> I don't think we want features for their own sake, though, and I'm
> not convinced that raw filesystems are actually useful. Course, it's
> not my itch, and PostgreSQL _is_ free software.
>
I agree that raw file systems are seldom useful with one caveat, more
advanced file systems are sometimes detrimental to database access.
Conceptually, a file system and a database are redundant, both are doing
their best to preserve data integrity. This is especially true with
journalling file systems. Not to mention technologies like reiserfs which
attempts to do sub-block allocation.
What I think would go a long way to improving database performance on
non-raw partitions would be a simplified file system -- SFS anyone? The
simplified file system would not track access time. It would not overly
try to manage disk space. The target applications are going to allocate
disk space on a block level, rather than quibble about 4K here or 8K here,
have a user defined standard allocation unit of 64K, 128K, or so on.
Reduction on allocation overhead also reduces meta-data updating I/O. I
can almost imagine 32BIT FAT with large clusers, only with real inodes.
The idea would be that a database, like PostgreSQL, would be managing the
data not the file system. The file systems job would only to be the most
minimalist interface to the OS.
The benefts would be awesome, near-raw partition access and standard OS
tools for maintainence.
