Postgresql 7.4 migration to (partially) new disks

Στις Παρασκευή 15 Σεπτέμβριος 2006 13:44, ο/η Achilleas Mantzios έγραψε:

Hi,

Our main postgresql/jboss/lotus notes server is configured as follows.

OS : Debian GNU linux 3.0
PgSQL: 7.4.13

The FS structure of the system has as follows:

Filesystem 1k-blocks Used Available Use% Mounted on
/dev/sda1 9614116 6528132 2597612 72% /
/dev/sdf 10321208 5801628 3995292 60% /raid2
/dev/sdg 6956424 4720060 1882988 72% /raidlog

where /dev/sda1 is the boot SCSI disk, while
/dev/sdf and /dev/sdg reside on two external EMC logical disks connected
with qlogic interfaces.

PgSQL is installed at the default location /usr/local/pgsql,
data is on the ~ of postgres user : /var/lib/pgsql/data

My main DB's (dynacom) data are held in
$PGDATA2 location at
/raid2/var/lib/postgres-data

also the commit log and transaction log directories are linked to:
pg_clog -> /raidlog/sma/var/lib/pgsql/data/pg_clog
and
pg_xlog -> /raidlog/sma/var/lib/pgsql/data/pg_xlog

We have planned to do the following task on this Sunday:
Migrate from Debian GNU linux 3.0 to SUSE SLES 9 (thats just a wierd lotus
notes "requirement"), and the sysadms have decided to do that on the same
HW, EMC disk arrays, by only replacing the root (/) disk.

The normal (safe) way to do that would be by following the normal
backup/install/configure/restore path.

However (just with any upgrade, and with lotus notes things get really
scary at times), there is always the possibility that the whole upgrade
procedure holds untill monday morning, when there would be an order from
our boss to rollback to the old system, or maybe repeat the same procedure
every night of the next days of the week until we succeed in Lotus Notes
upgrade!

In this scenario,If my new suse pgsql installation was some hours alive at
the meantime,
i would have to do the whole reverse backup/restore procedure again,
and this normally takes several minutes to comlete.
The DB is something about 2.5 Gbytes on .sql dump and 6 Gbytes on disk.

So one thought passing thru was to keep the alive postgresql dirs without
dumps/restores. That is to just retain the whole pgsql
directory /var/lib/pgsql/data on both systems, by copying back and fourth,
while leaving data $PGDATA2 (/raid2/var/lib/postgres-data) on the same EMC
disks.

I guess (from the zero responses) that maybe i was not clear enuf, so
i will try to set the question like this:

What considerations should someone take if he is about to just
cp -r <path_to_some_other_pgsql_data_path>/data .
.i.e. creating the PGDATA dir *not* following the standard initdb procedure

Is there any implication having to do with timestamps? modification dates?
ownerships? other?

if i just do
cp -a <path_to_some_other_pgsql_data_path>/data .
(-a implies -R)
and then
chown -R postgres:postgres ./data
(new postgres user on the new SUSE might have a different uid/gid than the
original Debian),

then will i have a healthy data dir
for postgresql to reliably work?

I'd suppose this might work on a test system, but should someone
rely on it for a critical production system?

Thanx

That is no backup restore at all.
(Providing ofcourse that the /var/lib/pgsql/data owner/group are also to be
setup correctly).

Does any one have done anything similar with (long term success),
or does anyone sees any potential problems with the later approach?

Thanx a lot for any thoughts.

--
Achilleas Mantzios

Achilleas Mantzios <achill@matrix.gatewaynet.com> writes:

What considerations should someone take if he is about to just
cp -r <path_to_some_other_pgsql_data_path>/data .
.i.e. creating the PGDATA dir *not* following the standard initdb procedure

If you copy the *entire* data tree (data, xlog, clog, any outside
tablespaces), and use exactly the same Postgres executables at both
ends, then it should pretty much Just Work.

chown -R postgres:postgres ./data
(new postgres user on the new SUSE might have a different uid/gid than the
original Debian),

This is OK.

I'd suppose this might work on a test system, but should someone
rely on it for a critical production system?

As long as the DBA has practiced beforehand ;-), sure.

regards, tom lane

Στις Παρασκευή 15 Σεπτέμβριος 2006 17:47, ο/η Tom Lane έγραψε:

Achilleas Mantzios <achill@matrix.gatewaynet.com> writes:

What considerations should someone take if he is about to just
cp -r <path_to_some_other_pgsql_data_path>/data .
.i.e. creating the PGDATA dir *not* following the standard initdb
procedure

If you copy the *entire* data tree (data, xlog, clog, any outside
tablespaces), and use exactly the same Postgres executables at both
ends, then it should pretty much Just Work.

Thats what i feared from the begining. (some connection between the data dir
and the executable, really, why does the executable need to be exactly the
same,
if we are talking about minor postgresql upgrade 7.4.12->7.4.13?)

The old system is kind of old, and there would be libraries
incompatibilities if a was to retain the old executable.

So i'll do it the formal way, and pray lotus notes will behave :)
Thanx.

chown -R postgres:postgres ./data
(new postgres user on the new SUSE might have a different uid/gid than
the original Debian),

This is OK.

I'd suppose this might work on a test system, but should someone
rely on it for a critical production system?

As long as the DBA has practiced beforehand ;-), sure.

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 2: Don't 'kill -9' the postmaster

--
Achilleas Mantzios

Achilleas Mantzios <achill@matrix.gatewaynet.com> writes:

If you copy the *entire* data tree (data, xlog, clog, any outside
tablespaces), and use exactly the same Postgres executables at both
ends, then it should pretty much Just Work.

Thats what i feared from the begining. (some connection between the data dir
and the executable, really, why does the executable need to be exactly the
same,
if we are talking about minor postgresql upgrade 7.4.12->7.4.13?)

Well, a minor version update won't in itself break anything; what would
break things is changing certain build options (such as
integer-datetimes) or switching to a different architecture (such as
x86 -> x86_64).

regards, tom lane