Strings in Java and many other JVM-languages consist of Unicode content and are encoded as utf-16. It was fantastic to already consider Unicode when introducing Java in the 90es and to make it the only reasonable way to use strings, so there is no temptation to start with a „US-ASCII“-version of a software that never really gets enhanced to deal properly with non-English content and data. And it also avoids having to deal with many kinds of String encodings within the software. For outside storage in databases and files and for communication with other processes and over the network, these issues of course remain. But Java strings are always encoded utf-16. We can be sure. Most common languages can make use of these Strings easily and handling common letter based languages from Europe and western Asia is quite strait forward. Rendering may still be a challenge, but that is another issue. A major drawback of this approach is that more memory is used. This usually does not hurt too much. Memory is not so expensive and a couple of Strings even with utf-16 will not be too big. With 64-bit Java, which should be used these days, the memory limitations of the JVM are not relevant any more, they can basically use as much memory as provided.
But some applications to hit the memory limits. And since usually most of the data we are dealing with is ultimately strings and combinations of relatively long strings with some reference pointers, some integer numbers and more strings, we can say that in typical memory intensive applications strings actually consume a large fraction of the memory. This leads to the temptation of using or even writing a string library that uses utf-8 or some other more condensed internal format, while still being able to express any Unicode content. This is possible and I have done it. Unfortunately it is very painful, because Strings are quite deeply integrated into the language and explicit conversions need to be added in many places to deal with this. But it is possible and can save a lot of memory. In my case we were able to abandon this approach, because other optimizations, that were less painful, proved to be sufficient.
An interesting idea is to compress strings. If they are long enough, algorithms like gzip work on a single string. As with utf-8, selectively accessing parts of the string becomes expensive, because it can only be achieved by parsing the string from the beginning or by adding indexing structures. We just do not know which byte to go to for accessing the n-th character, even with utf-8. In reality we often do not have long strings, but rather many relatively short strings. They do not compress well by themselves. If we know our data and have a rough idea about the content of our complete set of strings, custom compression algorithm can be generated. This allows good results even for relatively short strings, as long as they are within the „language“ that we are prepared for. This is more or less similar to the step from utf-16 to utf-8, because we replace common byte sequences by shorter byte sequences and less common sequences may even get replaced by something longer. There is no gain in utf-8, if we have mostly strings that are in non-Latin alphabets. Even Cyrillic or Greek, that are alphabets similar to the Latin alphabet, will end up needing two bytes for each letter, which is not at all better than utf-16. For other alphabets it will even become worse, because three or four bytes are needed for one symbol that could easily be expressed with two bytes in utf-16. But if we know our data well enough, the approach with the specific compression will work fine. The „dictionary“ for the compression needs to be stored only once, maybe hard-coded in the software, and not in each string. It might be of interest to consider building the dictionary dynamically at run-time, like it is done with gzip, but keeping it in a common place for all strings and thus sharing it. The custom strings that I used where actually using a hard coded compression algorithm generated using a large amount of typical data. It worked fine, but was just too clumsy to use because Java is not prepared to replace String with something else without really messing around in the standard run-time libraries, which I would neither recommend nor want.
It is important to consider the following issues:
- Is the memory consumption of the strings really a problem?
- Are there easier optimizations that solve the problem?
- Can it just be solved by adding more hardware? Yes, this is a legitimate question.
- Are there solutions for the problem in the internet or even within the current organization?
- A new String class is so fundamental that excellent testing is absolutely mandatory. The unit tests should be very extensive and complete.