[Linux-HA] Explaining new features in 2.0.8

[Andrew Beekhof]

On 2/9/07, Max Hofer <max.hofer at apus.co.at> wrote:
> Could anyone point me to the right direction where i can find
> documentation about the added features of release 2.0.8?

the sad fact is that we've barely got the features from 2.0.0
documented properly
(although Novell is busily writing a nice book for us)

>
> I'm specially interested in:
> > Allow resource colocation based on node attributes other than #id
> >    (node_attribute option added to rsc_colocation constraint).
>
> -- example for use of "node_attribute" would be fine
>
> I already used node attributes with 2.0.7 which worked pretty well.
> Why do we need "node_attribute"?

its another level of indirection.

say you have two groups of servers and two resources that don't have
to be on the same machine, but must be on the same _group_ of
machines.

ie. they need to be connected to the same SAN

>
> The following example moves resource "dummy_resource" to another
> node if the disk usage of the root file system goes over 90% with the
> help of CIB attributes mapped via attrd_updater (I wrote a RA mapping
> file systems into the CIB).
>
> <constraints>
>    <!-- add bonus if we have enough disk space
>          mapping of attribute "df:/" is done with RA -->
>   <rsc_location id="on_more_freespace" rsc="dummy_resource">
>     <rule id="du_less_than_90" score="100">
>        <expression attribute="df:/" operation="defined"/>
>        <expression attribute="df:/" operation="lt" value="90"/>
>      </rule>
>   </rsc_location>
> </constraints>
>
> > PE/CRM: Binaries now support being called with "metadata" command to
> >    document the complete list of options in the CIB.
>
> Could anyone explain this? Which binaries (crm, lrm, ...)? And what advantage
> do we get?

crmd and pengine

it produces RA-like metadata explaining the various options that can
be configured and what the defaults are

(the idea is to have it be self-describing so people dont have to ask :-)

>
> > CTS/PE: Introduce CRM option "startup_fencing" to disable fencing of
> >    unseen nodes, because CTS can't handle this.
> Is this option seomwhere described in more detail?

briefly in the above metadata but otherwise no
the reason being that it's extremely dangerous and I've only ever met
one person that _really_ needed heartbeat to do exactly what the
description says.

>
> > PE: Split-off "network-delay" option from global_timeout to separate
> >    network delays from action timeouts
> Same again - could anyone describe this option?

think about it...

before there was one timer value that covered the time between when an
action was sent from the DC to another node and when the DC got an
answer back.

now we can specify one timeout value that also gets passed to the LRM,
and an additional timer that accounts for delays crossing the network
(twice) and processing time inside the CIB.

>
> > PE: Handle asymmetric clusters where RAs are not installed on all
> > nodes.
> I assume this has something to do with the "probe". How is this handled?

yes.  we'll we dont throw a fit and assume the resource is actually running.
we now recognize that it is in fact just not there.

> > PE: Allow resource stickiness and failure stickiness to change based
> > on node attributes.
> Could anyone provide a simpel example?

if you search the archives there was a long discussion with someone
who wanted to do this.  i dont recall exactly why.