Jurgen Defurne wrote:
>
> carl garland wrote:
>
> > > Don't even think about 100000 separate tables in a database :-(. It's
> > >not so much that PG's own datastructures wouldn't cope, as that very
> > >few Unix filesystems can cope with 100000 files in a directory. You'd
> > >be killed on directory search times.
> > I understand the concern for directory search times but what if your
> > partition for the db files is under XFS or some other journaling fs that
> > allows for very quick search times on large directories. I also
> > saw that there may be concern over PGs own datastructures in that the
> > master tables that hold the table and index tables requires a seq
> > search for locating the tables. Why support a large # of tables in PG
> > if after a certain limit causes severe performance concerns. What if
> > your data model requires more 1,000,000 tables?
> >
>
> If the implementation is like above, there is much less concern with directory
> search times, although a directory might get fragmented and be spread out
> across the disk (with 1000000+ tables it will be fragmented).
If the filesystem uses block allocation. If the filesystem uses
extent-based allocation this wouldn't be a concern.
(I'm no expert on filesystems. Moshe Bar just happened to write an
article on filesystems in this month's Byte - www.byte.com).
> ... With the directory search above
> deleted, you still have to search your inode table.
Which could be enormous. Yuck.
Are there clever ways of managing huge numbers of inodes?
-Ron-