Re: Single-file DBs WAS: Need concrete 'Why Postgres

Started by Andrew Dunstanover 22 years ago3 messages
#1Andrew Dunstan
andrew@dunslane.net

Is anyone seriously suggesting that postgres should support either raw
devices or use some sort of virtual file system? If not, this whole
discussion is way off topic. And if they are my response would be that it
would at best be a serious waste of time - there is far more important
work to do.

cheers

andrew

#2Andrew Dunstan
andrew@dunslane.net
In reply to: Andrew Dunstan (#1)

This makes sense to me. I sense a TODO item :-)

(My dim and possibly incorrect memory of administering Ingres around 10
years ago was that it supported both raw devices and file system based
databases. We opted for a file system base, for reasons others have
mentioned here, but I seem to recall we used a raw device for the
transaction log for performance reasons. But, as the saying goes, that was
a long time ago, and in another country.)

andrew

Tom wrote

Show quoted text

"Andrew Dunstan" <andrew@dunslane.net> writes:

Is anyone seriously suggesting that postgres should support either raw
devices or use some sort of virtual file system? If not, this whole
discussion is way off topic.

I have zero interest in actually doing it. However, it'd be nice if
the existing "storage manager" API were clean enough that our response
to this type of question could be "sure, go implement it, and when
you're done let us know what performance improvement you see". We've
allowed the smgr API to degenerate over the years. CREATE/DROP
DATABASE both bypass it, and the support for alternate database
locations messes up the API pretty thoroughly (not that there's
anything clean about that feature at all), and I think there are some
other issues with specific commands bypassing the smgr abstractions.

I think it would be reasonable to fix this as part of the "tablespaces"
work that people keep wanting to do.

regards, tom lane

#3Tom Lane
tgl@sss.pgh.pa.us
In reply to: Andrew Dunstan (#1)

"Andrew Dunstan" <andrew@dunslane.net> writes:

Is anyone seriously suggesting that postgres should support either raw
devices or use some sort of virtual file system? If not, this whole
discussion is way off topic.

I have zero interest in actually doing it. However, it'd be nice if the
existing "storage manager" API were clean enough that our response to
this type of question could be "sure, go implement it, and when you're
done let us know what performance improvement you see". We've allowed
the smgr API to degenerate over the years. CREATE/DROP DATABASE both
bypass it, and the support for alternate database locations messes up
the API pretty thoroughly (not that there's anything clean about that
feature at all), and I think there are some other issues with specific
commands bypassing the smgr abstractions.

I think it would be reasonable to fix this as part of the "tablespaces"
work that people keep wanting to do.

regards, tom lane