Re: xlog switch & last record before the switch

Simon Riggs wrote:
> On Thu, 2007-07-12 at 16:17 +0200, Florian G. Pflug wrote:
>
>> To test my PITR-slave readonly-query patch, I continously do
>> insert into test ...
>> pg_switch_xlog()
>> sleep 1
>> on the master, and let the slave process the generated xlogs
>>
>> The log output on the slave looks the following (unnecessary lines remove)
>>
>> <snipped log output>>
>>
>> As you can see, the COMMIT records seems to end up being logged *after*
>> the xlog switch. I would have expected the order
>> "heap-insert, commit, switch, heap-insert, commit, switch, ...", not
>> "heap-insert, switch, commit, heap-insert, switch, commit, ...".
>>
>> Is this the expected behaviour, or just an artefact of the implementation
>> of xlog switches?
>
> Can you show the exact SQL executed?
>
> If you do INSERT ... ;select pg_switch_xlog() then the COMMIT will occur
> after the switch. If you do:
> INSERT ...;
> select pg_switch_xlog();
> then the COMMIT will occur before the switch.

Ah, you caught me - I had forgotten that pgsql -c "<statement>"
executes the statement in one transaction.

I was just going to suggest that pg_start_backup, pg_stop_backup
and pg_switch_xlog emit a warning or even an error if called
from within a larger transaction, because that's quite certainly
not what the user wants. But since those are just plain functions,
I guess checking for that might not be trivial...

> If it were otherwise this would imply statements were executed prior to
> the previous commit, which I hope and pray is never the case.
You can relax, the bug was located in front of the screen :-)

greetings, Florian Pflug