Usually, "select" isn't broken is a good motto to follow. Occasionally, there are cases where this is case. In particular, it may not be that it is broken, it may very well be that the way it works doesn't match the things that we need it to do.
I spoke about an optimization story that happened recently, in which we managed to reduce the average time from 5 - 10 seconds to 5 - 15 milliseconds.
What we needed was to walk a tree structure, which was stored in a database, and do various interesting tree based operations on it. The most natural way of working with trees is with recursion, and SQL is just not the right way of dealing with it.
Deciding to load the entire table to memory, build a real tree structure and perform all the operations on that tree structure has paid off tremendously. What is important to remember is that we hadn't had to do anything radical to the data model or the way the application worked. We only had to modify the implementation of the component that exposed that tree to the application.
One of the things that we had to deal was the case where the amount of data we have to deal with would exceed available memory. At least, we thought we had to deal with it.
But our tree was very simple, it consisted of a few properties and that it. Let us do the math about this, shall we?
- Name - 50 chars, unicode - 100 bytes
- 4 decimal fields - 16 bytes each = 64 bytes
- 3 boolean fields - 3 bytes
- Parent point - 4 bytes
- Set of children - average of 10 per node - 40 bytes + ~50 bytes bookkeeping
This is very rough, of course, but that would do. It puts the memory cost of a node at just under 256 bytes. We will use that number, because it is easier to work with.
Now, with 256 bytes per node, how many can we reasonably use?
Well, 100 MB will take 409,600 nodes or so. Which is pretty good number, I say. A table of that size is considered big, by most people. A GB of memory will give us 4,194,304 items in the tree, and keep the traversal speed near instantaneous. At that point, I would start thinking about the size of the node, because 256 bytes is big size. More realistic size would be 64 bytes or so (drop the name, pack the decimals, use linked list for children) which would give me 16,777,216 nodes for the same memory requirement.
All of those numbers are greater than the current and expected size of the data set, so there isn't a reason to care much beyond that.
The important thing here is to understand that the usual truth about "let the tool do the optimization" doesn't really hold true when you have specific scenarios. For solving very specific, very narrow circumstances, you can generally come up with a much better approach than the generic one.
Of course, this approach would not allow any generalization, and it doesn't have other benefits that using the common platform might have offered (needing to do our own transactions, for example).
Keep that in mind.