[OmniOS-discuss] Internal pkg error during a test r151010 to r151014 upgrade

[Richard Elling]

> On Apr 7, 2015, at 9:40 AM, Chris Siebenmann <cks at cs.toronto.edu> wrote:
> 
>> History lesson: until people could afford to purchase more than one
>> disk and before Sun invented the diskless workstation (with shared
>> /usr), everything was under /.
> 
> As Richard knows but other people may not, this is ahistorical on
> Unix. From almost the beginning[*] Unix had a split between the root
> filesystem and the /usr filesystem, based (as far as I understand
> it) on the physical disks involved at Bell Labs CSRG on their Unix
> machine. This is part of why the split of commands between /bin and
> /usr/bin existed for years. Sun's diskless machines did not invent a
> split /usr, they just took advantage of existing practice and made it
> read-only and shared.

Splitting some hairs... originally the OS was in / and user programs in /usr
(hence the name) It was later that the thing we now call the "OS" moved
to /usr. Now the "OS" is moving elsewhere, invading as it goes, as Volker
described rather well.

The key point is that trying to use filesystem(5) as written only works if
everyone uses it, and they don't :-( with /opt being the perfect example of
organizational dysfunction. The packaging system doesn't matter as it just
sweeps the dust under the rug.

IMHO the companies that solve this take the reductionist path: one file
system. When done well, upgrades and installation are painless and my
grandmother has no problem upgrading her phone without assistance.

My approach to creating filesystems is based on policies to be applied,
where those policies are fundamental to filesystems. Harkening back to
the diskless workstation example or the more modern SmartOS model,
there can be a readonly policy for the fixed OS bits that are replaced en
masse. Other common policy knobs include: quota, reservation, backup,
dedup, compression.

Back to the problems at hand:
1. BE managing /opt as if it was its own, exclusive, waterfront resort.
   IMHO trying to assert an upgrade en masse policy to /opt is futile.
   Oracle's hack in Solaris 11.2 just kicks the can down the street.

2. Saving space for dumps.
   Don't waste time dumping to ZFS, setup a dump device on a raw partition
   somewhere. No need to mirror it or back it up.

 -- richard