Question

Given that strings are immutable in .NET, I’m wondering why they have been designed such that string.Substring() takes O(substring.Length) time, instead of O(1)?

i.e. what were the tradeoffs, if any?

Answer 1

UPDATE: I liked this question so much, I just blogged it. See Strings, immutability and persistence

---

The short answer is: O(n) is O(1) if n does not grow large. Most people extract tiny substrings from tiny strings, so how the complexity grows asymptotically is completely irrelevant.

The long answer is:

An immutable data structure built such that operations on an instance permit re-use of the memory of the original with only a small amount (typically O(1) or O(lg n)) of copying or new allocation is called a “persistent” immutable data structure. Strings in .NET are immutable; your question is essentially “why are they not persistent”?

Because when you look at operations that are typically done on strings in .NET programs, it is in every relevant way hardly worse at all to simply make an entirely new string. The expense and difficulty of building a complex persistent data structure doesn’t pay for itself.

People typically use “substring” to extract a short string — say, ten or twenty characters — out of a somewhat longer string — maybe a couple hundred characters. You have a line of text in a comma-separated file and you want to extract the third field, which is a last name. The line will be maybe a couple hundred characters long, the name will be a couple dozen. String allocation and memory copying of fifty bytes is astonishingly fast on modern hardware. That making a new data structure that consists of a pointer to the middle of an existing string plus a length is also astonishingly fast is irrelevant; “fast enough” is by definition fast enough.

The substrings extracted are typically small in size and short in lifetime; the garbage collector is going to reclaim them soon, and they didn’t take up much room on the heap in the first place. So using a persistent strategy that encourages reuse of most of the memory is also not a win; all you’ve done is made your garbage collector get slower because now it has to worry about handling interior pointers.

If the substring operations people typically did on strings were completely different, then it would make sense to go with a persistent approach. If people typically had million-character strings, and were extracting thousands of overlapping substrings with sizes in the hundred-thousand-character range, and those substrings lived a long time on the heap, then it would make perfect sense to go with a persistent substring approach; it would be wasteful and foolish not to. But most line-of-business programmers do not do anything even vaguely like those sorts of things. .NET is not a platform that is tailored for the needs of the Human Genome Project; DNA analysis programmers have to solve problems with those string usage characteristics every day; odds are good that you do not. The few who do build their own persistent data structures that closely match their usage scenarios.

For example, my team writes programs that do on-the-fly analysis of C# and VB code as you type it. Some of those code files are enormous and thus we cannot be doing O(n) string manipulation to extract substrings or insert or delete characters. We have built a bunch of persistent immutable data structures for representing edits to a text buffer that permit us to quickly and efficiently re-use the bulk of the existing string data and the existing lexical and syntactic analyses upon a typical edit. This was a hard problem to solve and its solution was narrowly tailored to the specific domain of C# and VB code editing. It would be unrealistic to expect the built-in string type to solve this problem for us.