Add Support for the Jinja2 Template Engine

We've decided some time ago to remove the legacy support for the ClearSilver template engine, for Trac 1.0 (r10570). Clearsilver had its share of inconveniences, enough that we decided to switch to the nicer ​Genshi template engine in 0.11, but to be honest ClearSilver was very fast and memory lenient. While we managed to keep Genshi memory usage somewhat in control (remember #6614?), the speed was never really adequate, especially for big changesets and for displaying source files over a few thousand lines of code (see TracDev/Performance#Genshi for details).

So one solution would be to switch once again, to a template engine that would combine the advantages of Genshi (pure Python, nice templates, flexible) and ClearSilver (speed!). Such a beast seems to exist now: ​Jinja2.

Several points remain to be clarified:

• what will be the upgrade path for plugins that came to rely on IStreamFilters?
• how to handle themeing?
• should we rewrite tag builders or use lightweight string templates?
• others?

See also ​this Trac-Dev discussion from 2010, which is still pertinent. Well, obviously we managed to release Genshi 0.6 since then, but the issue is a recurring one, see this recent (2016-01) ​Genshi question on Trac-Users.

Experimenting with Jinja2 (2.8)

Nothing like a few numbers to make a point ;-)

These are the timings for rendering r3871, with the diff options set to side-by-side, in place modifications, served by tracd on my development laptop. This generates a page weighing 11.5MB (Genshi) to 10.3MB (Jinja2) in size.

	Genshi				Jinja2
	stream		blob		generate		stream (5)		stream (10)		stream (100)		stream (1000)		blob
	1st	2nd	1st	2nd	1st	2nd	1st	2nd	1st	2nd	1st	2nd	1st	2nd	1st	2nd
TTFB	16600	15670	25530	24460	2020	1160	2030	1160	2070	1170	2150	1230	2280	1230	3370	2450
CD	16090	16050	387	1240	2820	2720	2730	2640	2730	2680	2470	2390	2350	2250	488	1060
Total	32690	31720	25917	25700	4840	3880	4760	3800	4800	3850	4620	3620	4630	3480	3850	3510
Rdr	—	—	23533	23273	—	—	—	—	—	—	—	—	—	—	1477	1263

Some explanations:

• Genshi (0.7 with speedups)
  • stream means we return content via Stream.serialize and send chunks as we have them
  • blob means we first generate all the content in memory with Stream.render, then send it at once
• Jinja2 (2.8 with ​speedups)
  • generate means we use Template.generate and send chunks as we have them
  • stream means we use the TemplateBuffer wrapper on the above, which groups a few chunks (given by the number in parenthesis) together before we send them; for a chunk size of 100, we get the best compromise: still a very low TTFB and a reduced Content download time; actually the sweet spot is probably between 10 and 100, and will most certainly depend on the actual content (I just tested 75 which gives 1160/2430 for example)
  • blob means we first generate all the content in memory with Template.render
• both:
  • 1st, is the time in ms for the first request, sent right after a server restart
  • 2nd, is the time in ms for the second request, sent just after the first (usually the 3rd and subsequent requests would show the same results as this 2nd request)

We measure:

• TTFB (Time to first byte), as given by Chrome network panel in the developer window
• CD (Content download), idem
• Rdr (template rendering time), mostly significant for the "blob" method otherwise it also takes the network latency into account

Note that even if the total "blob" time seems better than the total "stream" one, the lower TTFB is nevertheless a major benefit for the streaming variant, as this means the secondary requests can start earlier (and in this case, finish before the main request).

In addition, while I didn't measure precisely the memory usage, Genshi made the python.exe process jump from 109MB to 239MB while rendering the request (blob). The memory seems to be freed afterwards (there were no concurrent requests). By contrast, with Jinja2 the memory spike was 106MB to 126MB.

In summary, this means that for the big problematic pages, we can easily have a 10x speedup and more, by migrating to Jinja2,and this with a much lighter memory footprint. For smaller pages, the speed-up is between 5x to 10x as well.

Implementation details

Themeing

Genshi theme

I've never really tried the ​TH:ThemeEnginePlugin plugin, or alternatives, so I can't be sure if I got it right, but from what I can see in Trac's code base itself, the idea with Genshi-based themeing (and page architecture in general) was the following:

Let's take the example of a simple "end user" page, the search.html page. What happens is that:

• search.html includes
  • layout.html, which includes
    • theme.html (as this is our default theme page)

In more details, this is what happens with search.html​:

• it starts by including the layout.html page:

  <xi:include href="layout.html" />
  

  • the layout.html​ page:
    • defines <py:match> filters for transforming the content provided in <head> and <body> elements
    • then dynamically includes some "theme" page:

      <xi:include href="$chrome.theme"><xi:fallback /></xi:include>
      

      • by default, this will be our theme.html​:
        • it defines a <py:match> filter on the <body> tag (the one that will be produced by the previously applied filters, i.e. the output of the <py:match> from the layout.html) and it will embed that element into some predefined HTML structure (inside a div element), and it will prepend and append other divs around it:

          <body>
            <div id="banner">
              ...
            </div>
            <div id="main">
              ... 
              (here goes the content of the <body></body> produced by layout.html)
            </div>
            <div id="footer">
              ...
            </div>
          </body>
          

• the search.html page then provide the <head> and <body> elements specific to that search page; these elements will be processed by the <py:match>es filters defined so far in the order in which they have been included (so first the ones defined in layout.html, then the one for the body defined in theme.html)

The scheme can be extended with more intermediate levels, for example what we have for the admin panels, e.g the admin_basics.html​ panel:

• it starts by including the admin.html page:

  <xi:include href="admin.html" />
  

  • the admin.html​ page
    • defines <py:match> filters for <head> and <body>, which in turn will produce modified <head> and <body> elements
    • it then includes the layout.html page (see above)
• then it provides the <head> and <body> elements specific to that panel (they will be process by the <py:match> filters in the order seen, so first those from admin.html, then those from layout.html and finally those from the dynamically included theme page

Feel free to brush up your Genshi craft by reading (​GenshiTutorial#AddingaLayoutTemplate), as I just did ;-)

So this dynamic theme template has the last say, and can theoretically re-order the content generated by the previous filters any way it likes. In practice (at least in our default theme.html), it's just a <py:match once> template which appends a footer to the content of the <body> element produced so far.

An example of another theme page: ​trac-bootstrap-theme's theme.html.

Jinja2 theme

Fortunately Jinja2 can do dynamic includes (or extends in this case) as well, so the same idea can be transposed: have the end user page extend the layout page, then have the layout page extend whatever has been defined to be the theme page.

The differences with Genshi are subtle: while in both case the control of the output is delegated to the more generic page, with Jinja2 the parent only controls what it puts around blocks. It can put some default content in these blocks, but the end user page has the final say about what to do with this default content, as it can reuse it inside its block (by calling super()) or not.

So let's transpose the example. For as long as we have both template engines coexisting, we prefix the new Jinja2 templates with a j, so we're now discussing:

• jsearch.html, which extends
  • jlayout.html, which extends
    • jtheme.html (as this is our default theme page)

In more details, we describe what happens with the jsearch.html page:

• it starts by extending the jlayout.html page:

  # extends 'jlayout.html'
  

  • the jlayout.html page:
    • first dynamically extends in turn some "theme" page

      # extends ('j' + chrome.theme)
      

      • by default, this will be our jtheme.html page:
        • it defines a head block inside of an otherwise empty <head> element; this means this a "slot" that will be filled by the content of the head block in the extending templates (in this case, jlayout.html)
        • it contains a <body> element; as we want to replicate what the original theme.html did, what we want to achieve here is to provide a slot at the place where we want to substitute in the content produced by the jlayout.html template:

          <body>
            # block body
            <div id="banner">
              ...
            </div>
            <div id="main">
              ... 
              # block content
              (here goes the content of the content block produced by layout.html)
              # endblock content
            </div>
            <div id="footer">
              ...
            </div>
            # block body
          </body>
          

          I didn't name that inner block "body", as this could be confusing: we're not in control of the <body> element there, just of a fraction of it, the bottom part of the main div. If we want to be in complete control of the body in the extended template, we could (as opposed to what happens in Genshi), we would simply have to redefine the body block which contains all of the default structure (possibly reusing the content of that block by a call to ${ super() })
    • then the jlayout.html template defines the head and content blocks; it doesn't need to redefine the body block, as it's happy with what jtheme does with it; that content block is filled with some predefined, generic content, mostly the same stuff that could be found in the corresponding <py:match> filters in layout.html;
• the end user template jsearch.html has to redefine the head and content blocks, and has to place a ${ super() } expression in order to insert the default content proposed by the extended template at the right place

Depending how one looks at it, it seems this approach is even more flexible than what we had in Genshi, as the end user template can decide which bits of the parent template it wants or not ("bottom-up" control), something that was not readily doable with Genshi ("top-down" control).

I omitted the discussion of the title block (not even implemented yet) and the replacement for the <xi:include href="site.html> template (# include site_head.html and # include site_body.html).