Tactics for using PHP in a high-load site

No two sites are alike. You really need to get a tool like jmeter and benchmark to see where your problem points will be. You can spend a lot of time guessing and improving, but you won't see real results until you measure and compare your changes.

For example, for many years, the MySQL query cache was the solution to all of our performance problems. If your site was slow, MySQL experts suggested turning the query cache on. It turns out that if you have a high write load, the cache is actually crippling. If you turned it on without testing, you'd never know.

And don't forget that you are never done scaling. A site that handles 10req/s will need changes to support 1000req/s. And if you're lucking enough to need to support 10,000req/s, your architecture will probably look completely different as well.

Databases

• Don't use MySQLi -- PDO is the 'modern' OO database access layer. The most important feature to use is placeholders in your queries. It's smart enough to use server side prepares and other optimizations for you as well.
• You probably don't want to break your database up at this point. If you do find that one database isn't cutting, there are several techniques to scale up, depending on your app. Replicating to additional servers typically works well if you have more reads than writes. Sharding is a technique to split your data over many machines.

Caching

• You probably don't want to cache in your database. The database is typically your bottleneck, so adding more IO's to it is typically a bad thing. There are several PHP caches out there that accomplish similar things like APC and Zend.
• Measure your system with caching on and off. I bet your cache is heavier than serving the pages straight.
• If it takes a long time to build your comments and article data from the db, integrate memcache into your system. You can cache the query results and store them in a memcached instance. It's important to remember that retrieving the data from memcache must be faster than assembling it from the database to see any benefit.
• If your articles aren't dynamic, or you have simple dynamic changes after it's generated, consider writing out html or php to the disk. You could have an index.php page that looks on disk for the article, if it's there, it streams it to the client. If it isn't, it generates the article, writes it to the disk and sends it to the client. Deleting files from the disk would cause pages to be re-written. If a comment is added to an article, delete the cached copy -- it would be regenerated.

I'm a lead developer on a site with over 15M users. We have had very little scaling problems because we planned for it EARLY and scaled thoughtfully. Here are some of the strategies I can suggest from my experience.

SCHEMA First off, denormalize your schemas. This means that rather than to have multiple relational tables, you should instead opt to have one big table. In general, joins are a waste of precious DB resources because doing multiple prepares and collation burns disk I/O's. Avoid them when you can.

The trade-off here is that you will be storing/pulling redundant data, but this is acceptable because data and intra-cage bandwidth is very cheap (bigger disks) whereas multiple prepare I/O's are orders of magnitude more expensive (more servers).

INDEXING Make sure that your queries utilize at least one index. Beware though, that indexes will cost you if you write or update frequently. There are some experimental tricks to avoid this.

You can try adding additional columns that aren't indexed which run parallel to your columns that are indexed. Then you can have an offline process that writes the non-indexed columns over the indexed columns in batches. This way, you can control better when mySQL will need to recompute the index.

Avoid computed queries like a plague. If you must compute a query, try to do this once at write time.

CACHING I highly recommend Memcached. It has been proven by the biggest players on the PHP stack (Facebook) and is very flexible. There are two methods to doing this, one is caching in your DB layer, the other is caching in your business logic layer.

The DB layer option would require caching the result of queries retrieved from the DB. You can hash your SQL query using md5() and use that as a lookup key before going to database. The upside to this is that it is pretty easy to implement. The downside (depending on implementation) is that you lose flexibility because you're treating all caching the same with regard to cache expiration.

In the shop I work in, we use business layer caching, which means each concrete class in our system controls its own caching schema and cache timeouts. This has worked pretty well for us, but be aware that items retrieved from DB may not be the same as items from cache, so you will have to update cache and DB together.

DATA SHARDING Replication only gets you so far. Sooner than you expect, your writes will become a bottleneck. To compensate, make sure to support data sharding early as possible. You will likely want to shoot yourself later if you don't.

It is pretty simple to implement. Basically, you want to separate the key authority from the data storage. Use a global DB to store a mapping between primary keys and cluster ids. You query this mapping to get a cluster, and then query the cluster to get the data. You can cache the hell out of this lookup operation which will make it a negligible operation.

The downside to this is that it may be difficult to piece together data from multiple shards. But, you can engineer your way around that as well.

OFFLINE PROCESSING Don't make the user wait for your backend if they don't have to. Build a job queue and move any processing that you can offline, doing it separate from the user's request.

I've worked on a few sites that get millions/hits/month backed by PHP & MySQL. Here are some basics:

1. Cache, cache, cache. Caching is one of the simplest and most effective ways to reduce load on your webserver and database. Cache page content, queries, expensive computation, anything that is I/O bound. Memcache is dead simple and effective.
2. Use multiple servers once you are maxed out. You can have multiple web servers and multiple database servers (with replication).
3. Reduce overall # of request to your webservers. This entails caching JS, CSS and images using expires headers. You can also move your static content to a CDN, which will speed up your user's experience.
4. Measure & benchmark. Run Nagios on your production machines and load test on your dev/qa server. You need to know when your server will catch on fire so you can prevent it.

I'd recommend reading Building Scalable Websites, it was written by one of the Flickr engineers and is a great reference.

Check out my blog post about scalability too, it has a lot of links to presentations about scaling with multiple languages and platforms: http://www.ryandoherty.net/2008/07/13/unicorns-and-scalability/