On Mon, Oct 16, 2017 at 04:27:35PM +0200, Patrik Lundin wrote:
> On 2017-10-12 11:38:57, Nico Williams wrote:
> > On Thu, Oct 12, 2017 at 05:38:32PM +0200, Patrik Lundin wrote:
> > > 
> > > To summarize this tread: The --reset flag should currently not be used
> > > in a production systems since ipropd-master is unable to parse the
> > > resulting log file.
> > 
> > No, the master is perfectly able to to parse the new log.  The issue is
> > that it's not enough to ensure that slaves get full props.
> > 
> 
> A 7.4.0 ipropd-master chokes on the log when it is truncated with
> --reset as I showed in my initial message to the list:
> http://www.h5l.org/pipermail/heimdal-discuss/2017-October/000277.html:
> ===
> If the slave log has a version number greater than 0 the master will
> notice it is out of sync and attempt to update it only to fail in
> send_diffs() because kadm5_log_get_version_fd() returns
> HEIM_ERR_EOF (and this is documented behaviour for that function when
> the log is truncated and LOG_VERSION_FIRST is supplied).
> ===

Ah, OK, I'll fix this too.  Thanks.

> > > #1. Stop the current master.
> > > #2. Dump the database on the current master.
> > > #3. Load the database on the new master.
> > > #4. Do some random modification of the database on the new master via
> > >     kadmin -l in order to set the log version to at least 2.
> > 
> > Just version 1 will do, but yes.
> > 
> 
> No, if you leave it at version 1 the slave will "get in sync with
> version 1 by means of replaying log entries" which leaves it with an
> empty database:

Oh, nice failure mode.  But version 1 should suffice, so I'll make it
so.

Thanks!

Nico
-- 

Reply via email to