Perfomance on large pdf

[alecasciaro]

Hi,

i'm trying to convert an html page that generate a 52 pages pdf.
Weasyprint needs 1.40 minute to create the the 52 page pdf.
I looked for a solution, removing image and all css (thinking that maybe some css convertion could be slow) but performance increase only 20 seconds.
Someone can give me some advice to improve the performance?
Thanks

[liZe]

WeasyPrint is known to be slow. The only general advice I can give you is to use Python 3.6 and the latest version of WeasyPrint. If you can share your HTML source, I can try to find why your document takes an especially long time to get rendered.

[alecasciaro]

@liZe this is the page http://www.placesrank.it/pdf-destination?place=6290&campaign=8
Thanks

[liZe]

@alecasciaro Thank you for this example. It takes about 2:20 minutes for me, with:

• about 20 seconds to get, parse and apply HTML/CSS,
• about 20 seconds to create the formatting structure,
• about 40 seconds to create the layout,
• about 60 seconds to create the fixed boxes and page margins layout,
• a couple of seconds to generate the PDF.

There's a big performance problem about the fixed boxes or page margins. I'll try to find where it comes from.

[alecasciaro]

Ok thank you @liZe, i realy appreciate it

[eligiobz]

@liZe Is there any advance on this issue? I'd like to know if there's something in particular to look at to help improve this

[liZe]

OK, I've found the source of the problem. More than half of the time is taken by downloading photos.

A solution to avoid that is to render PDFs on the same server than your website, and use something that gets the resources on disk rather than through network. That's what Flask-WeasyPrint or Django-WeasyPrint do for example, you can take a look on how they work and duplicate the behaviour for your own use.

[fcaldas]

@liZe Just a question do WeasyPrint download the images in parallel or serially?

[alecasciaro]

@liZe , i tried to remove all image and all css but performance increase only 20 seconds.
In your case, the server is a test environment, that why it's slow. I think the issue it's not strictly related to images. If you want i can remove images to test better.

Thanks for your time.

[liZe]

Just a question do WeasyPrint download the images in parallel or serially?

They're downloaded serially. I would be better to have parallel download, but that would not be as interesting as in a browser: layout is only done once in WeasyPrint and depends on image sizes, when it's done again and again in a browser while images are still downloading.

I can't find any simple solution right now, but we may find a smart way to asynchronously download images when creating the boxes and wait for them if they're not downloaded yet during the layout. If you're interested, that's a cool new issue to open 😉!

[liZe]

@liZe , as i said i tried to remove all image and all css but performance increase only 20 seconds.

Sorry, I forgot this comment.

With only background-image disabled, WeasyPrint generates a 91-page PDF in 1:05 for me (2:20 with background-image enabled). It's slower than a browser, but it's in WeasyPrint's poor standards (between 0.1 and 1 second per page for large real-life documents).

Other bugs have been reported with performance problems (see #553 or #483 for example), but these issues were focused on special cases that were outstandingly slow. Your case just shows that WeasyPrint is generally slow 😢, there's unfortunately no magic wand to fix your problem.

I try hard to improve performance and memory use on a regular basis (see #384 and #70 for example), but it's an endless, very frustrating work. I can sometimes spend days getting a 5% speed improvement … or often nothing.

I'm open to suggestions (and parallel download of resources is a good one), but I didn't find another specific point of failure I can work on for your document.

[Saksow]

Hi @liZe it's been a while 😄 I am interested in learning the best practices when designing you HTML/CSS template and using WeasyPrint, in order to get the fastest result possible. Are there written guidelines or can you list few points? Thanks!

[jnoortheen]

@liZe As you know, In Python world, often speedup involves writing part of code in Cython. Eventhough it is tedious at first, it will not affect the users installing weasyprint through pip if binary wheels are available. Another option is to try nuitka for certain performance related modules and compile them alone.