Hi all,
After your comments about how to upgrade postgres from 9.2 to 9.6 in an overloaded server I've been learning and testing streaming replication. But the info I've found about this topic is not enough for me (or I'm not able to completely understand it).
(I'm bad with terminology as I've always had a problem trying to remember the technical name of things so I apologyze if I make any mistake with the name of things.)
One of the parameters that have to be changed to get streaming
replication work is wal_archive. I've been testing with hot_standby
option, as I've found it to be the simplest way to get streaming
replication work. With this option there's not archive of WAL
files, so we only have the files that exist in pg_xlog directory.
And the number of these files is defined by wal_keep_segments.
I'm explaining the replication process, so you can tell if is
there something that I'm missing or not understanding. There are
two ways of creating a streaming replica: one is setting wal_level
to hot_standby and the other one setting it to archive.
Let's use hot_standby option and let's suppose we make
pg_basebackup that takes 1hr: after this base backup, slave has to
synchronize with the master to get the changes that have been made
in this hour. And as there is no repository for WAL files, we must
ensure that pg_xlog contains enough files to this synchronization
is achieved succesfully (wal_keep_segments). Is this right?
I've been looking at the files in this directory, in master's
PGDATA, and ls -lt gives me 62 files created in 1hr so being
optimistic this config of wal files would be enough for my slave
to synchronize with the master, am I right? (when taking this to
production scenario I would add some more files to be sure slave
doesn't miss anything)
Would it be better to use wal archiving? The final goal is to
create the replica having master and slave in the same server so I
guess archive command would be something like "cp <source>
<destination>" without any file transfer involved. I suppose
archiving folder can be in a different path than PGDATA.
Thanks for your feedback,
Ekaterina
PS: Just in case anyone wants to know, this is part of the
process of upgrade a server with 9.2 version that has no free
space in PGDATA and that can't be stopped for much time. After
asking here, the strategy to upgrade will be: replicate this DB to
a path with space to grow, switch clusters (slave becoming master
and creating a new slave to have just-in-case), and pg_upgrade
slave/new master with --link option.
