Re: further clarification: alter table alter column set not null - table scan is skipped

[Shuyu Pan]

I like your versions that emphasize: don’t drop the constraint in the same 
alter table set no null command.
Similar to David’s point, I spent some time trying to figure out a simple 
refactoring to carry the optimization all the way to the end but it might 
require executing “set not null” sooner which has a big impact. Another option 
is only implement a special treatment for this specific use case but it is a 
code smell to me. I believe a small clarification for the doc entry is the most 
efficient thing.

Sent from Yahoo Mail for iPhone


On Thursday, July 31, 2025, 09:01, David G. Johnston 
 wrote:

On Thursday, July 31, 2025, Álvaro Herrera  wrote:

On 2025-Jul-30, David G. Johnston wrote:

> On Wed, Jul 30, 2025, 13:55 PG Doc comments form 
> wrote:

> > The "table scan is skipped" optimization can use some clarification
> >
> > https://www.postgresql.org/doc s/current/sql-altertable.html# 
> > SQL-ALTERTABLE-DESC-SET-DROP- NOT-NULL
> > My proposal is "then the table scan is skipped if the alter statement
> > doesn't drop the constraint."

> I'm kinda hoping this is actually just a fixable bug...

I don't think so -- it's just the way ALTER TABLE is designed to work.
We don't promise that the subcommands are going to be executed in the
order that they are given, and thus this sort of thing can happen.
I suspect a mechanism that would throw an error at trying to drop the
constraint would be too complicated / brittle / laborious to write.

I wouldn’t want an error.  At the start of the command the constraint existed 
and its presence then would be enough.  It is immaterial that it went away 
during the command.  But it’s definitely not something that seems worth 
spending a non-trivial amount of effort on. 

(This is correct for 18; for 17 and earlier, the mention of NOT VALID
needs to be removed.)  Of course, in 18 you'd rely on ADD NOT NULL NOT
VALID instead of using a separate CHECK constraint.

Yeah, the main question here is whether we want to document for v17 and earlier 
what the article points out regarding locks.


Not sure if this reads better:

   if a valid CHECK constraint is
   found (and is not dropped in the same command) which
   proves no NULL can exist, then



If a valid check constraint exists (and is not dropped in the same command) 
which proves the absence of NULLs, then
I do agree the parenthetical should appear closer to the word constraint.
David J.

Re: further clarification: alter table alter column set not null - table scan is skipped

[Shuyu Pan]

Thanks a lot Álvaro for preparing the clarification so quickly. Looking forward 
to the release.
If we mark a column to skip table scan during drop (phase 0) and later skip the 
table scan when setting attributes (phase 7), we will not risk data corruption 
if the Access Exclusive lock is never released between phase 0 and 7.


Sent from Yahoo Mail for iPhone


On Monday, August 4, 2025, 04:32, Álvaro Herrera  wrote:

On 2025-Jul-31, Shuyu Pan wrote:

> I like your versions that emphasize: don’t drop the constraint in the
> same alter table set no null command.  Similar to David’s point, I
> spent some time trying to figure out a simple refactoring to carry the
> optimization all the way to the end but it might require executing
> “set not null” sooner which has a big impact. Another option is only
> implement a special treatment for this specific use case but it is a
> code smell to me.

Oh yeah, delaying the drop is much more likely to break other things.  I
was more thinking along the lines of maintaining a list of columns that
are known non-null at the start of the command (a bitmapset actually).
This could be computed in ALTER TABLE phase 1, and used later to
determine that no scans are needed.  But this is a lot of mechanism
which is useless 99% of the time, and moreso now that you can directly
add the NOT NULL constraints as NOT VALID to start with, which saves
having to mess with a separate CHECK constraint.

> I believe a small clarification for the doc entry is the most efficient thing.

Okay, I've pushed the change to all branches using David Johnston's
suggested wording.

Thank you all!

-- 
Álvaro Herrera        Breisgau, Deutschland  —  https://www.EnterpriseDB.com/