Ian,
>My next test, hopefully tomorrow, will be to turn Forced Writes
off, and kill the link in the 5 second time between doing stuff >and
the OS deciding to do anything with it, but I think I'm still on a
hiding to nothing unless I can get the packets to drop part >way
through the splurge of writing.
It will be a really interested experiment - please keep us informed.
>We are not worried about HA, we are just trying to get real-time
replication for persistence of data - and I've no idea how to >kill it!
If you are interested in more ways to kill Firebird - please provide an
exact version of your Firebird, and I will try to find a script to kill
it or at least freeze. Please contact me: ak at ib-aid.com
Regards,
Alexey Kovyazin
IBSurgeon
> So, did you make experiments with sudden reboot of one of the
nodes with simultaneous high load
> (inserting or updating a lot of records)?
Hi Alexey,
That's pretty much the only test that we've tried in several
different ways. The problem is that Firebird is just too reliable,
so I don't have a mental model of how to break it. We've been
using it for 15 years and only ever had problems with the
generation of HDDs in the early 00s that reported successful write
to the OS but cached forever - specifically Maxtors. Apart from
that we had a power supply blow once on a 10GB database that
corrupted just a single record at the moment of death, and all
that took was a careful extract of the data from that table either
side of the bad record.
For testing DRBD we've tried pulling power during heavy activity,
and then repeated this with iptables dropping all traffic between
the nodes to simulate to the secondary the total immediate failure
of the primary in a more test friendly way. So far Firebird just
shrugs a bit and gets back on with the work on the secondary.
My next test, hopefully tomorrow, will be to turn Forced Writes
off, and kill the link in the 5 second time between doing stuff
and the OS deciding to do anything with it, but I think I'm still
on a hiding to nothing unless I can get the packets to drop part
way through the splurge of writing.
We are not worried about HA, we are just trying to get real-time
replication for persistence of data - and I've no idea how to kill it!
Ian