前往 Weblate! Menu

Speed improvements in Weblate

分享
Michal Čihař
博客功能
2013年7月19日

Yesterday, I've spent some time trying to improve speed of translation import in Weblate. It got annoyingly slow while adding new features and I felt there has to be some area for improvements even with all the features.

I was testing on development server with debug stuff enabled (what slows down things quite a lot already) and first import took about 30 seconds.

Quick look revealed that for linked repositories, the link lookup happens too often, so some caching could be used. This has saved about 3 seconds...

With offloaded indexing (what is anyway only useful setup for production use), every reindex request was checking whether there is already one before creating new one. In most cases there is none before, so it's simply faster to try to create one and ignore error in case it is already there. We're now 5 seconds faster.

Another big resource consumer are quality checks. They provide useful feedback to translators, but it means that quite a lot of code needs to be executed for every single string. First thing I've realized is that unit check state is recalculated everytime regardless the fact whether there was any change, what was easy to fix. The other speed issue is in consistency check - it does too complex SQL query, which usually does not match anything. Converting this to much simpler one (which will in most cases still have no results) improves performance for usual path while having minimal impact for the real checking. Another 4 seconds down on my test import.

Last thing which I changed is way how per translation statistics are updated. In past this required quite a lot of queries, but yesterday I've realized this can be much reduced. During test suite run on PostgreSQL, I saw that it's not that easy - PostgreSQL does not allow to sum boolean fields, but fortunately this can be fixed as well. This has saved about 2 seconds.

In the end the tested import took 16 seconds instead of original 30 and issued 2500 SQL queries instead of 5000 at the beginning. I hope you will enjoy regained speed of imports :-).

简体中文