[Lustre-discuss] OST outage = client freeze?

Johann Lombardi johann at sun.com
Fri Oct 17 02:34:43 PDT 2008


On Fri, Oct 17, 2008 at 10:47:16AM +0200, Lukas Hejtmanek wrote:
> So I created a file (it seems that the file is stored only to one OST, is it
> expected?), I unmounted the OST from the metadata server, I tried to read the
> file and the reading process is frozen (is is about 18 hours now).

By default, the client just waits for the OST to be available again (it
also tries to reach the OST through the failover partners, if any) and
does not return any errors to the application (needed for transparent failover).

The lustre manual provides more information about this:
http://manual.lustre.org/manual/LustreManual16_HTML/Failover.html#50642999_pgfId-5529

More particularly:
http://manual.lustre.org/manual/LustreManual16_HTML/Failover.html#50642999_pgfId-1287626
http://manual.lustre.org/manual/LustreManual16_HTML/Failover.html#50642999_pgfId-1287643

> I would expect that I/O error will be returned.

If you want to get EIO, you should use the failout mode.
This can be set at mkfs time:
mkfs.lustre ...  --param="failover.mode=failout" $dev
or with tunefs.lustre:
tunefs.lustre --writeconf --param="failover.mode=failout" $dev

Johann



More information about the lustre-discuss mailing list