#2594 closed Bug (wontfix)
Template system should handle whitespace better
| Reported by: | Owned by: | jshedd | |
|---|---|---|---|
| Component: | Template system | Version: | dev |
| Severity: | Normal | Keywords: | |
| Cc: | mccutchen@…, gary.wilson@…, cgrady@…, Tai Lee, kmike84@…, andrep, Stephane "Twidi" Angel, Florian.Sening@…, Waldir Pimenta, ldiqual@…, python@…, ua_django_bugzilla@…, s.shanabrook@…, chipx86@…, tomasbabej@…, trbs@…, adehnert | Triage Stage: | Accepted |
| Has patch: | yes | Needs documentation: | no |
| Needs tests: | no | Patch needs improvement: | yes |
| Easy pickings: | no | UI/UX: | no |
Description
So, I think that the template system should not leave blank lines
behind if the tag(s) on that line evaulated to the empty string. To
make sure we are all on the same page, here is an example:
This template code:
<h1>My list</h1>
<ul>
{% for item in items %}
<li>{{ item }}</li>
{% endfor %}
</ul>
Would normally evaluate to:
<h1>My list</h1>
<ul>
<li>item 1</li>
<li>item 2</li>
<li>item 3</li>
</ul>
And with the {% spaceless %} tag evaluates to:
<h1>My list</h1> <ul> <li>item 1</li> <li>item 2</li> <li>item 3</li> </ul>
And with the StripWhitespaceMiddleware evaluates to
<h1>My list</h1>
<ul>
<li>item 1</li>
<li>item 2</li>
<li>item 3</li>
</ul>
But what I think should evaluate to:
<h1>My list</h1>
<ul>
<li>item 1</li>
<li>item 2</li>
<li>item 3</li>
</ul>
In other words, leave lines that I add, but remove the empty lines that
had template tags on them.
Attachments (13)
Change History (74)
comment:1 Changed 17 years ago by
Changed 17 years ago by
| Attachment: | template_whitespace-2006-08-22.diff added |
|---|
comment:2 Changed 17 years ago by
| Cc: | mccutchen@… added |
|---|---|
| Summary: | template system should handle whitespace better → [patch] template system should handle whitespace better |
(just indicating that this ticket has a patch)
comment:3 Changed 17 years ago by
| Cc: | gary.wilson@… added |
|---|
comment:4 Changed 17 years ago by
Although, this patch does not work as desired in all cases. See the google group link above.
comment:5 Changed 17 years ago by
| Summary: | [patch] template system should handle whitespace better → Template system should handle whitespace better |
|---|
The speed issue is the killer here. We can't have a 100% performance hit on template rendering times -- speed in that area is one of Django's selling points. I'm going to remove the "patch" keyword for now, Gary, since I don't think we can take this patch. Leaving the ticket open, though, since the idea is a good one and if we can fix it, we should.
I'll try to work up something a bit faster; I had something that was almost right a while back, how hard can it be? :-(
comment:6 Changed 17 years ago by
Here are some performance measures I got when I tested a few weeks ago:
Using the command: $ python timeit.py -n 50 -s "from test.views import my_render" "my_render()" without patch (r3735): 50 loops, best of 3: 6.34 msec per loop 50 loops, best of 3: 6.23 msec per loop 50 loops, best of 3: 6.24 msec per loop with patch: 50 loops, best of 3: 8.67 msec per loop 50 loops, best of 3: 8.46 msec per loop 50 loops, best of 3: 8.46 msec per loop
So about 33-39% slower. And it doesn't work in all cases. :(
The test code looked like:
from django.template import Context, Template def my_render(): s = """\ <html> <body> <div id="content"> {% if items %} <ul> <li>Single item:</li> <li> {{ item.0.name }} </li> {% for item in items %} {% if item.name %} <li>{{ item.name }}</li> {% endif %} {% if not item.name %} <li>No name item</li> {% endif %} {% for item in fooitems %} <li>{{ item.name }}</li> {% endfor %} {% if item.des %} <li>{{ item.des }}</li> {% endif %} {% endfor %} {% if items %} <li>Spaceless items to follow:</li> {% spaceless %} {% for item in items %} <li>{{ item.name }}</li> {% endfor %} {% endspaceless %} {% endif %} </ul> {% endif %} {% if items %} <ul> {% for item in items %} {{ item.name }} {% endfor %} {% for foo in foos %} {{ foo }} {% endfor %} {% for item in items %} <li>huh{{ foo }}</li> {% endfor %} </ul> {% endif %} {% if items %}<p>We have items.</p>{% endif %} {% if items %} <p>We have items.</p> {% endif %} {% if foo %}<p>We have foo.</p>{% endif %} {% if foo %} <p>We have foo.</p> {% endif %} </div> </body> </html> """ class Item: pass items = [] for x in range(4): item = Item() item.name = 'item' + str(x) items.append(item) t = Template(s) c = Context({'items': items}) return t.render(c)
comment:7 Changed 17 years ago by
I've had a long ongoing discussion with Gary and came up with a completely different version to solve the same problem (actually, a couple of different versions, but this one is the best).
We did some profiling of this one too, and it turned out to be about 50% slower. If someone wants to have a look at optimization, I'd be most appreciative.
comment:8 Changed 17 years ago by
I'm happy to talk about my patch with anyone interested - smileychris at gmail.
How it works:
- When a tag is found on its own line, the lexer adds a token before and after containing the relevant whitespace to hold on to.
- The parser turns these into nodes.
- The nodelist works out at render time whether or not the whitespace nodes should be used.
- If an EndNode is reached before a StartNode is found, the EndNode is discarded.
- If an EndNode is reached and the previous rendered block ended in a new line, both nodes are discarded.
- Otherwise, the whitespace nodes are rendered.
comment:9 Changed 17 years ago by
| Owner: | changed from Adrian Holovaty to Malcolm Tredinnick |
|---|
Changed 17 years ago by
| Attachment: | template_whitespace_2.patch added |
|---|
Optimized version of my alternate patch
comment:10 Changed 17 years ago by
This one doesn't cause anywhere near as huge of a performance hit as the last one. And it's patched versus some recent changes.
comment:11 Changed 17 years ago by
| Has patch: | set |
|---|---|
| Patch needs improvement: | set |
| Triage Stage: | Unreviewed → Design decision needed |
comment:13 Changed 16 years ago by
| Triage Stage: | Design decision needed → Someday/Maybe |
|---|
I like the idea, but there's no way we can make the template engine 50% slower for this.
comment:14 Changed 16 years ago by
| Cc: | korpios@… added |
|---|
comment:15 Changed 15 years ago by
| Cc: | korpios@… removed |
|---|
comment:16 Changed 14 years ago by
| Owner: | changed from nobody to jshedd |
|---|
Changed 14 years ago by
| Attachment: | template_whitespace_ticket_2594.diff added |
|---|
Causes template tag blocks to collapse the whitespace around them when they are on their own line.
Changed 14 years ago by
| Attachment: | template_whitespace_ticket_2594_tests.diff added |
|---|
Tests for whitespace collapsing.
Changed 14 years ago by
| Attachment: | template_whitespace_ticket_2594_with_tests.diff added |
|---|
Everything in one diff.
Changed 14 years ago by
| Attachment: | whitespace_benchmarks.zip added |
|---|
Benchmarks from unladen-swallow for the template system.
comment:17 Changed 14 years ago by
| Patch needs improvement: | unset |
|---|
For the patch, I ran a copy of unladen-swallow's django template benchmark, attached as a zip to the ticket.
| Run | Django 1.1.1 | Django Trunk | Django Trunk with patch |
| 1 | 1.0504629612 | 1.20866298676 | 1.20275402069 |
| 2 | 1.07058095932 | 1.20971012115 | 1.20344805717 |
| 3 | 1.1476020813 | 1.2085249424 | 1.20287799835 |
| 4 | 1.06535291672 | 1.21073198318 | 1.20276403427 |
| 5 | 1.05421996117 | 1.21186900139 | 1.20313000679 |
Best I can tell, my patch is actually slightly faster, so yay!
comment:18 Changed 14 years ago by
As requested by Alex_Gaynor, running Gary Wilson's benchmark:
python -m timeit -n 50 -s "from views import my_render" "my_render()" Django 1.1.1: 50 loops, best of 3: 3.64 msec per loop 50 loops, best of 3: 3.66 msec per loop 50 loops, best of 3: 3.65 msec per loop Trunk, without patch: 50 loops, best of 3: 4.12 msec per loop 50 loops, best of 3: 4.11 msec per loop 50 loops, best of 3: 4.14 msec per loop Trunk, with patch: 50 loops, best of 3: 4.13 msec per loop 50 loops, best of 3: 4.11 msec per loop 50 loops, best of 3: 4.13 msec per loop
Script looked like:
from django.template import Context, Template from django.conf import settings settings.configure() def my_render(): s = """\ <html> <body> <div id="content"> {% if items %} <ul> <li>Single item:</li> <li> {{ item.0.name }} </li> {% for item in items %} {% if item.name %} <li>{{ item.name }}</li> {% endif %} {% if not item.name %} <li>No name item</li> {% endif %} {% for item in fooitems %} <li>{{ item.name }}</li> {% endfor %} {% if item.des %} <li>{{ item.des }}</li> {% endif %} {% endfor %} {% if items %} <li>Spaceless items to follow:</li> {% spaceless %} {% for item in items %} <li>{{ item.name }}</li> {% endfor %} {% endspaceless %} {% endif %} </ul> {% endif %} {% if items %} <ul> {% for item in items %} {{ item.name }} {% endfor %} {% for foo in foos %} {{ foo }} {% endfor %} {% for item in items %} <li>huh{{ foo }}</li> {% endfor %} </ul> {% endif %} {% if items %}<p>We have items.</p>{% endif %} {% if items %} <p>We have items.</p> {% endif %} {% if foo %}<p>We have foo.</p>{% endif %} {% if foo %} <p>We have foo.</p> {% endif %} </div> </body> </html> """ class Item: pass items = [] for x in range(4): item = Item() item.name = 'item' + str(x) items.append(item) t = Template(s) c = Context({'items': items}) return t.render(c)
comment:19 Changed 14 years ago by
| Triage Stage: | Someday/Maybe → Unreviewed |
|---|
comment:20 Changed 14 years ago by
| Triage Stage: | Unreviewed → Accepted |
|---|
The idea has already been notionally accepted by Jacob - patch now needs review and triage.
comment:21 Changed 13 years ago by
| Needs documentation: | set |
|---|
On first inspection it looks reasonably good. The performance stats are promising, too.
I can see a couple of minor areas for improvement:
- Trailing whitespace - a lot of the diff is actually just your editor adding spaces to the end of lines.
- Edge cases in tests - futzing with regular expressions makes me nervous, so I'd like to see a lot more tests to verify that nothing weird is going on. In particular, the test cases are fairly minimal - most of them start with the newline that is going to be stripped. I'd be interested in seeing more verification that if there is text/newlines *before* the newline that is to be stripped, that text won't be affected adversely.
- Documentation. This behaviour needs to be documented -- nothing major; a quick note in the description of the template language should do.
The other thing holding up this ticket is the start of feature development for 1.3. We can't apply this until be branch for 1.2 support.
comment:22 Changed 13 years ago by
| Patch needs improvement: | set |
|---|
comment:23 Changed 13 years ago by
Could you just do something like this?
t = Template(s)
c = Context({'items': items})
result = t.render(c)
return os.linesep.join([s for s in result.splitlines() if s.strip()])
comment:24 Changed 13 years ago by
I wrote some more tests for this and it turned out that the presence of a tag would strip out all whitespace occurring before it, not just one line. I don't know if that was intentional, but I thought it surprising.
Another thing I noticed is that trailing whitespace is not stripped. I don't know how relevant that is, but I mimiced that behavior in this patch.
BEFORE PATCH:
kde-devel@bishop:~/Random/djtbenchmarks$ python -m timeit -n 50 -s "from bm import my_render" "my_render()"
50 loops, best of 3: 3.29 msec per loop
AFTER PATCH:
kde-devel@bishop:~/Random/djtbenchmarks$ python -m timeit -n 50 -s "from bm import my_render" "my_render()"
50 loops, best of 3: 3.51 msec per loop
comment:25 Changed 13 years ago by
| Cc: | Tai Lee added |
|---|
comment:26 Changed 13 years ago by
The patch attached brings in jshedd's work and steveire's tests. I updated to trunk, removed trailing white space, and tweaked regex slightly.
Pre regex tweaking had failures on templatetag-whitespace numbers 14-17, 21-22, 25-30, 32, 34.
Tweaked regex fails now on 26-30, 32, 34.
I plan on breaking out the regular expression with comments to better understand it, and try to fix the remaining failures.
Changed 13 years ago by
Patched against trunk (and template shuffling), removed trailing white space, and adjusted to add comments and make pass with steveire's tests.
comment:27 Changed 13 years ago by
My local testing shows:
python -m timeit -n 50 -s "from views import my_render" "my_render()" Django 1.2.3 50 loops, best of 3: 3.75 msec per loop 50 loops, best of 3: 3.74 msec per loop 50 loops, best of 3: 3.76 msec per loop Django trunk without patch 50 loops, best of 3: 3.72 msec per loop 50 loops, best of 3: 3.73 msec per loop 50 loops, best of 3: 3.73 msec per loop Django trunk with patch 50 loops, best of 3: 3.73 msec per loop 50 loops, best of 3: 3.73 msec per loop 50 loops, best of 3: 3.72 msec per loop
comment:28 Changed 13 years ago by
I seriously doubt anybody has used vertical tabs (\v) or form feeds (\f) in Django templates ever. There's no need to put them in the reg-exp.
Also, realise that this is backwards incompatible and not in a trivial way. Anybody using templates to generate output in a format where whitespace is always significant (plain text email is one very big case) would have their output broken by this (unclear if Jacob and Russell both thought this was an acceptable sacrifice -- I definitely don't -- or if it was overlooked). It has to be optional and off by default in the rendering process. (The other case where I can immediately think of this breaking something is Django's debug page's "paste traceback" button -- I remember Erik Karulf spending ages to make sure the whitespace in what we POST'ed to dpaste was sensible.
comment:29 Changed 13 years ago by
I agree with Malcolm that outright changing this is a backwards-compatibility issue. Unfortunately.
On the other hand, I think the new behavior is clearly superior in pretty much every case, and a major win, so I think it's worth finding a way to do it via an optional flag (an argument to Template.init? Would then need to be added to several shortcuts as well...), and then consider (later?) possible deprecation paths to migrate towards making it the default.
comment:30 Changed 13 years ago by
I'd like to see this enabled with a template tag, as it is template authors who will be needing to use it most. I'd also like to see a deprecation path so that this can become the default behaviour.
Similar to when autoescaping was introduced, we could add a template tag {% stripwhitespace on|off %}{% endstripwhitespace %} that people can put in their base templates to preserve the existing behaviour when the default changes or enable the new behaviour before it becomes the default. We could raise a deprecation warning whenever a template is rendered without this tag.
comment:31 Changed 13 years ago by
I'm not very familiar with the behavior of the blocktrans tag, but I guess there would need to be some handling of the whitespace at the end of a msgid being present in the po file, but not in the compiled template. I don't know if gettext is that strict. I guess the message extractor script should handle template syntax on its own line in the same way that the latest patch here does.
If the django message extractor script already uses the same classes, then there may be no problem, but it's something to look into.
This is related to #5849.
Changed 13 years ago by
| Attachment: | whitespace.patch added |
|---|
comment:32 Changed 13 years ago by
| Needs documentation: | unset |
|---|---|
| Patch needs improvement: | unset |
Changed 13 years ago by
| Attachment: | 0001-Commiting-whitespace-patch-from-django-ticket-2594.patch added |
|---|
Modifies DebugLexer too, not just Lexer
Changed 13 years ago by
| Attachment: | 0001-Commiting-whitespace-patch-from-django-ticket-2594.2.patch added |
|---|
Same as previous, with fixed whitespace issue
comment:33 Changed 12 years ago by
| Type: | defect → Bug |
|---|
comment:34 Changed 12 years ago by
| Severity: | normal → Normal |
|---|
comment:35 Changed 12 years ago by
| Easy pickings: | unset |
|---|---|
| Patch needs improvement: | set |
0001-Commiting-whitespace-patch-from-django-ticket-2594.2.patch fails to apply cleanly on to trunk
comment:37 Changed 12 years ago by
I just started a thread about an alternative solution to this problem.
Newline stripping in templates: the dnl way
http://groups.google.com/group/django-developers/browse_thread/thread/8d6b6a3aae82ee03
comment:38 Changed 12 years ago by
| Cc: | kmike84@… added |
|---|
comment:39 Changed 11 years ago by
| Cc: | andrep added |
|---|
comment:40 Changed 11 years ago by
| Cc: | Stephane "Twidi" Angel added |
|---|
comment:41 Changed 11 years ago by
Please someone from django core sacrifice a day, and test and ad this to 1.4.x. Thanks a lot!
comment:42 Changed 11 years ago by
| Cc: | Florian.Sening@… added |
|---|
comment:43 Changed 11 years ago by
| Cc: | Waldir Pimenta added |
|---|
comment:45 Changed 11 years ago by
| Cc: | python@… added |
|---|
Changed 11 years ago by
| Attachment: | ticket-2594-template-strip-leading-whitespace-2012-06-28.diff added |
|---|
Update patch against 1.5-dev
comment:46 Changed 11 years ago by
I've spent the morning updating the patch against the latest development version.
It was fairly straightforward, except for the Lexer.create_token() method, which took some time. It was already fairly hairy in 1.4, and the addition of the verbatim tag in 1.5-dev had made it even more so. I've added comments to make it easier for the next person to unravel the logic.
All tests pass, pre- and post-patch, and the results of Gary Wilson's benchmark remain good.
python -m timeit -n500 -s 'from bm import my_render' 'my_render()' Without patch: 500 loops, best of 3: 1.96 msec per loop 500 loops, best of 3: 1.95 msec per loop 500 loops, best of 3: 1.96 msec per loop With patch, TEMPLATE_STRIP_LEADING_WHITESPACE = True: 500 loops, best of 3: 1.96 msec per loop 500 loops, best of 3: 1.96 msec per loop 500 loops, best of 3: 1.97 msec per loop With patch, TEMPLATE_STRIP_LEADING_WHITESPACE = False: 500 loops, best of 3: 1.99 msec per loop 500 loops, best of 3: 2 msec per loop 500 loops, best of 3: 2.01 msec per loop
comment:47 Changed 11 years ago by
| Cc: | ua_django_bugzilla@… added |
|---|
comment:48 Changed 11 years ago by
| Cc: | s.shanabrook@… added |
|---|
What needs to be done to get this merged still?
comment:49 Changed 11 years ago by
You could try reading through the ticket and django-developers group history so that you're familiar with all the options and opinions that have been presented. Then outline a proposal on the django-developers group to move the ticket forward by choosing a preferred option (probably the one which had the most support), summarise why it has been chosen over any others, update any patches so that they apply cleanly and ask for a code review and any final objections that would delay this being committed.
comment:50 Changed 11 years ago by
Why not implement the {%- tag -%} support in Jinja2, and let people choose themselves?
comment:51 Changed 10 years ago by
| Cc: | chipx86@… added |
|---|
comment:52 Changed 10 years ago by
| Cc: | tomasbabej@… added |
|---|
comment:53 Changed 10 years ago by
There's no way I'd want to have _only_ a global flag to enable this behaviour. Certainly, there are cases where I'd rather avoid leading spaces and blank lines, but equally there are cases where it's important.
Not all the world is HTML. Django templates are frequently used for content types where the leading whitespace may be important, in the same projects where it's used for HTML, where it's less important.
Typically to avoid these problems I've started using constructs like:
<h1>The title</h1>
<ul>{% for item in somelist %}
<li>{{ item }}</li>{% endfor %}
</ul>
Perhaps not as obvious with the template tags, but clearer with the HTML.
comment:54 Changed 10 years ago by
Just wanted to note that this is exactly the problem I'm facing. Not everything is HTML and sometimes whitespace matters. The templates for these cases are very hard to read. :/
comment:55 Changed 10 years ago by
| Cc: | trbs@… added |
|---|
comment:56 Changed 9 years ago by
| Resolution: | → wontfix |
|---|---|
| Status: | new → closed |
As far as I know, the consensus among the core team is that:
- the Django template language is good enough for generating HTML, which isn't sensitive to whitespace;
- the long term plan is to replace it with a better engine (most likely Jinja), not to keep working on it;
- if you have more specialized needs (want to generate RTF?) just use another template engine, there are several to choose from.
Hence, closing wontfix. Thanks everyone who tried contributing to this ticket, and sorry it didn't happen.
comment:57 Changed 9 years ago by
Is there an open ticket for tracking the progress of integrating Django's newly blessed template engine (Jinja)? Or any indication on how "long term" that plan is? Was there a google groups discussion about this long term plan?
I did a quick search and couldn't find either. That being the case, it seems that if no new improvements will be made to Django's template engine because it might be replaced in the next 5-10 years (seems like a big job), that's simply leaving everyone who finds the current engine lacking in any way, stuck in limbo indefinitely.
Is the existing template engine in permanent feature freeze now? Only bug fixes allowed?
comment:58 Changed 9 years ago by
Let me restate this with more plainly: we're not interested in this feature and we don't provide an alternative.
comment:59 Changed 9 years ago by
I think that what @aaugustin is trying to say is:
Google.com/search?q=how+do+i+use+jinja2+with+django
There, another bug fixed; right?
comment:60 Changed 8 years ago by
To achieve this from jinja2, pass both trim_blocks=True and lstrip_blocks=True when you instantiate either your Environment or Template.
see: http://jinja.pocoo.org/docs/dev/templates/#whitespace-control
comment:61 Changed 11 months ago by
| Cc: | adehnert added |
|---|
the thread:
http://groups.google.com/group/django-developers/browse_thread/thread/137eac5d3c9ad0fb/#