[Info-vax] HP Integrity rx2800 i4 (2.53GHz/32.0MB) :: PAKs won't load

[G Cornelius]

On 02/16/2016 09:39 PM, Stephen Hoffman wrote:
> Ayup.  That being part of the utterly absurd implementation of
> clustering through as many as 20 shared configuration files and
> variously more, and largely a result of file accretion and not
> of design, and of compatibility and not of simplicity nor
> maintainability, and certainly not of ease of use.

Someone finally admitting this.

And it is such a simple problem compared to volume shadowing or to
clustering itself.

Clustering was wonderfully transparent until we went to multisite clusters
and what was effectively a single point of failure: the shared shadow set with
all the configuration files, the one for which the initial mount would be
problematical because you did not know if you had the other site(s) at all,
and could not be sure _your_ site did not have the stale shadow member.

> It works.  When it works.   But it's a train-wreck to deal with.
> And it's a mess to extend and to update, and a mess for the system
> manager to upgrade.

Tell me about it.  For me it's been job security in that I'm the only
one who understands the SLOGICALS.COM that mounts the common disk,
a script that has gotten ever smarter over the years, first trying
to mount the virtual unit without shadow members specified in hopes
it is already mounted somewhere, retrying a few times by killing the
mount subprocess and restarting it if it is hung while allowing
OPCOM and other interventions from other nodes or even AMDS,
and ultimately dropping into an interactive subprocess to let the
operator select from various choices including going into DCL to
resolve the problem, typically a $ MOUNT/CLUSTER/CONFIRM specifying
all shadow members.

Earlier versions knew which node was in what data center and would
try to wait for maybe two data centers of three present before trying
the mount.

Ugly, and if I had specs for IO$_PACKACK, or trusted my little demo
program that verified I could use physical I/O operations to read
individual members' shadow control blocks, I could have made it
almost transparent for the case of all members visible but not
mounted. [Of course there's never a kernel mode hacker around when
you need one!]

Guess it works.  Cluster has been up 18 years since last cluster
reboot, with a few planned "walking dead" intervals when converting
to SAN storage, or from one flavor of SAN storage to another.

Cluster is now just performing a utility function or two plus providing
access to the tape database and, theoretically, the tape library, given
that the last production application has gone away, with maintenance
dropped effective this month. Yep, I'm now system manager without
portfolio.

> But then I'm being polite.

Don't sugar coat it, Hoff, tell it like it is!

George