On 12/14/2005 12:24 PM, Christopher Browne wrote:
> Before I submitted the RC2 tarball, I ran tests based on the "new
> generation testbed" against the following new PG builds:
> 
> 1.  PG 7.3.12 - Linux/Debian/IA-32
> 
>   test1, testddl, testschemanames, testtabnames, testutf8
> 
> 2.  PG 8.0.5 - Linux/Debian/IA-32
> 
>   test1, testddl, testschemanames, testtabnames, testutf8
> 
> 3.  PG 8.1.1 - Linux/Debian/IA-32
> 
>   test1, testddl, testschemanames, testtabnames, testutf8
> 
> Seneca Cunningham ran tests on AIX 5.3, ML3
> -bash-2.05b$ oslevel -r
> 5300-03
> 
> She ran just PG 8.1.1, with tests test1, testddl, testschemanames,
> testtabnames, testutf8
> 
> Jan has run a set of failover tests with the RC2 tarball on FreeBSD; it
> would be nice to get some further details as to what was involved there...

PG 8.1.1

All tests use a pgbench database. Node 1 is the initial origin. Node 2 
is a direct subscriber. Node 3 is either another direct subscriber or a 
cascaded subscriber of 2. These are actually 2 different test runs and I 
will change the test setup for the future into a 4 node test with two 
direct and one cascaded subscribers.

After disconnecting node 1 from the network, I fail over to node 2. This 
leaves the configuration in node 2 being origin, node 3 being a direct 
subscriber.

After drop node 1, the system is reconnected to the network and the node 
rebuilt from scratch (dropdb, createdb, ...) as a subscriber of node 2. 
When the rebuild is complete and node 1 has caught up (pgbench is 
running all the time against the current origin), a switchback is issued 
to make node 1 the origin again, leaving configuration as 1->2->3. The 
final step is chaning the provider of node 3 to node 1, leading to the 
final configuration 1->2, 1->3.

After terminating pgbench and giving the system time to catch up, all 
three databases check out equal.


Jan

-- 
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#================================================== JanWieck at Yahoo.com #