Thread: pg_depend patch

pg_depend patch

From
Rod Taylor
Date:
While looking forward to the domain patch set being applied, I started
pg_depend support.

2 main functions, dependCreate() and dependDelete().

dependCreate() generates a dependence between two system objects, and
can optionally be forced to 'always cascade' the drop for items such as
complex types, toast tables or indexes.

dependDelete() removes or restricts the removal of all dependent objects
on the one being dropped informing the user of the cascade.

Affected areas (thus far):
- drop type
- drop view
- drop trigger
- drop sequence
- drop rule
- drop operator
- drop language
- drop function
- drop index
- drop aggregate
- create language
- create index
- create type

Type to array type relation is now done using an 'always cascaded'
dependent relationship.


To be completed (currently uses old 'ignorance is bliss' methods):
- Drop serial on column drop (tables cascade to drop all columns)
- Drop triggers via always cascade relationship (uses hard coded method)
- create [ view | trigger | table | sequence | rule | operator |
function | aggregate ] need to record dependencies on creation time.
- RESTRICT / CASCADE keywords should be used with drop statements
(Always restricts, unless it's an implicit cascade)
- BOOTSTRAPPED objects need their dependencies recorded.


Problems:
Regression tests can fail as the OID in names of some objects (toast
tables, indexes, etc) are never the same.  Solution is to not mention
implicit cascades, or potentially to mark specific implicit cascades as
'silent drop' while others are 'informed drop'.


Thanks for taking a look.  I'm sure I did some weird stuff.


BTW.  In dependDelete() each loop rescans incase while running down the
tree something else depended on an object I wanted to drop.  A good case
was an index, function, type loop.  Dropping them without rescanning and
a CommandCounterIncrement() would cause double tuple update issues.

Attachment

Re: pg_depend patch

From
"Christopher Kings-Lynne"
Date:
I was thinking.  With a patch such as this, this is the situation:

You have a view or a function or something that depends on another view.
You want to update the view that these depend on.  Up until this patch is
applied, you can just drop the view and recreate it and the things that
depend on it will carry on working.

However, after this patch you only have two choices: drop the view and have
everything that depends on it disappear, or have your drop attempt denied by
restrict.

ie. After this patch is applied, it will be essential that CREATE [OR
REPLACE] is implemented for all types of objects involved in the dependency
hierarchy.

Is this correct?

Chris

> -----Original Message-----
> From: pgsql-patches-owner@postgresql.org
> [mailto:pgsql-patches-owner@postgresql.org]On Behalf Of Rod Taylor
> Sent: Thursday, 14 March 2002 10:50 PM
> To: pgsql-patches@postgresql.org
> Subject: [PATCHES] pg_depend patch
>
>
> While looking forward to the domain patch set being applied, I started
> pg_depend support.
>
> 2 main functions, dependCreate() and dependDelete().
>
> dependCreate() generates a dependence between two system objects, and
> can optionally be forced to 'always cascade' the drop for items such as
> complex types, toast tables or indexes.
>
> dependDelete() removes or restricts the removal of all dependent objects
> on the one being dropped informing the user of the cascade.
>
> Affected areas (thus far):
> - drop type
> - drop view
> - drop trigger
> - drop sequence
> - drop rule
> - drop operator
> - drop language
> - drop function
> - drop index
> - drop aggregate
> - create language
> - create index
> - create type
>
> Type to array type relation is now done using an 'always cascaded'
> dependent relationship.
>
>
> To be completed (currently uses old 'ignorance is bliss' methods):
> - Drop serial on column drop (tables cascade to drop all columns)
> - Drop triggers via always cascade relationship (uses hard coded method)
> - create [ view | trigger | table | sequence | rule | operator |
> function | aggregate ] need to record dependencies on creation time.
> - RESTRICT / CASCADE keywords should be used with drop statements
> (Always restricts, unless it's an implicit cascade)
> - BOOTSTRAPPED objects need their dependencies recorded.
>
>
> Problems:
> Regression tests can fail as the OID in names of some objects (toast
> tables, indexes, etc) are never the same.  Solution is to not mention
> implicit cascades, or potentially to mark specific implicit cascades as
> 'silent drop' while others are 'informed drop'.
>
>
> Thanks for taking a look.  I'm sure I did some weird stuff.
>
>
> BTW.  In dependDelete() each loop rescans incase while running down the
> tree something else depended on an object I wanted to drop.  A good case
> was an index, function, type loop.  Dropping them without rescanning and
> a CommandCounterIncrement() would cause double tuple update issues.
>


Re: pg_depend patch

From
Peter Eisentraut
Date:
Christopher Kings-Lynne writes:

> You have a view or a function or something that depends on another view.
> You want to update the view that these depend on.  Up until this patch is
> applied, you can just drop the view and recreate it and the things that
> depend on it will carry on working.

No they won't.  That's exactly the reason why we need this dependency
tracking.

--
Peter Eisentraut   peter_e@gmx.net


Re: pg_depend patch

From
"Rod Taylor"
Date:
Ahh.. Thanks.

I'm completely confident it'll work however when the grammer portions
are added and everything is tracked at creation time.  It's been setup
so the calling function must pass the keyword currently, it's a simple
matter of pulling it from the grammer.

However, I'll configure that element (and the domain stuff) to
actually use it -- although they won't cascade very far until the
items they're cascading through also have the ability.
--
Rod Taylor

This message represents the official view of the voices in my head

----- Original Message -----
From: "Christopher Kings-Lynne" <chriskl@familyhealth.com.au>
To: "Rod Taylor" <rbt@zort.ca>
Sent: Thursday, March 14, 2002 8:32 PM
Subject: RE: [PATCHES] pg_depend patch


> > To be completed (currently uses old 'ignorance is bliss' methods):
> > - Drop serial on column drop (tables cascade to drop all columns)
> > - Drop triggers via always cascade relationship (uses hard coded
method)
> > - create [ view | trigger | table | sequence | rule | operator |
> > function | aggregate ] need to record dependencies on creation
time.
> > - RESTRICT / CASCADE keywords should be used with drop statements
> > (Always restricts, unless it's an implicit cascade)
>
> If you want something to experiement with, the ALTER TABLE/DROP
CONSTAINT
> code currently REQUIRES the restrict/cascade keyword but just
ignores it...
>
> Chris
>
>


Re: pg_depend patch

From
"Rod Taylor"
Date:
> You have a view or a function or something that depends on another
view.
> You want to update the view that these depend on.  Up until this
patch is
> applied, you can just drop the view and recreate it and the things
that
> depend on it will carry on working.
>
> However, after this patch you only have two choices: drop the view
and have
> everything that depends on it disappear, or have your drop attempt
denied by
> restrict.
>
> ie. After this patch is applied, it will be essential that CREATE
[OR
> REPLACE] is implemented for all types of objects involved in the
dependency
> hierarchy.

> Is this correct?

That would be correct. And with good reason.  If you have a function
which depends on a view, and you drop and re-create the view you also
need to recompile the function.  Right now we don't (also need to drop
and recreate function).

Atleast now (soon) you will be able to find out what depended on the
view in the first place to tell them to 'reset' if necessary.


> > -----Original Message-----
> > From: pgsql-patches-owner@postgresql.org
> > [mailto:pgsql-patches-owner@postgresql.org]On Behalf Of Rod Taylor
> > Sent: Thursday, 14 March 2002 10:50 PM
> > To: pgsql-patches@postgresql.org
> > Subject: [PATCHES] pg_depend patch
> >
> >
> > While looking forward to the domain patch set being applied, I
started
> > pg_depend support.
> >
> > 2 main functions, dependCreate() and dependDelete().
> >
> > dependCreate() generates a dependence between two system objects,
and
> > can optionally be forced to 'always cascade' the drop for items
such as
> > complex types, toast tables or indexes.
> >
> > dependDelete() removes or restricts the removal of all dependent
objects
> > on the one being dropped informing the user of the cascade.
> >
> > Affected areas (thus far):
> > - drop type
> > - drop view
> > - drop trigger
> > - drop sequence
> > - drop rule
> > - drop operator
> > - drop language
> > - drop function
> > - drop index
> > - drop aggregate
> > - create language
> > - create index
> > - create type
> >
> > Type to array type relation is now done using an 'always cascaded'
> > dependent relationship.
> >
> >
> > To be completed (currently uses old 'ignorance is bliss' methods):
> > - Drop serial on column drop (tables cascade to drop all columns)
> > - Drop triggers via always cascade relationship (uses hard coded
method)
> > - create [ view | trigger | table | sequence | rule | operator |
> > function | aggregate ] need to record dependencies on creation
time.
> > - RESTRICT / CASCADE keywords should be used with drop statements
> > (Always restricts, unless it's an implicit cascade)
> > - BOOTSTRAPPED objects need their dependencies recorded.
> >
> >
> > Problems:
> > Regression tests can fail as the OID in names of some objects
(toast
> > tables, indexes, etc) are never the same.  Solution is to not
mention
> > implicit cascades, or potentially to mark specific implicit
cascades as
> > 'silent drop' while others are 'informed drop'.
> >
> >
> > Thanks for taking a look.  I'm sure I did some weird stuff.
> >
> >
> > BTW.  In dependDelete() each loop rescans incase while running
down the
> > tree something else depended on an object I wanted to drop.  A
good case
> > was an index, function, type loop.  Dropping them without
rescanning and
> > a CommandCounterIncrement() would cause double tuple update
issues.
> >
>
>