Python 3.5, bytes, and %-interpolation (aka PEP 461)

[Ethan Furman]

Greetings!

A PEP is under discussion to add %-interpolation back to the bytes type in Python 3.5.

Assuming the PEP is accepted, what *will* be added back is:

Numerics:

b'%d' % 10 --> b'10'
b'%02x' % 10 --> b'0a'

Single byte:

b'%c' % 80 --> b'P'

and generic:

b'%s' % some_binary_blob --> b'tHE*&92h4' (or whatever)

What is under debate is whether we should also add %a:

b'%a' % some_obj --> b'some_obj_repr'

What %a would do:

get the repr of some_obj

convert it to ascii using backslashreplace (to handle any code points over 127)

encode to bytes using 'ascii'

Can anybody think of a use-case for this particular feature?

 

[Marko Rauhamaa]

Can anybody think of a use-case for this particular feature?

Internet protocol entities constantly have to format (and parse)
ASCII-esque octet strings:

headers.append(b'Content-length: %d\r\n' % len(blob))

headers.append(b'Content-length: {}\r\n'.format(len(blob)))

Now you must do:

headers.append(('Content-length: %d\r\n' % len(blob)).encode())

headers.append('Content-length: {}\r\n'.format(len(blob)).encode())

That is:

1. ineffient (encode/decode shuffle)

2. unnatural (strings usually have no place in protocols)

3. confusing (what is stored as bytes, what is stored as strings?)

4. error-prone (UTF-8 decoding exceptions etc)


To be sure, %s will definitely be needed as well:

uri = b'http://%s/robots.txt' % host


Marko

 

[random832]

That is:

1. ineffient (encode/decode shuffle)

2. unnatural (strings usually have no place in protocols)

That's not at all clear. Why _aren't_ these protocols considered text
protocols? Why can't you add a string directly to headers?

 

[Ethan Furman]

That's not at all clear. Why _aren't_ these protocols considered text
protocols? Why can't you add a string directly to headers?

Because text is a high-order abstraction. You don't store text in files, you don't transmit text over the wire or
through the air -- those actions are done with a lower abstraction, that of bytes.

You're framework may allow you to add a string, but under the covers it's converting to bytes -- at which point is up to
the framework.

 

[Marko Rauhamaa]

(e-mail address removed):

That's not at all clear. Why _aren't_ these protocols considered text
protocols? Why can't you add a string directly to headers?

Text expresses a written human language. In prosaic terms, a Python
string is a sequence of ISO 10646 characters, whose codepoints are not
octets.

Most network protocols are defined in terms of octets, although many of
them can carry textual, audio or video payloads (among others). So when
RFC 3507 (ICAP) shows an example starting:

RESPMOD icap://icap.example.org/satisf ICAP/1.0
Host: icap.example.org
Encapsulated: req-hdr=0, res-hdr=137, res-body=296

it consists of 8-bit octets and not some human language.

In practical terms, you get the bytes off the socket as, well, bytes. It
makes little sense to "decode" those bytes into a string for
manipulation. Manipulating bytes directly is both more efficient and
more natural from the point of view of the standard.

Many internet protocols happen to look like text. It makes it nicer for
human network programmers to work with them. However, they are primarily
meant for computers, and the message formats are really a form of binary
code.


Marko

 

[Steven D'Aprano]

Greetings!

A PEP is under discussion to add %-interpolation back to the bytes type
in Python 3.5.

Assuming the PEP is accepted, what *will* be added back is:

Numerics:

b'%d' % 10 --> b'10'
b'%02x' % 10 --> b'0a'

Single byte:

b'%c' % 80 --> b'P'

Will %c also accept a length-1 bytes object?

b'%c' % b'x'
=> b'x'

and generic:

b'%s' % some_binary_blob --> b'tHE*&92h4' (or whatever)

Will b'%s' take any arbitrary object, as in:

b'Key: %s' % [1, 2, 3, 4]
=> b'Key: [1, 2, 3, 4]'


or only something which is already bytes (i.e. a bytes or bytearray
object)?

What is under debate is whether we should also add %a:

b'%a' % some_obj --> b'some_obj_repr'

What %a would do:

get the repr of some_obj

convert it to ascii using backslashreplace (to handle any code points
over 127)

encode to bytes using 'ascii'

Can anybody think of a use-case for this particular feature?

Not me.

 

[Steven D'Aprano]

(e-mail address removed):

You cannot mix text strings and byte strings in Python 3. Python 2 allows
you to do so, and it leads to hard-to-diagnose bugs and confusing
behaviour. This is why Python 3 insists on a strict separation between
the two.

But of course you can add *byte* strings directly to byte headers. Just
prefix your strings with a b, as in b'Header' instead of 'Header', and it
will work fine.

However, you don't really want to be adding large numbers of byte strings
together, due to efficiency. Better to use % interpolation to insert them
all at once. Hence the push to add % to bytes in Python 3.


Marko replied:

Text expresses a written human language. In prosaic terms, a Python
string is a sequence of ISO 10646 characters, whose codepoints are not
octets.

Almost correct, but not quite. Python strings are Unicode, not ISO-10646.
The two are not the same.

http://www.unicode.org/faq/unicode_iso.html

Most network protocols are defined in terms of octets, although many of
them can carry textual, audio or video payloads (among others). So when
RFC 3507 (ICAP) shows an example starting:

RESPMOD icap://icap.example.org/satisf ICAP/1.0 Host:
icap.example.org
Encapsulated: req-hdr=0, res-hdr=137, res-body=296

it consists of 8-bit octets and not some human language.

Not really relevant. In practical terms, whether they are implemented as
octets or not, the sequence "Host" *is* human language, specifically it
is the English word Host that just happens to be encoded in ASCII.
Likewise the sequence "Encapsulated" *is* the English word Encapsulated
encoded in ASCII.

In practical terms, you get the bytes off the socket as, well, bytes. It
makes little sense to "decode" those bytes into a string for
manipulation. Manipulating bytes directly is both more efficient and
more natural from the point of view of the standard.

But not necessarily more natural from the point of the programmer, which
is what matters.

I agree that if you don't need to interpret the data as Unicode text,
then there's no real benefit to decoding to text. (In fact, if your data
can contain arbitrary bytes, you may not be able to decode to text, since
not all byte sequences are legal UTF-8.)

Many internet protocols happen to look like text. It makes it nicer for
human network programmers to work with them. However, they are primarily
meant for computers, and the message formats are really a form of binary
code.

The reason that, say, the subject header line in emails starts with the
word "Subject" rather than some arbitrary binary code is because it is
intended to be human-readable. Not just human-readable, but *semantically
meaningful*. That's why the subject line is labelled "Subject" rather
than "Field 23" or "SJT".

Fortunately, such headers are usually (always?) ASCII, and byte strings
in Python privilege ASCII-encoded text. When you write b'Subject', you
get the same sequence of bytes as 'Subject'.encode('ascii').

 

[Ethan Furman]

Greetings!

A PEP is under discussion to add %-interpolation back to the bytes type
in Python 3.5.

Assuming the PEP is accepted, what *will* be added back is:

Numerics:

b'%d' % 10 --> b'10'
b'%02x' % 10 --> b'0a'

Single byte:

b'%c' % 80 --> b'P'

Will %c also accept a length-1 bytes object?

b'%c' % b'x'
=> b'x'
Yes.

and generic:

b'%s' % some_binary_blob --> b'tHE*&92h4' (or whatever)

Will b'%s' take any arbitrary object, as in:

b'Key: %s' % [1, 2, 3, 4]
=> b'Key: [1, 2, 3, 4]'
No.

or only something which is already bytes (i.e. a bytes or bytearray
object)?

It must already be bytes, or have __bytes__ method (that returns bytes, obviously .

Not me.

I find that humorous, as %a would work with your list example above.

 

[Steven D'Aprano]

On 02/24/2014 03:55 PM, Steven D'Aprano wrote:

Will b'%s' take any arbitrary object, as in:

b'Key: %s' % [1, 2, 3, 4]
=> b'Key: [1, 2, 3, 4]'

No.

Very glad to hear it.


[...]

I find that humorous, as %a would work with your list example above.

I know. But why would I want to do it? "It won't fail" is not a use-case.
I can subclass int and give it a __getitem__ method that raise SystemExit,
but that's not a use-case for doing so

I cannot think of any reason to want to ASCII-ise the repr of arbitrary
objects, and on the rare occasion that I did, I could say

repr(obj).encode('ascii', 'backslashescape')


I don't object to this feature, but nor do I want it.

 

[wxjmfauth]

Le mardi 25 février 2014 00:55:36 UTC+1, Steven D'Aprano a écrit :

However, you don't really want to be adding large numbers of byte strings

together, due to efficiency. Better to use % interpolation to insert them

all at once. Hence the push to add % to bytes in Python 3.

Interpolation will not help.

What is wrong by design will always stay wrong by design.

jmf

 

[Mark Lawrence]

What is wrong by design will always stay wrong by design.

Why are you making the statement that PEP 461 is wrong by design?