Linus Torvalds (torvalds@linux-foundation.org) on 1/10/09 wrote:
---------------------------

>I do agree with you that random writes are important,
>btw. I just don't think it's usually just totally
>random. In fact, the filesystem will hopefully turn even
>separate small files into somewhat contiguous data, although
>things like parallel makes can easily cause problems at
>least with traditional filesystems.
>
>Linus

It seems that frequently one wants to guess how data will be accessed and modified so it can be arranged on disk appropriately. Now for source files the standard file and directory abstraction seem to give the filesystem pretty much everything it would want to know but when it comes to database files or any other kind of structured data abstracting that data as a stream of bytes not only seems to be a poor fit but also makes it hard for the filesystem to properly predict file access patterns.

Now it seems to me it would have been better if disk storage had used a record based abstraction or offered some other more structured API for permanent storage. Not only does this seem more elegant to me but the filesystem would know enough about file access patterns to much better optimize their layout and if it the abstraction was powerful enough it seems one might be able to eliminate the whole mess with high end databases needing to bypass the standard file access layer. Also it would provide a much easier way to browse non-text files not to mention offering the possibility of a much more useful permissions system.

Is it simply inertia that favors the traditional unix file abstraction or some disadvantage that isn't occurring to me?