-----Original Message----- From: wikitech-l-bounces@lists.wikimedia.org [mailto:wikitech-l-bounces@lists.wikimedia.org] On Behalf Of Aryeh Gregor Sent: 16 December 2008 01:49 To: Wikimedia developers Subject: Re: [Wikitech-l] Future of Javascript and mediaWiki
On Mon, Dec 15, 2008 at 6:05 PM, Jared Williams jared.williams1@ntlworld.com wrote:
The idea being you could get a sdch capable user agent to
download the
concated & gzipped javascript in a single request (called a dictionary), which quick testing is about 15kb for
en.mediawiki.org,
cache that on the client for a long period.
Except that if you can do that, you can fetch the ordinary JS anyway and keep it until it expires. If clients actually did this, you'd be talking about single-digit JS requests per client per week, and further optimization would be pointless. The fact of the matter is that a significant percentage of views have no cached files from Wikipedia, for whatever reason. Those are the ones we're concerned with, and your idea does nothing to help them.
Technically it is possible to remove some of the JS & CSS roundtrips, as inlining the JS & CSS no longer comes at a download cost, the cost of copying something from the dictionary is around 3-5 bytes.
Also possible to put the static (inline html) in templates into the dictionary, together with a lot of the translated messages,
to try and
reduce the HTML size, though not sure how effective it'd be.
It looks to me like you're talking about hugely increasing the size of the first page view on the optimistic assumption that first page views account for only a very small percentage of hits. Why do you think that assumption is warranted?
The SDCH protocol has a time delay built in, so it wouldn't download the dictionary unless the browser has remained on a page (that has requested the
client download the dictionary) for a time.
One possible use of SDCH is that it would allow us to refrain from re-transmitting the redundant parts of each response, most of the skin bits. But that's probably under a couple of KB gzipped, so we're not talking about much savings here. It might even take longer to compute and apply the diffs than to transmit the extra data, over typical hardware.
Naively tried, was around 3kb for monobook. Hence my pessimism in my previous post, about beating plain gzip. The ratio of wikitext article output to skin bits is too high, I think.
On Mon, Dec 15, 2008 at 7:57 PM, Jared Williams jared.williams1@ntlworld.com wrote:
Only needs to download it once, and kept on client for months, as server can send a diff to update to the most recent version at any time, rather than having to send a whole new js file.
What leads you to believe that the client will actually *keep* it for months? Caches on clients are of finite size and things get evicted. The oldest file I have in Firefox's cache right now is from December 12, three days ago:
$ ls -lt /home/aryeh/.mozilla/firefox/7y6zjmhm.default/Cache | tail -rw------- 1 aryeh aryeh 22766 2008-12-12 09:34 14C66B37d01 -rw------- 1 aryeh aryeh 28914 2008-12-12 09:34 20916AFCd01 -rw------- 1 aryeh aryeh 59314 2008-12-12 09:34 57BFF8A6d01 -rw------- 1 aryeh aryeh 39826 2008-12-12 09:34 9EB67039d01 -rw------- 1 aryeh aryeh 52145 2008-12-12 09:34 A2F75899d01 -rw------- 1 aryeh aryeh 18072 2008-12-12 09:34 C11C3B29d01 -rw------- 1 aryeh aryeh 25590 2008-12-12 09:34 C845BED2d01 -rw------- 1 aryeh aryeh 32173 2008-12-12 09:34 D69FA933d01 -rw------- 1 aryeh aryeh 45189 2008-12-12 09:34 F2AD5614d01 -rw------- 1 aryeh aryeh 25164 2008-12-12 09:34 DBA05B7Bd01
Dictionaries don't seem to get deleted when clearing cache atleast. They have their own mechanism to specify how long they should be kept.
A fair percentage of computers might be even worse, like public computers configured to clean cache and other private data when the user logs out.
Concatenating the (user-independent) wikibits.js and so on to the (user-dependent) page HTML will prohibit intermediate proxies from caching the user-invariant stuff, too, since the entire thing will need to be marked Cache-Control: private. It will prevent our own Squids from caching the CSS and JS, in fact.
Why would you concatenate user-independent with user-dependent HTML?
I think I only suggested skin dependant, and language dependant dictionary.
Also would want the translation text from \languages\messages\Messages*.php in there too I think.
Handling the
$1 style placeholders is easy, its just determining what
message goes
through which wfMsg*() function, and if the WikiText
translations can be preconverted to html.
What happens if you have parser functions that depend on the value of $1 (allowed in some messages AFAIK)? What if $1 contains wikitext itself (I wouldn't be surprised if that were true somewhere)? How do you plan to do this substitution anyway, JavaScript? What about clients that don't support JavaScript?
The substitution doesn't really happen, as its done server side, translations are essentially preparsed, saving time in template rendering.
For example
http://www.amazon.com/exec/obidos/ISBN=$1
http://www.amazon.com/exec/obidos/ISBN= would be put into the dictionary, and server metadata for the translation becomes...
'Amazon.com' => array( array(offset, 39), // COPY 39 bytes from dictionary offset 1 // Output value of parameter 1. );
I do have working PHP code, That can parse PHP templates & language strings to generate the dictionary, and a new set of templates rewritten to output the vcdiff efficiently.
Jared