Christopher Kings-Lynne wrote:
>
> Hmmm. Personally, I think that a DROP COLUMN that cannot reclaim space is
> kinda useless - you may as well just use a view!!!
>
> So how would this occur?:
>
> 1. Lock target table for writing (allow reads)
> 2. Begin a table scan on target table, writing
> a new file with a particular filenode
> 3. Delete the attribute row from pg_attribute
> 4. Point the table in the catalog to the new filenode
> 5. Release locks
> 6. Commit transaction
> 7. Delete orhpan filenode
>
> i. Upon postmaster startup, remove any orphaned filenodes
>
> The real problem here is the fact that there are now missing attnos in
> pg_attribute. Either that's handled or we renumber the attnos - which is
> also quite hard?
The attnos should be renumbered and it's easy.
But the above seems only 20% of the total implementation.
If the attnos are renumbered, all objects which refer to
the numbers must be invalidated or re-compiled ...
For example the parameters of foreign key constraints
triggers are consist of relname and colnames currently.
There has been a proposal that change to use relid or
column numbers instead. Certainly it makes RENAME happy
but DROP COLUMN unhappy. If there's a foreign key a_rel/1/3
and the second column of the relation is dropped the
parameter must be changed to be a_rel/1/2. If neither
foreign key stuff nor DROP COLUMN take the other into
account, the consistency is easily broken.
regards,
Hiroshi Inoue
