We recently stumbled upon a phenomen we absolutely weren’t aware of: You can kill any Java IDE and also any Java process with a simple regular expression…
Back in university, I was taught that regular expressions, which are called regular grammers or type 3 grammers always end up in an finite state automaton and can therefore be processed in linear time (input length doubles, processing time doubles). However, that’s only true for “sane” expressions. A regular expression can also result in an non-deterministic finite state automaton and things can get messed up quite bad.
Consider the expression: (0*)*A This will any number of zeros, followed by an upper case A. Now if you use Matcher.find() for this expression, everything is fine as long as there is a match in the input. However, if you call this, with ” 00000000000000000000″ as input, your program will hang (and so will the regex console in Eclipse or IntelliJ and every (Java-based) online regex service).
What at first glance looks like an infinite loop, truns out to be catastrophic backtracking. What this basically means is, that the matcher detects, that no A was found at the end of the input. Now the outer quantifier goes on step back – the inner one forward and again – no result. Therefore the matcher goes back step by step retrying all combinations to find a match. It will eventually return (without a match) but the complexity (and therefore the runtime) of this is expotential (adding one character to the input doubles the runtime). A detailed description can be found here: catastrophic backtracking
Here are some runtimes I measured (which almost exactly double for each character added):
0000000000: 0.1ms 00000000000: 0.2ms 000000000000: 0.7ms 0000000000000: 1.3ms 00000000000000: 1.7ms 000000000000000: 3.5ms 0000000000000000: 7.2ms 00000000000000000: 13.9ms 000000000000000000: 27.5ms 0000000000000000000: 55.5ms 00000000000000000000: 113.0ms 000000000000000000000: 226.4ms 0000000000000000000000: 439.1ms 00000000000000000000000: 886.0ms
As a little side-note: For micro benchmarks like this, you always need to “warm” up the JVM as the HotSpot JIT will jump in at some point and optimize the code. Therefore the first run looks like this:
0000000000: 6.8ms 00000000000: 11.8ms 000000000000: 25.5ms 0000000000000: 39.5ms 00000000000000: 6.3ms <- JIT jumped in and started to translate 000000000000000: 5.4ms to native code. 0000000000000000: 7.1ms 00000000000000000: 14.2ms 000000000000000000: 26.8ms 0000000000000000000: 54.4ms 00000000000000000000: 109.6ms 000000000000000000000: 222.1ms 0000000000000000000000: 439.2ms 00000000000000000000000: 885.6ms
So what’s the take-away here? If you’re running a server application or anything critical used by many users, don’t let them enter regular expressions unless you really trust them. There are regex implementations out there, which detect this problem and abort, but Java (up to JDK 8) doesn’t.
Note: You can test this with your local IDE or a small Java program to your hearts content – but please don’t start to knock out all the regex tester websites out there. Those guys provide a nice tool free of charge, so it would be quite unfair..
Here is the tiny benchmark I used:
public class Test {
public static void main(String[] args) {
for (int runs = 0; runs < 2; runs++) {
Pattern pattern = Pattern.compile("(0*)*A");
// Run from 5 to 25 characters
for (int length = 5; length < 25; length++) {
// Build input of specified length
String input = "";
for (int i = 0; i < length; i++) { input += "0"; }
// Measure the average duration of two calls...
long start = System.nanoTime();
for (int i = 0; i < 2; i++) {
pattern.matcher(input).find();
}
System.out.println(input + ": "
+ ((System.nanoTime() - start) / 2000000d)
+ "ms");
}
}
}
}
fyi: http://madbean.com/2004/mb2004-20/
another Java Regex “feature” ;)
Yes, this is known problem with anything that can be treated like “program code”. And the RegEx pattern literal certanly is. But A) IntelliJ IDEA won’t hang on this – and most likely on any – regular expression used in its “find”. And not because of multithreading or something. We just won’t let the evaluating thread to hang. B) There is a totally generic solution that allows this – if you think of the problem root. The problem root is that you loosing execution control whenever you invoke any lib function. And the more complex its code is the more… Read more »
I agree with the take-away and I think it fits in the more general approach of not “trusting” any user input.
Looking at the benchmark code, the issue there is mainly the use of greedy quantifiers. Using possessive quantifiers could prevent this behavior: http://docs.oracle.com/javase/tutorial/essential/regex/quant.html
Hey Andreas, This is very interesting. Thanks a lot for sharing it and it’s a good thing you mentioned about not testing this code online.
Even after using Regex for quite some time it’s only recently that I discovered catastrophic backtracking. In my case it broke one batch processor:
http://vladmihalcea.com/2014/02/24/the-regex-that-broke-a-server/
I don’t understand why matcher continues when it doesn’t find ‘A’ character in the end of the string
jmzc, the pattern evaluator doesn’t keep track of what it’s already processed. so it sees that there’s no A at the end, but it doesn’t know if there was an earlier A that was greedily consumed. thus it backtracks to the last wildcard and tries a shorter match. once again there is no A, so it backtracks again and tries a shorter match, etc., without ever figuring out how pointless this is.