While writing Inklog, I’ve been debating with myself regarding my use of the filesystem as a datastore. While the filesystem certainly makes creating, updating, visualizing, backing up, and restoring data much easier than it would be in a database, it adds many hardships. First of all, the convenience of SQL is thrown out the window. While it is nice that using the filesystem doesn’t require a database server, not being able to use a database server means that more programming is involved. Additionally, things like searching through all the entries in the system become difficult, not to mention slow. Another downfall is that the filesystem limits the amount of metadata for each entry that can be kept in a simple fashion.
However, the biggest question on my mind was whether or not using a database server would be faster or slower when performing the most commonly requested actions: getting a list of recent items from the entire system, getting a list of items from a category, and getting one item. I decided to write a test case.
I created 6000 empty files in 20 directories. I also created a table in a mysql database that simulated the filesystem: name, mtime, dir, data. I added indexes on dir, name, and mtime. Then I started testing. In each case the test is run 10 times. Then the average is displayed. For the database tests, mysql_connect is called each time.
Getting the filenames of the 10 most recent entries from the entire system.
TOT TIME: 17.201191902161
AVG TIME: 1.5637447183782DATABASE
TOT TIME: 0.012644052505493
AVG TIME: 0.0011494593186812Getting the filenames of the 10 most recent files in a single directory.
TOT TIME: 0.46042513847351
AVG TIME: 0.041856830770319DATABASE
TOT TIME: 0.059021830558777
AVG TIME: 0.0053656209598888Getting one item.
TOT TIME: 0.0020089149475098
AVG TIME: 0.0001826286315918DATABASE
TOT TIME: 0.019086122512817
AVG TIME: 0.0017351020466198The database was 1360 times faster than the filesystem when looking for the 10 most recent items in the entire system. The database was 7.8 times faster when looking for the 10 most recent items in a single directory. However, the filesystem was 9.5 times faster at getting a single file.
These numbers skew greater and greater towards the database as the number of items increases. And, in the one place that the filesystem wins, the operation being performed is so un-time-consuming in general, that the increase in the speed of the filesystem doesn’t amount to much.
These tests were performed with the database being on the same server as the running script. Additionally, the server performing these actions was, basically, not performing anything else at the time. If your database server is only accessible over a 2400bps modem link, your results will differ greatly. Additionally, if your database server is heavily loaded, while your web server isn’t, you may also see very different results.
Benchmarks are crap, for the most part. They don’t really mean a whole lot, unless they represent the exact cases in which you will be using the functions being tested. However, in this case, they DO represent exactly what I will be doing.
What does this mean? Inklog will no longer use the file system as its main method of data storage.