attr.ib(validator=instance_of(float))

The fact that we were using attrs means we have a place to put this extra validation: we can just add type information to each attribute as we need it. Some of these facilities let us avoid other common mistakes. For example, this is a popular “spot the bug” Python interview question:

1
2
3
4
5
6
7

class Bag:
    def __init__(self, contents=[]):
        self._contents = contents
    def add(self, something):
        self._contents.append(something)
    def get(self):
        return self._contents[:]

Fixing it, of course, becomes this:

1
2
3
4
5

class Bag:
    def __init__(self, contents=None):
        if contents is None:
            contents = []
        self._contents = contents

adding two extra lines of code.

contents inadvertently becomes a global varible here, making all Bag objects not provided with a different list share the same list. With attrs this instead becomes:

1
2
3
4
5
6
7

@attr.s
class Bag:
    _contents = attr.ib(default=attr.Factory(list))
    def add(self, something):
        self._contents.append(something)
    def get(self):
        return self._contents[:]

There are several other features that attrs provides you with opportunities to make your classes both more convenient and more correct. Another great example? If you want to be strict about extraneous attributes on your objects (or more memory-efficient on CPython), you can just pass slots=True at the class level - e.g. @attr.s(slots=True) - to automatically turn your existing attrs declarations a matching __slots__ attribute. All of these handy features allow you to make better and more powerful use of your attr.ib() declarations.

The Python Of The Future

Some people are excited about eventually being able to program in Python 3 everywhere. What I’m looking forward to is being able to program in Python-with-attrs everywhere. It exerts a subtle, but positive, design influence in all the codebases I’ve seen it used in.

Give it a try: you may find yourself surprised at places where you’ll now use a tidily explained class, where previously you might have used a sparsely-documented tuple, list, or a dict, and endure the occasional confusion from co-maintainers. Now that it’s so easy to have structured types that clearly point in the direction of their purpose (in their __repr__, in their __doc__, or even just in the names of their attributes), you might find you’ll use a lot more of them. Your code will be better for it; I know mine has been.

Remember that thing I said in my pycon talk about native packaging being the main thing to worry about, and single-file binaries being at best a stepping stone to that and at worst a bit of a red herring? You don’t have to take it from me. From the authors of a widely-distributed command-line application that was rewritten from Python into Go specifically for easier distribution, and then rewritten in Python:

... [the] majority of people prefer native packages so distributing precompiled binaries wasn’t a big win for this type of project¹ ...

I don’t want to say “I told you so”, but... no. Wait a second. That is exactly what I want to do. That is what I am doing.

I told you so.

Hello lazyweb,

I want to run some “legacy” software (Trac, specifically) on a Swarm cluster. The files that it needs to store are mostly effectively write-once (it’s the attachments database) but may need to be deleted (spammers and purveyors of malware occasionally try to upload things for spamming or C&C) so while mutability is necessary, there’s a very low risk of any write contention.

I can’t use a networked filesystem, or any volume drivers, so no easy-mode solutions. Basically I want to be able to deploy this on any swarm cluster, so no cheating and fiddling with the host.

Is there any software that I can easily run as a daemon that runs in the background, synchronizing the directory of a data volume between N containers where N is the number of hosts in my cluster?

I found this but it strikes me as ... somehow wrong ... to use that as a critical part of my data-center infrastructure. Maybe it would actually be a good start? But in addition to not being really designed for this task, it’s also not open source, which makes me a little nervous. This list, or even this smaller one is huge and bewildering. So I was hoping you could drop me a line if you’ve got an idea what I could use for this.

I like keeping a comprehensive an accurate addressbook that includes all past email addresses for my contacts, including those which are no longer valid. I do this because I want to be able to see conversations stretching back over the years as originating from that person.

Unfortunately this causes problems when sending mail sometimes. On macOS, at least as of El Capitan, neither the Mail application nor the Contacts application have any mechanism for indicating preference-order of email addresses that I’ve been able to find. Compounding this annoyance, when completing a recipient’s address based on their name, it displays all email addresses for a contact without showing their label, which means even if I label one “preferred” or “USE THIS ONE NOW”, or “zzz don’t use this hasn’t worked since 2005”, I can’t tell when I’m sending a message.

But it seems as though it defaults to sending messages to the most recent outgoing address for that contact that it can see in an email. For people I send email to regularly to this is easy enough. For people who I’m aware have changed their email address, but where I don’t actually want to send them a message, I think I figured out a little hack that makes it work: make a new folder called “Preferred Addresses Hack” (or something suitably), compose a new message addressed to the correct address, then drag the message out of drafts into the folder; since it has a recent date and is addressed to the right person, Mail.app will index it and auto-complete the correct address in the future.

However, since the previous behavior appeared somewhat non-deterministic, I might be tricking myself into believing that this hack worked. If you can confirm it, I’d appreciate it if you would let me know.