Hi all!
At the hackathon in DC, I have been discussing options for caching Wikidata
content with a few people, primarily RobLa and Ryan Lane. Here's a quick
overview of the problem, and what I took away from the conversation.
So, what we need is this:
* Wikidata objects are multi-lingual
* Wikidata generates language specific views of the data objects ("items" and
other "entities") as HTML.
* To provide good response times, we need to cache that generated HTML on some
level, but that cache needs to be specific to (i.e. varying on) the user's language.
* anonymous users must have a way to set their preferred language, so they are
able to read pages on wikidata in their native language.
Now, there are (at least) two levels of caching to be considered:
* the parser cache, containing the HTML for every page's content. This cache
uses keys based on the ParserOptions, which includes a field for specifying the
language. However, currently only a "canonical" version of the rendered page is
cached, disregarding the user language.
* the squid cache, containing full HTML pages for each URL. Currently, our squid
setup doesn't support language versions of the same page; it does however
support multiple versions of the same page in general - a mechanism that is
already in use for caching different output for different mobile devices.
For each cache, we have to decide whether we want to split the cache by
language, or not use that cache for wikidata. After some discussion with RobLa
and Ryan, we came to the following conclusion:
* We will not use the parser cache for data pages (normal wikitext pages on the
wikidata site, e.g. talk pages, will use it). The reasons are:
a) memory for the parser cache is relatively scarce, multiplying the number of
cached objects by the number of languages (dozens, if not hundreds) is not a
good idea.
b) generating HTML from Wikidata objects (which are JSON-like array structures)
is much quicker than generating HTML from wikitext
c) Wikidata will have relatively view direct page view, it will mainly be used
via Wikipedia pages or the API.
Splitting the parser cache by language may be desired in the future, depending
on what caching mechanisms will be available for the parser cache, and how
expensive HTML generation actually is. We agreed however that this could be left
for later and only needs to be addressed in case of actual performance problems.
* We will split the squid cache by language, using a cookie that specifies the
user's language preference (for both logged in users and anons). The same URL is
used for all language versions of the page (this keeps purging simple). The
reasons are:
a) bypassing caching completely would open us up to DDoS attacks, or even
unintentional DoS by spiders, etc.
b) the amount of pages in the cache need not be large, only a small subset will
be actively viewed.
c) squid memory is relatively plenty, the experience with splitting the cache by
mobile device shows that this is feasible.
Note that logged in users currently bypass the squid cache. This may change in
the future, but the above scheme is oblivious to this: it will cache based on
the language cookie. Looking more closely, there is four cases to consider:
* anonymous with language cookie: vary cache on the language in the cookie, keep
URL unchanged
* logged in with language cookie: once squid cache is implemented for logged in
users, vary cache on the language in the cookie, keep URL unchanged. Until then,
bypass the cache.
* anonymous with no cookie set (first access, or the user agent doesn't support
cookies): this should use the primary content language (english). Squids could
also detect the user language somehow, set the cookie, and only then pass the
request to the app server - this would be nice, but puts app logic into the
squids. Can be added later if desired, MediaWiki is oblivious to this.
* logged in with no cookie: skip cache. the response will depend on the user's
language preference stored in the database, caching it without the language
cookie to vary on would poison the cache. We could try to come up with a scheme
to mitigate the overhead this generates, but that only becomes relevant once
squid layer caching is supported for logged in users at all.
So, that's it: no parser cache, split squid cache, use cookie smartly.
What do you think? Does that sound good to you?
-- daniel