On Fri, Jun 18, 2010 at 12:59 PM, Andrew Dunstan
<andrew@dunslane.net> wrote:
David E. Wheeler wrote:
On Jun 18, 2010, at 9:34 AM, Andrew Dunstan wrote:
I'd be perfectly happy to hear a reasonable alternative. Assuming we use some integer representation, given two labels represented by n and n+1, we can't add a label between them without rewriting the tables that use the type, whether it's my representation scheme or some other. Maybe we could have a FORCE option which would rewrite if necessary.
People would likely always use it.
Why not use a decimal number?
Probably it'd be the same as the decimal suggestion above, but we can use floating-point data type.
It will allow injection of a new label at any stage.
CREATE leads to
Label1 -> 1.0
Label2 -> 2.0
Label3 -> 3.0
ALTER ... ADD Label4 AFTER Label2; leads to
Label1 -> 1.0
Label2 -> 2.0
Label4 -> 2.5
Label3 -> 3.0
ALTER ... ADD Label5 AFTER Label2; leads to
Label1 -> 1.0
Label2 -> 2.0
Label5 -> 2.25
Label4 -> 2.5
Label3 -> 3.0
Since floating-point implementation is architecture dependent, the ALTER command should check that the injected value does not equate to any value around it (eg. comparisons of (2.5 == 2) and (2.25 == 2.5) should not yield 0); and if it does, then throw an error and ask the user to use the rewrite-the-table version of the command.
And since it is still 32 bit, and comparisons done by machine, performance should be acceptably close to current integer comparisons, and much faster that the cache lookups etc. being proposed.
This is very similar to Andrew's original suggestion of splitting 32 bits into 16+16, but managed by the machine hence no complicated comparison algos needed on our part. Also, since this is all transparent to the SQL interface, our dump-reload cycle or Slony replication, etc. should not be affected either.
Regards,
--
gurjeet.singh
@ EnterpriseDB - The Enterprise Postgres Company
http://www.EnterpriseDB.comsingh.gurjeet@{ gmail | yahoo }.com
Twitter/Skype: singh_gurjeet
Mail sent from my BlackLaptop device