Обсуждение: Possible enhancement : replace view ?
Dear all, The current implementation of views uses OIDs, not table/view names. As a consequence, when you happen to replace (drop then create) an underlying table or view, you also have to drop and recreate all views using this table|view (and this recursively, of course ...). I stumbled on this while doing repeat analyses (involving repeated uses of aggregation) of the same dataset using slight variations of the subset of interest. When my dataset was small, I used to do that in (yuck !) MS-Access by creating a view defining the subset of interest, then creating views based on this view, and so on... Now that my dataset is too large to be Access-manageable, I migratedit to PostgreSQL (which, BTW, gave me nice performance enhancements), but I had to change my working habits. I have now to create a scriptdefining my views, then to run it at each and every variation of the subset of interest ... To be able to conserve existing views would definitely be a bonus. Of course, the overhead is necessary to handle the general case. However, there is a special case where this is unnecessary : when the new table or view "class" definition is a (possibly improper) subclass of the original one, or, if you prefer, when the column set of the new definition is a (possibly improper) superset of the old one. For tables, this case is already handled by a judicious use of alter table, at least in its present form (adding DROP COLUMN, which might be an interesting feature for other reasons, entails the risk of invalidating existing views ...). However, there is currently no easily reachable way to do that for a view (I suppose that the special case of modifying the definition of a view creating the same columns in the old and new definitions might be possible with a "clever" (ab)use of system catalogs, but I tend to be *very* wary of such hacks ...). Of course, I am aware that view definitions aren't just stored, but that a lot of rewriting is involved before storing theactual execution plan. Modifying a view definition would entail re-processing of other view definitions. But so is the case with the modification of a table ... What do you think ? -- Emmanuel Charpentier
On Wed, 2002-08-14 at 04:08, Emmanuel Charpentier wrote: > Dear all, > ... > Of course, I am aware that view definitions aren't just stored, but that > a lot of rewriting is involved before storing the actual execution > plan. Modifying a view definition would entail re-processing of other > view definitions. But so is the case with the modification of a table ... > > What do you think ? I'm trying to propose a scenario where 1. The SELECT clause defining the view is preserved 2. DROP of undrlying table/column will _not_ drop the view, but just mark it dirty 3. Using the view checks for the dirty flag and if it is set tries to recreate the view from its plaintext definition. --------------- Hannu
Emmanuel Charpentier <charpent@bacbuc.dyndns.org> writes:
> What do you think ?
I think Gavin Sherry is already working on this.
regards, tom lane
Hannu Krosing wrote: > On Wed, 2002-08-14 at 04:08, Emmanuel Charpentier wrote: > >>Dear all, >> > > ... > > >>Of course, I am aware that view definitions aren't just stored, but that >> a lot of rewriting is involved before storing the actual execution >>plan. Modifying a view definition would entail re-processing of other >>view definitions. But so is the case with the modification of a table ... >> >>What do you think ? > > > I'm trying to propose a scenario where > > 1. The SELECT clause defining the view is preserved > > 2. DROP of undrlying table/column will _not_ drop the view, but just > mark it dirty > > 3. Using the view checks for the dirty flag and if it is set tries to > recreate the view from its plaintext definition. I might be dense, but why not try to recreate it directly after the table/column modification ? > --------------- > Hannu > > > ---------------------------(end of broadcast)--------------------------- > TIP 5: Have you checked our extensive FAQ? > > http://www.postgresql.org/users-lounge/docs/faq.html
Tom, I submitted a patch for this a few days ago. Did it not hit pgsql-patches? Gavin On Tue, 13 Aug 2002, Tom Lane wrote: > Emmanuel Charpentier <charpent@bacbuc.dyndns.org> writes: > > What do you think ? > > I think Gavin Sherry is already working on this. > > regards, tom lane > > ---------------------------(end of broadcast)--------------------------- > TIP 4: Don't 'kill -9' the postmaster >
On Wed, 2002-08-14 at 04:23, Emmanuel Charpentier wrote: > Hannu Krosing wrote: > > > > I'm trying to propose a scenario where > > > > 1. The SELECT clause defining the view is preserved > > > > 2. DROP of undrlying table/column will _not_ drop the view, but just > > mark it dirty > > > > 3. Using the view checks for the dirty flag and if it is set tries to > > recreate the view from its plaintext definition. > > I might be dense, but why not try to recreate it directly after the > table/column modification ? If it is a DROP TABLE/CREATE TABLE sequence you have no idea that you have to recreate a view. ----------- Hannu
Hannu Krosing wrote:
> On Wed, 2002-08-14 at 04:23, Emmanuel Charpentier wrote:
>
>>Hannu Krosing wrote:
>>
>>>I'm trying to propose a scenario where
>>>
>>>1. The SELECT clause defining the view is preserved
>>>
>>>2. DROP of undrlying table/column will _not_ drop the view, but just
>>>mark it dirty
>>>
>>>3. Using the view checks for the dirty flag and if it is set tries to
>>>recreate the view from its plaintext definition.
>>
>>I might be dense, but why not try to recreate it directly after the
>>table/column modification ?
>
>
> If it is a DROP TABLE/CREATE TABLE sequence you have no idea that you
> have to recreate a view.
Right. But I was wary of delaying recreation : views are more often than
not created by programmers/DBAs/someone somewhat competent in DB design
and use, in order to be used by people not necessarily aware of the real
struxture of data (that's the whole point of having views, BTW).
Delaying recreation entails the risk of overlooking a problem and
getting a nice phone call at 2 AM from the maintainance guy stuttering
that he can no longer access its (vital, of course) data ...
Tradeoffs, again ... What about emitting warnings after table drop
(easy)/creation (not so easy !) ?
BTW : since drop column and alter various attributes (not null, primary
key, etc ...) will be possible, shoudn't the need to drop/recteate a
table drastically decrease ?
E. g. : I recently created a log table wit a field date timestamptz
default now(), only to discover that, due to current limitations of the
ODBC driver, I should have used timestamptz[0] (ODBC doesn't like
fraction of seconds in datetime). I kludged away bby updating (set
date=date_trunc('second',date)) and altering default to
date_trunc('second',date) (Yuck !), but the real solution would have
been of course to recreate the column with the right attribute, which
currently involves dropping/recreating the table, therefore losing all
defined views.
What a ten-thumbs programmer such as me would love to see in such a
scenario would be something along the lines of :
# Create table T (date as timestamp defailt now(), ...) ...;
CREATE
# Create view X as select date, ... from T join ...;
CREATE
# Create view Y as select <anthing but date> ... from T where ...;
CREATE
Create view Z as select date, ... from T join ...;
# CREATE
Create view U as select ... from Z left outer join ...;
---
--- Insert data here
---
...
---
--- Later ! Insert ODBC epiphany here
---
# alter table T add column newdate timestamptz[0];
ALTER --- I can't remember the exact acknowledgement sent for alter column
update T set newdate=date;
UPDATE (somenumber) 0
alter table T rename column date to olddate;
ALTER --- ditto
WARNING : View X might have become invalid. Please check it or drop it !
WARNING : View Z might have become invalid. Please check it or drop it !
WARNING : View U might have become invalid. Please check it or drop it !
alter table T rename newdate to date;
ALTER --- ditto;
WARNING : View X successfully recreated from it's original SQL
definition. Please check it or drop it !
WARNING : View Z successfully recreated from it's original SQL
definition. Please check it or drop it !
WARNING : View U successfully recreated from it's original SQL
definition. Please check it or drop it !
Alter table T drop column olddate;
ALTER
Exercise left for the reader : what about inheritance ?
Another exercise : what about adding/dropping indices (indexes ?) ?
Your thoughs ?
__
Emmanuel Charpentier
> Tradeoffs, again ... What about emitting warnings after table drop > (easy)/creation (not so easy !) ? The warnings are certainly there now. Dependency code won't let you do such a thing without specifying CASCADE. Hopefully CREATE OR REPLACE VIEW will be applied soon, which solves part two of the problem.
Gavin Sherry <swm@linuxworld.com.au> writes:
> I submitted a patch for this a few days ago. Did it not hit pgsql-patches?
It did ... I had some gripes about it ... I thought you were working on
the gripes?
regards, tom lane
Ah, that's why it wasn't in my mailbox. Gavin, http://groups.google.com/groups?hl=en&lr=&ie=ISO-8859-1&q=replace+view+gavin&btnG=Google+Search&meta=group%3Dcomp.databases.postgresql.* Gavin Sherry <swm@linuxworld.com.au> writes: > It passes all regression tests. There's only one really sketchy part of > the patch: UpdateAttributeTuples(). This routine is fairly dangerous since > it simply removes a given relid's pg_attribute entries and creates a new > set basic on a given TupleDesc. Naturally, it is only useful for views. You can NOT allow CREATE OR REPLACE VIEW to change the tupledesc of the view, so I stopped reading right here --- take it out and install prevention instead. Why do you think that REPLACE VIEW is interesting? It's so you can modify a view without breaking things that depend on it ... and things that depend on it depend on the tupledesc. This is exactly analogous to not allowing REPLACE FUNCTION to change the return type of the function. regards, tom lane --------------------------------------------------------------------------- Tom Lane wrote: > Gavin Sherry <swm@linuxworld.com.au> writes: > > I submitted a patch for this a few days ago. Did it not hit pgsql-patches? > > It did ... I had some gripes about it ... I thought you were working on > the gripes? > > regards, tom lane > > ---------------------------(end of broadcast)--------------------------- > TIP 6: Have you searched our list archives? > > http://archives.postgresql.org > -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001+ If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania19073