experiences with autocommit functionality in 7.3

From: Barry Lind <barry(at)xythos(dot)com>
To: pgsql-hackers(at)postgresql(dot)org
Subject: experiences with autocommit functionality in 7.3
Date: 2002-10-13 08:41:39
Message-ID: 3DA931C3.9020807@xythos.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers


I was spending some time investigating how to fix the jdbc driver to
deal with the autocommit functionality in 7.3. I am trying to come up
with a way of using 'set autocommit = on/off' as a way of implementing
the jdbc symantics for autocommit. The current code just inserts a
'begin' after every commit or rollback when autocommit is turned off in
jdbc.

I can continue to use the old way and just issue a 'set autocommit = on'
at connection initialization, but I wanted to investigate if using 'set
autocommit = off' would be a better implementation.

The problem I am having is finding a way to turn autocommit on or off
without generating warning messages, or without having the change
accidentally rolled back later.

Below is the current behavior (based on a fresh pull from cvs this morning):

Key: ACon = autocommit on
ACoff = autocommit off
NIT = not in transaction
IT = in transaction
IT* = in transaction where a rollback will change autocommit state

Current State Action End State
ACon and NIT set ACon ACon and NIT
set ACoff ACoff and IT*
ACon and IT set ACon ACon and IT
set ACoff ACoff and IT*
ACon and IT* set ACon ACon and IT*
set ACoff ACoff and IT
ACoff and NIT set ACon ACon and NIT
set ACoff ACoff and IT
ACoff and IT set ACon ACon and IT*
set ACoff ACoff and IT
ACoff and IT* set ACon ACon and IT
set ACoff ACoff and IT*

There are two conclusions I have drawn from this:

1) Without the ability for the client to know the current transaction
state it isn't feasible to use set autocommit = on/off in the client.
There will either end up being spurious warning messages about
transaction already in process or no transaction in process, or
situations where a subsequent rollback can undo the change. So I will
stay with the current functionality in the jdbc driver until the FE/BE
protocol provides access to the transaction status.

2) In one place the current functionality doesn't make sense (at least
to me).
ACon and NIT set ACoff ACoff and IT*

If I am running in autocommit mode and I issue a command I expect that
command to be committed. But that is not the case here. I would have
expected the result to be: ACoff and NIT

thanks,
--Barry

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Masaru Sugawara 2002-10-13 10:07:41 Re: Transactions through dblink_exec()
Previous Message Joe Conway 2002-10-13 06:44:18 Re: Transactions through dblink_exec()