On 9/4/17 13:23, Artem Tomyuk wrote: > I'am wondering, what is the fastest backup method do you know, for me it is: > (Assume that we running on 16 * CPU box) > > pg_dump -Z 4 -j 16 -F c -F d > > -Z 4 compression level > > -j 16 number of parallel jobs > > -F c, d custom+directory format
I think you misunderstood that last part. There can only be one output format, and the last option wins. So you are really running in directory format only.
As for performance, you generally have the right idea. Specifics depend on the hardware.
My experience is that any compression at all can be bad for performance. But the I/O cost might balance that in some cases.
In my experience and experimentation, using a hot standby in conjunction with PITR almost completely insulates the master from backup load. The recipe is simply obviously configure for WAL archiving, have a hot standby for backups, execute pg_start_backup(), wait on hot standby for the LSN on the hot standby, execute pg_xlog_replay_pause() on hot standby, execute pg_stop_backup() on master, do the necessary backup on hot standby, and wrap it up with pg_xlog_replay_resume() on hot standby. It's been a while so that was from memory.
I have performed this test in high transaction environment and tested the restore with much success. Aside from the checkpoint on the master and archive_command, minimal to no load incurred on the master. I don't believe xlog replay pause/resume is strictly necessary but seemed like a good idea.