On-disk Tuple Size - Mailing list pgsql-hackers
From | Curt Sampson |
---|---|
Subject | On-disk Tuple Size |
Date | |
Msg-id | Pine.NEB.4.43.0204201608060.467-100000@angelic.cynic.net Whole thread Raw |
Responses |
Re: On-disk Tuple Size
(Tom Lane <tgl@sss.pgh.pa.us>)
|
List | pgsql-hackers |
[I've moved this discussion about changing the line pointer from four bytes to two from -general to -hackers, since it's fairly technical. The entire message Tom is responding to is appended to this one.] On Sat, 20 Apr 2002, Tom Lane wrote: > Curt Sampson <cjs@cynic.net> writes: > > ... Then we could declare that all tuples must be aligned on a > > four-byte boundary, use the top 14 bits of a 16-bit line pointer as the > > address, and the bottom two bits for the LP_USED and LP_DELETED flag. > > This would slightly simplify the code for determining the flags, and > > incidently boost the maximum page size to 64K. > > Hmm. Maybe, but the net effect would only be to reduce the minimum row > overhead from 36 to 34 bytes. Not sure it's worth worrying about. Well, unless the implementation is hideously complex, I'd say that every byte is worth worrying about, given the amount of overhead that's currently there. 36 to 34 bytes could give something approaching a 5% performance increase for tables with short rows. (Actually, do we prefer the tables/rows or relations/tuples terminology here? I guess I kinda tend to use the latter for physical stuff.) If we could drop the OID from the tuple when it's not being used, that would be another four bytes, bringing the performance increase up towards 15% on tables with short rows. Of course I understand that all this is contingent not only on such changes being acceptable, but someone actually caring enough to write them. While we're at it, would someone have the time to explain to me how the on-disk CommandIds are used? A quick look at the code indicates that this is used for cursor consistency, among other things, but it's still a bit mysterious to me. > > ... I don't see why we would then > > need the LP_DELETED flag at all. > > I believe we do want to distinguish three states: live tuple, dead > tuple, and empty space. Otherwise there will be cases where you're > forced to move data immediately to collapse empty space, when there's > not a good reason to except that your representation can't cope. I don't understand this. Why do you need to collapse empty space immediately? Why not just wait until you can't find an empty fragment in the page that's big enough, and then do the collapse? Oh, on a final unrelated note, <john@akadine.com>, you're bouncing mail from my host for reasons not well explained ("550 Access denied.") I tried postmaster at your site, but that bounces mail too. If you want to work out the problem, drop me e-mail from some address at which you can be responded to. cjs -- Curt Sampson <cjs@cynic.net> +81 90 7737 2974 http://www.netbsd.org Don't you know, in this new Dark Age, we're alllight. --XTC ------- Previous Message --------
pgsql-hackers by date: