Thread: Removing savepointLevel from TransactionState

Gurjeet Singh <singh.gurjeet@gmail.com> writes:
> I noticed that the savepointLevel member of TransactionStateData struct is
> initialized to 0 from TopTransactionStateData, and never incremented or
> decremented afterwards.

> Since this is a file-local struct I think we can simply get rid of all
> usages of this without any risk.

ISTM you have detected a bug, not just dead code that should be removed.
Surely those tests that throw error on savepointLevel change were
meant to do something important?
        regards, tom lane

On Thu, Sep 29, 2011 at 10:52 AM, Gurjeet Singh <singh.gurjeet@gmail.com> wrote:
> On Thu, Sep 29, 2011 at 1:11 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>>
>> Gurjeet Singh <singh.gurjeet@gmail.com> writes:
>> > I noticed that the savepointLevel member of TransactionStateData struct
>> > is
>> > initialized to 0 from TopTransactionStateData, and never incremented or
>> > decremented afterwards.
>>
>> > Since this is a file-local struct I think we can simply get rid of all
>> > usages of this without any risk.
>>
>> ISTM you have detected a bug, not just dead code that should be removed.
>> Surely those tests that throw error on savepointLevel change were
>> meant to do something important?
>
> That's surprising for dead-code removal!
>
> Not sure which failures you're pointing to. `make check` before and after
> the patch on master says 'All 126 tests passed.'
>

I think Tom is talking about the test in the code and not a regression
test itself. Looking at the archives and git log, it seems though that
we never completed this feature and the code was always dead since its
inception.

http://www.mail-archive.com/pgsql-hackers@postgresql.org/msg48321.html

Thanks,
Pavan
--
Pavan Deolasee
EnterpriseDB     http://www.enterprisedb.com

Excerpts from Tom Lane's message of jue sep 29 02:11:52 -0300 2011:
> Gurjeet Singh <singh.gurjeet@gmail.com> writes:
> > I noticed that the savepointLevel member of TransactionStateData struct is
> > initialized to 0 from TopTransactionStateData, and never incremented or
> > decremented afterwards.
> 
> > Since this is a file-local struct I think we can simply get rid of all
> > usages of this without any risk.
> 
> ISTM you have detected a bug, not just dead code that should be removed.
> Surely those tests that throw error on savepointLevel change were
> meant to do something important?

This is the patch I submitted that introduced that bit:
http://archives.postgresql.org/pgsql-patches/2004-07/msg00292.php
which got committed as cc813fc2b8d9293bbd4d0e0d6a6f3b9cf02fe32f.

Amusingly, the savepointLevel thing was introduced there; I don't
remember the details but I think what it was intended to implement is
some sort of restriction laid out by the SQL standard's spelling of
savepoint commands.

... in fact, SQL 2008 talks about savepoint levels in "4.35.2
Savepoints".  And as far as "Part 2: Foundation" is concerned, I think
only <routine invocation> can cause the savepoint level to be changed.
That is, if you have a function that declares itself to have NEW SAVEPOINT
LEVEL, then that function is not allowed to roll back savepoints that
were created before it started.

Now, we already disallow functions from doing this at all; so it seems
that the missing feature for us is OLD SAVEPOINT LEVEL (which, according
to the standard, is the default behavior).  Since this is not
implementable on the current SPI abstraction, we cannot do much about
this.  But if we ever have transaction-controlling SPs, then it seems to
me that we ought to keep this and enable those to use it as appropriate.

-- 
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support