But seriously… • We only care about PHP! • So where does Flickr use it?
PHP is at the core of Flickr Photo Storage Database Node Service Business/Application Logic Parser Page Logic API Logic Templates Endpoints Email Flickr.com 3rd Party Apps Flickr Apps Users
Ok, ok – what besides PHP? • Smarty for templating • PEAR for XML and Email parsing • Java for… – Controlling ImageMagick (image processing) – Storage metadata – The node service • MySQL (4.0 / InnoDb) • Perl for deployment & testing tools • Apache 2, Redhat, etc. etc.
Medium sized application • Small team (3 programmers until recently) – 1 PHP, 1 Flash/DHTML, 1 Java • >60,000 lines of PHP code – >80 smarty extensions • >60,000 lines of templates • >250,000 users • >3,500,000 photos • >50,000,000 page views per month • Growing fast – Like, really fast • So these stats are out of date by now
Thinking outside the web app • Services – Atom/RSS/RDF Feeds – APIs • SOAP • XML-RPC • REST • We love PEAR::XML::Tree
More services • Email interface – Postfix – PHP – PEAR::Mail::mimeDecode • FTP • Uploading API • Authentication API • Unicode – (Not really a service, but common to all Flickr services)
Even more services • Real time application – The “node service” • ‘Cool’ flash apps – Which use the REST APIs • Blogging APIs – Blogger API (1 & 2) – Metaweblog API – Atom – LiveJournal
APIs are simple! • Modeled on XML-RPC (sort of) • Method cal s with XML responses • Named arguments (key/name pairs) – Tricky in WebServices.framework on Mac OS X • SOAP, XML-RPC and REST are just transports • PHP endpoints mean we can use the same application logic as the website – Endpoints talk to the business logic using PHP function calls • Essentially a really fast transport
XML isn’t simple :( • PHP 4 doesn’t have good a XML parser – PHP 5 is new and scares me • (and it wasn’t out when we started) • Expat is cool though (PEAR::XML::Parser) • Why doesn’t PEAR have XPath? – Because PEAR is stupid! – PHP 4 sucks! – Actual y, PHPXPath rocks • http://phpxpath.sourceforge.net/
Creating API methods • Stateless method-cal APIs are easy to extend – They don’t affect each other • Adding a method requires no knowledge of the transport – We just get passed arguments and return XML – The transport layer hides all that junk • Adding a method once makes it available to al the interfaces • Self documenting – method dispatch requires a list of methods – Because everyone hates writing documentation
Scaling the beast • Why PHP is great • MySQL scaling • Search scaling • Horizontal scaling
But first… • Why do we need to scale? – There are a lot of people on the Internet – They all want to use our “web services” – Whether they know it yet or not
Why PHP is great • Stateless – We can bounce people around servers – Everything is stored in the database – Even the smarty cache – “Shared nothing” • (so long as we avoid PHP sessions) • But what this really means… – …is we just have to deal with scaling elsewhere
A MySQL Scaling Haiku • Database server slow • Load of over two hundred • Replication wins!
MySQL Replication • But it only gives you more SELECT’s • Else you need to partition vertically • Re-architecting sucks :(
Looking at usage • But really, we SELECT much more than anything else –A snapshot says • SELECT’s 44m • INSERT’s 1.3m • UPDATE’s 1.7m • DELETE’s 0.3m • 19 SELECT’s for each IUD
Replication is really cool • A bunch of slave servers handle all the SELECT’s • A single master handles IUD’s • We can scale horizontally, at least for a while.
Searching • A simple text search • We were using RLIKE • Then switched to LIKE • Then disabled it all together
FULLTEXT Indexes • FULLTEXT saves the day! • But they’re only supported on MyISAM tables • And we use InnoDb for locking • We’re doomed :(
But wait! • Partial replication saves the day • Replicate the portion of the database we want to search • But change the table types on the slave to MyISAM • It can keep up because it’s only handling IUD’s on a couple of tables • And we can reduce the IUD’s with a little bit of vertical partitioning
JOIN’s are slow • “Normalised data is for sissies” • Erm, • “Selective de-normalisation can be a big win” • Keep multiple copies of data around • Makes searching faster • Have to ensure consistency in the application logic • For instance, have a concat’d field containing a bunch of child-row data, just for searching.
Our current setup DB1 IUD’s Master SELECT’s DB2 Main Slave Slave Farm DB3 Main Search slave Search SELECT’s Search Slave Farm
Our current, current setup DB1 IUD’s DB4 Master IUD’s Master SELECT’s SELECT’s DB2 DB5 Main Slave Search Cluster Main Slave Slave Farm DB3 Main Search Slave Farm slave Search SELECT’s Main Cluster Aux Cluster Search Slave Farm
Horizontal scaling • At the core of our design • Just add hardware! • Inexpensive • Not exponential • Avoid redesigns/re-architectures
Talking to the Node Service • Just another service with an API – But just internal at the moment • Everyone speaks XML (badly) • Just TCP/IP - fsockopen() • We’re issuing commands, not requesting data, so we don’t bother to parse the response – Just substring search for state=“ok” – This only works for a simple protocol
Still talking to the Node Service •Don’t rely on it! – Check the connection was established – Use a connection timeout – Use an IO timeout!
RSS / Atom / RDF • Different formats – (al quite bad) • We’re generating a lot of different feeds • Abstract the difference away using templates • No good way to do private feeds. Why is nobody working on this? (WSSE maybe?) – Most of the feed readers (including bloglines.com) support basic HTTP Auth • Easy to implement in PHP – We love PHP » It’s great!
Receiving email • We want users to be able to email photos to Flickr • Get postfix to pipe each mail to a PHP script • Parse the mail and find any photos • Cel ular phone companies hate you • Lots of mailers are retarded – Photos as text/plain attachments – Segments out of order – No mime types – UUEncoded and mime-less
Processing email • PEAR to the rescue • Mail::mime_decode – With some patches • UUEncoding • Relax the address atom parser • We need to convert character sets – ICONV loves you
Upload via FTP • PHP isn’t so great at being a daemon – PHP4, I mean. Maybe PHP 5 is great • Leaks memory like a sieve • No (easy) threads • Java to the rescue • Java just acts as an FTPd and passes al uploaded files to PHP for processing – This isn’t actually public • Not my idea – Bricolage does this I think. Maybe Zope?
Blogs • Why does everyone loves blogs so much? • Only a few APIs real y – Blogger – Metaweblog – Blogger2 – Movable Type – Atom – Live Journal
It’s all broken • Lots of blog software has broken interfaces • It’s a support nightmare • Manila is tricky • But it al works, more or less • Abstracted in the application logic • We just cal blogs_post_message(); • And so can you, via the API
Back to those APIs • We opened up the Flickr APIs a few months ago • Programmers mainly build tools for other programmers • We now have Perl, python, PHP, ActionScript, XMLHTTP, .NET, Objective-C, C++, C and Ruby interface libraries • But also a few actual applications
iPhoto Plugin • We developed a Mac uploader • But it wasn’t great • A user developed an iPhoto plugin • It was great • APIs encourage people to do your work for you
Flickr Carnivore • Uses Carnivore PE – Sniffs AIM traffic (amongst others) from the local net • Calculates the most popular words of the moment • Uses the Flickr API to display photos of those words • It’s like a really invasive zeitgeist
Flickr Tivo • A Tivo app which uses Flickr photos • Just Type in some tags • And your TV becomes a “digital picture frame”
So what next? • Even more scaling • PHP 5? • MySQL 5? –or NDB? • Taking over the world