csv writer needs more quoting rules

[samwyse]

BPO	23041
Nosy	@smontanaro, @rhettinger, @bitdancer, @berkerpeksag, @serhiy-storchaka, @yoonghm, @erdnaxeli, @msetina
PRs	gh-67230: add quoting rules to csv module #29469
Files	issue23041.patch: Patch for resolving issue 23041. According to message 232560 and 232563 issue23041_test.patch: Test cases added for testing as behaviour proposed in message 232560

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = 'https://github.com/smontanaro'
closed_at = None
created_at = <Date 2014-12-12.16:36:45.576>
labels = ['easy', 'type-feature', 'library', '3.11']
title = 'csv needs more quoting rules'
updated_at = <Date 2022-02-24.00:07:17.586>
user = 'https://bugs.python.org/samwyse'

bugs.python.org fields:

activity = <Date 2022-02-24.00:07:17.586>
actor = 'samwyse'
assignee = 'skip.montanaro'
closed = False
closed_date = None
closer = None
components = ['Library (Lib)']
creation = <Date 2014-12-12.16:36:45.576>
creator = 'samwyse'
dependencies = []
files = ['37444', '37445']
hgrepos = []
issue_num = 23041
keywords = ['patch', 'easy']
message_count = 26.0
messages = ['232560', '232561', '232563', '232596', '232607', '232630', '232676', '232677', '232681', '261141', '261146', '261147', '341460', '358461', '396621', '396641', '396642', '396643', '401607', '401608', '405951', '406013', '406015', '406017', '412463', '413867']
nosy_count = 11.0
nosy_names = ['skip.montanaro', 'rhettinger', 'samwyse', 'r.david.murray', 'berker.peksag', 'serhiy.storchaka', 'krypten', 'yoonghm', 'tegdev', 'erdnaxeli', 'msetina']
pr_nums = ['29469']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue23041'
versions = ['Python 3.11']

Linked PRs

• gh-67230: document new csv quoting modes in whatsnew (gh-103491) #103491
• gh-67230: Update Misc/whatsnew/3.12.rst for csv module changes #103598
• gh-67230: Add versionadded notes for QUOTE_NOTNULL and QUOTE_STRINGS #114816
• [3.12] gh-67230: Add versionadded notes for QUOTE_NOTNULL and QUOTE_STRINGS (GH-114816) #114840

[samwyse]

The csv module currently implements four quoting rules for dialects: QUOTE_MINIMAL, QUOTE_ALL, QUOTE_NONNUMERIC and QUOTE_NONE. These rules treat values of None the same as an empty string, i.e. by outputting two consecutive quotes. I propose the addition of two new rules, QUOTE_NOTNULL and QUOTE_STRINGS. The former behaves like QUOTE_ALL while the later behaves like QUOTE_NONNUMERIC, except that in both cases values of None are output as an empty field. Examples follow.

Current behavior (which will remain unchanged)

>>> csv.register_dialect('quote_all', quoting=csv.QUOTE_ALL)
>>> csv.writer(sys.stdout, dialect='quote_all').writerow(['foo', None, 42])
"foo","","42"

>>> csv.register_dialect('quote_nonnumeric', quoting=csv.QUOTE_NONNUMERIC)
>>> csv.writer(sys.stdout, dialect='quote_nonnumeric').writerow(['foo', None, 42])
"foo","",42

Proposed behavior

>>> csv.register_dialect('quote_notnull', quoting=csv.QUOTE_NOTNULL)
>>> csv.writer(sys.stdout, dialect='quote_notnull').writerow(['foo', None, 42])
"foo",,"42"

>>> csv.register_dialect('quote_strings', quoting=csv.QUOTE_STRINGS)
>>> csv.writer(sys.stdout, dialect='quote_strings').writerow(['foo', None, 42])
"foo",,42

[bitdancer]

As an enhancement, this could be added only to 3.5. The proposal sounds reasonable to me.

[samwyse]

David: That's not a problem for me.

Sorry I can't provide real patches, but I'm not in a position to compile (much less test) the C implementation of _csv. I've looked at the code online and below are the changes that I think need to be made. My use cases don't require special handing when reading empty fields, so the only changes I've made are to the code for writers. I did verify that the reader code mostly only checks for QUOTE_NOTNULL when parsing. This means that completely empty fields will continue to load as zero-length strings, not None. I won't stand in the way of anyone wanting to "fix" that for these new rules.

typedef enum {
QUOTE_MINIMAL, QUOTE_ALL, QUOTE_NONNUMERIC, QUOTE_NONE,
QUOTE_STRINGS, QUOTE_NOTNULL
} QuoteStyle;

static StyleDesc quote_styles[] = {
{ QUOTE_MINIMAL, "QUOTE_MINIMAL" },
{ QUOTE_ALL, "QUOTE_ALL" },
{ QUOTE_NONNUMERIC, "QUOTE_NONNUMERIC" },
{ QUOTE_NONE, "QUOTE_NONE" },
{ QUOTE_STRINGS, "QUOTE_STRINGS" },
{ QUOTE_NOTNULL, "QUOTE_NOTNULL" },
{ 0 }
};

        switch (dialect->quoting) {
        case QUOTE_NONNUMERIC:
            quoted = !PyNumber_Check(field);
            break;
        case QUOTE_ALL:
            quoted = 1;
            break;
        case QUOTE_STRINGS:
            quoted = PyString_Check(field);
            break;
        case QUOTE_NOTNULL:
            quoted = field != Py_None;
            break;
        default:
            quoted = 0;
            break;
        }

" csv.QUOTE_MINIMAL means only when required, for example, when a\n"
" field contains either the quotechar or the delimiter\n"
" csv.QUOTE_ALL means that quotes are always placed around fields.\n"
" csv.QUOTE_NONNUMERIC means that quotes are always placed around\n"
" fields which do not parse as integers or floating point\n"
" numbers.\n"
" csv.QUOTE_STRINGS means that quotes are always placed around\n"
" fields which are strings. Note that the Python value None\n"
" is not a string.\n"
" csv.QUOTE_NOTNULL means that quotes are only placed around fields\n"
" that are not the Python value None.\n"

[rhettinger]

Samwyse, are these suggestions just based on ideas of what could be done or have you encountered real-world CSV data exchanges that couldn't be handled by the CSV module?

[smontanaro]

It doesn't look like a difficult change, but is it really needed? I guess my reaction is the same as Raymond's. Are there real-world uses where the current set of quoting styles isn't sufficient?

[krypten]

Used function PyUnicode_Check instead of PyString_Check

[samwyse]

Yes, it's based on a real-world need. I work for a Fortune 500 company and we have an internal tool that exports CSV files using what I've described as the QUOTE_NOTNULL rules. I need to create similar files for re-importation. Right now, I have to post-process the output of my Python program to get it right. I added in the QUOTE_STRINGS rule for completeness. I think these two new rules would be useful for anyone wanting to create sparse CSV files.

[smontanaro]

If I understand correctly, your software needs to distinguish between

# wrote ["foo", "", 42, None] with quote_all in effect
"foo","","42",""

and

# wrote ["foo", None, 42, ""] with quote_nonnull in effect
"foo",,"42",""

so you in effect want to transmit some type information through a CSV file?

[samwyse]

Skip, I don't have any visibility into how the Java program I'm feeding data into works, I'm just trying to replicate the csv files that it exports as accurately as possible. It has several other quirks, but I can replicate all of them using Dialects; this is the only "feature" I can't. The files I'm looking at have quoted strings and numbers, but there aren't any quoted empty strings. I'm using a DictWriter to create similar csv files, where missing keys are treated as values of None, so I'd like those printed without quotes. If we also want to print empty strings without quotes, that wouldn't impact me at all.

Besides my selfish needs, this could be useful to anyone wanting to reduce the save of csv files that have lots of empty fields, but wants to quote all non-empty values. This may be an empty set, I don't know.

[smontanaro]

Thanks for the update berker.peksag. I'm still not convinced that the csv module should be modified just so one user (sorry samwyse) can match the input format of someone's Java program. It seems a bit like trying to make the csv module type-sensitive. What happens when someone finds a csv file containing timestamps in a format other than the datetime.datetime object will produce by default? Why is None special as an object where bool(obj) is False?

I think the better course here is to either:

• subclass csv.DictWriter, use dictionaries as your element type, and have its writerow method do the application-specific work.

• define a writerow() function which does something similar (essentially wrapping csv.writerow()).

If someone else thinks this is something which belongs in Python's csv module, feel free to reopen and assign it to yourself.

[serhiy-storchaka]

The csv module is already type-sensitive (with QUOTE_NONNUMERIC). I agree, that we shouldn't modify the csv module just for one user and one program.

If a standard CVS library in Java (or other popular laguages) differentiates between empty string and null value when read from CSV, it would be a serious argument to support this in Python. Quick search don't find this.