On 20/6/19 1:12 μ.μ., Jehan-Guillaume (ioguix) de Rorthais wrote:
On Thu, 20 Jun 2019 08:37:21 +0300
Achilleas Mantzios <achill@xxxxxxxxxxxxxxxxxxxxx> wrote:
On 20/6/19 12:40 π.μ., Pavan Kumar wrote:
Hello Rui,
yes those are committed transaction.
I have tested two cases
1. complete lost. (means complete data directory, pg_wal direcoty)
2. lost few directories (deleted few directories , except pg_wal).
both cases I do see latest data in the current wal is lost.
when I check the recovery process , wal archives are copying from
pgbackrest repository to pg_wal directory. however pgbackrest tool is not
copying current wal log at all. it is only copying archived wal's/
is there any way to copy current wal log with pgbackrest tool?
No tool that is based on archiving is able to do that. Incomplete wals don't
get archived till they are complete. So, you have the risk of losing
(maximum) 16MB worth of data.
No, this could be more than 16MB. Archiving might be lagging as well depending
on various circumstances and activity.
In this case we have more than one problem.
In such cases pg_wal will be quickly stuffed with unarchived files, this indicates another (usually network or some sudden unforeseen spike in activity) problem.
So yes, in such cases where we have a network or performance problem AND at the same time disaster in the cluster occurs, then it would be much more than 16MB, but then again we are having two
problems now to solve (DB loss + infrastructure).
If you want to minimize the effect of this, you
should set
https://www.postgresql.org/docs/11/runtime-config-wal.html#GUC-ARCHIVE-TIMEOUT
(|archive_timeout|) to a smaller value .
This will still allow data loss.
I said minimize.
If you don't want a secondary in sync, you might want to have a look at
pg_receivewal.
++
--
Achilleas Mantzios
IT DEV Lead
IT DEPT
Dynacom Tankers Mgmt
