Обсуждение: commiting transaction from outside

Hello,

my question my sound strange, but because of a bug in an application I
see the transaction was started and not commited (connection state is
IDLE in transaction). Is there a way to commit this transaction? I don't
want to loose data. I think if I close the application the transaction
would get  rollbacked.
I've tried google'ing but without results - this might mean, that there
is no way,  but I thought asking is an option..

Thank you in advance for any advice.

--
Julius Tuskenis
Programavimo skyriaus vadovas
UAB nSoft
mob. +37068233050

Julius Tuskenis <julius@nsoft.lt> wrote:

> my question my sound strange, but because of a bug in an
> application I see the transaction was started and not commited
> (connection state is IDLE in transaction). Is there a way to
> commit this transaction? I don't want to loose data. I think if I
> close the application the transaction would get  rollbacked.

I can't think of any way to issue the commit, unless the application
is running in an unusual environment which lets you break in and
issue ad hoc commands on its connection.  There is a way you could
fish out the effects of the transaction, although it might be a fair
bit of work.  Each tuple inserted or updated has the transaction ID
set as its xmin in the new tuple, and every tuple deleted or updated
has the transaction ID set as its xmax.  The old and new are
guaranteed not to go away until the transaction completes, one way
or the other.  With some clever programming you could capture the
net effect of the transaction, and duplicate that effect after the
transaction is rolled back.

Be aware that while the transaction is stuck "idle in transaction"
the cleanup of old tuples can't proceed normally; so if you're
continuing to modify any database in the cluster, it could be
accumulating bloat until you resolve this.

-Kevin

On 2011.10.03 20:57, Kevin Grittner wrote:
>
>
> I can't think of any way to issue the commit, unless the application
> is running in an unusual environment which lets you break in and
> issue ad hoc commands on its connection.  There is a way you could
> fish out the effects of the transaction, although it might be a fair
> bit of work.  Each tuple inserted or updated has the transaction ID
> set as its xmin in the new tuple, and every tuple deleted or updated
> has the transaction ID set as its xmax.  The old and new are
> guaranteed not to go away until the transaction completes, one way
> or the other.  With some clever programming you could capture the
> net effect of the transaction, and duplicate that effect after the
> transaction is rolled back.
>
> Be aware that while the transaction is stuck "idle in transaction"
> the cleanup of old tuples can't proceed normally; so if you're
> continuing to modify any database in the cluster, it could be
> accumulating bloat until you resolve this.
>
> -Kevin
>
Thank You, Kevin, Julien, Scott for the help.

Scott - your idea is worth remembering. The problem is I can not see the
tuples until they are committed, or can I ? And if I understand you
correctly I can not rely on xmin and xmax values after the commit.
Anyway Thank You for the Idea.

I solved the problem by using sql injection methods (I was lucky the
application was buggy enough).

--
Julius Tuskenis
Programavimo skyriaus vadovas
UAB nSoft
mob. +37068233050

On Oct 3, 2011, at 10:50 AM, Julius Tuskenis wrote:

> I've tried google'ing but without results - this might mean, that there is no way,  but I thought asking is an
option..

There's no way to do it, because it makes no sense. If the client app is buggy, then how do you know for sure that the
clientapp has completed the transaction, rather than making a partial update? 

--
Scott Ribe
scott_ribe@elevated-dev.com
http://www.elevated-dev.com/
(303) 722-0567 voice