I am already a bit late to write about my MySQL Users Conference impressions or input, but better later than never. My pet topic is scalability, or rather, how to build big cheap systems, and I’ve had many mixed thoughts after the event, which of course had many scalability gurus from nice companies. The biggest impression was that we all scale different applications and have different demands (some have many datacenters with applications distributed, some had two power failures in whole datacenter in single week and went down for few hours..).
And as I also had a presentation on Wikipedia scaling, I’ll try to mention some of issues discussed there.
Main thing is that rules do not matter, application (or rather a purpose) does. All techniques should be taken with grain of salt, MMORPG is different from e-banking, though both may require synchronized states. A blog is not a wiki, as you won’t have clashes or lock conflicts on same resources. And sure, in some cases high availability (percentage of uptime) is less important than general availability – percentage of reach.
Distributing the load
Second major idea is that load has to be split. Of course, it is mandatory in case of ‘scale out’, but there may be different paths to acquire different kinds of needs – efficiency, availability, redundancy, accuracy, yadda yadda. Like…
- If you know how to manage desynched slaves (what is really tied to application process), you can allow possibility of desynch (and hence not flush logs to disk after every transaction on any of your boxes).
- Queries for different data domains can be sent to different slaves. Let it be per user, per language or any other fragmentation. Even if data is all there, just touching specific parts of it may improve cache locality. One of main things to consider then, is that data is clustered together, and a single row fetch will read 16KB block from disk and store it in memory. “Every third article/product/user/…” should be replaced by “every third thousand of articles/products/users”, or even some semi-dynamic bucket allocation. For Wikipedia it’s rather easy, we may just direct different languages to different servers, as per-language projects are quite self-contained.
- Different types of queries can be sent to different slaves. Even if same data is there, you can still hit it with different patterns, and keep different indexes in hottest caches.
- Not having the data will be always faster than just not reading it. If there’s enough of redundancy, data from different domains/types can be simply purged. Of course, purging the data that is not needed completely from the system is even more efficient approach.
- You’ve already got RAIS – Redundant Array of Inexpensive Servers, so you can take out R from RAID and use just stripes of disks, forming AID, for performance of course.
Weakest slave will be slowing down capability of whole system, so doing less work on it not only in terms of requests per second, but also of how much of data it has to handle, may revive it for a new life (and, hehe, that way our poor 4GB old DB servers do have lots of juice :).
Caching is essential in high-performance environments (unless the service is random number generator two-point-oh). It is a common practice to add big nice caching layer (in memcacheds or squids or wherever else), but to leave data in core databases as it is. If efficient caching allows not to access data inside database too often, there’s no need to keep it on core database systems, as those are designed to work with data that needs work. Any box that has some idle resource like storage (most application servers usually do), may handle piece of rarely accessed but heavily cached elsewhere content.
Tools for the task
Different tasks may require different tools for the job. Lots of semi-static data can often be stored on application servers, usually as lightweight hash databases, just a proper method of migrating dynamic changes from core databases is required. It may be a simple rsync after a change was made, but it will save a roundtrip afterwards. Instead of updating full text indexes inside database, streams of changes may go to Lucene-based search application. And of course, sometimes just putting changes into background queues or off-peak schedules may improve responsiveness.
Speed vs power – both important
In scaled out environments adding more hardware often helps, but shouldn’t always be the main solution of the problem. Micro-optimizations have the purpose – besides obvious “saves resources” they also increase efficiency of individual nodes. Having the query served faster means also less locking or occupation of common resources (such as DB threads, waits on network), as well as far more improved user experience. This is where you might want to use high-power cores as in Opteron instead of lots of Niagara or Celeron ones (even if that may look much cheaper). Having 100 requests per second at 0.1s each rather than 100 requests per second at 1s each is quite a difference, and it counts.
It is critical to avoid slow tasks on high performance systems. If there’re queries that are too expensive, just… kill them. Once you become overloaded you might want to start killing queries that run too long. Just KILLing the thread is not enough, either it has to be optimized (indexes, summary tables, etc), or eliminated. One cheap way is to delete outdated data, but if it is not possible, just having another table with hottest aggregated data may double or triple the performance. Here again, once data is aggregated into commonly used format, main source can be retired from hot memory to disks or to other cheaper services.
My common illustration is – you don’t want to see elephant walking in a highway. Elephant will have absolutely different access pattern, occupy too much space and move too slowly, where usually lots of throughput exists, not only blocking few lanes, but also attracting attention by drivers in opposite direction. Kill the elephant! Or rather, leave in natural habitat.
One of the magic weapons is compression. Gzip is fast, bzip2 is not, and many common perceptions of compression is that it is slow. No, it’s bzip2 that is slow, gzip is fast. Additionally, it may pack your data into a single block on file system instead of two or three. In some cases that may mean three times less seeks – milliseconds saved at a tiny fraction of CPU costs. In cases where there’s lots of similar text – like comments quoting other comments, different revisions for entry – concatenate it all and then compress. Gzip will love the similarity and produce ten or twenty times smaller BSOB (Binary Small Object).
There’re various profiling advices around, but what I hit multiple times, is that one hundred requests profiled separately one by one may provide more insights than a generic collection of min/max/avg/deviation for a million requests. Moreover, profile from production system may give lots of issues unspotted in development. It doesn’t mean though, that generic or development profiling should not be done. There should be no prejudices in process of profiling – worse than that is just optimization without profiling. Instead of “it has to be so” there should always a question if specific task can be improved.
Conclusions (or bragging (or whining))
Site handles now (it rises quite fast) over 12000 HTTP requests per second (out of which around 4000 are pageviews), on a cluster that could be built with ~500k$. At one talk in UC it was told that our platform isn’t very good. Sorry, with few volunteers working on that, we have to choose priorities for development. And it is a pity, that most of scaling management software is usually closed asset of the big players. Um, I’d love it to be open, really, pretty pretty please!