[CentOS] Re: Why is yum not liked by some? -- CVS analogy (and why you're not getting it)

Fri Sep 9 18:28:33 UTC 2005
Bryan J. Smith <b.j.smith at ieee.org>

Les Mikesell <lesmikesell at gmail.com> wrote:
> Please quantify easy.  Will you do it for me every time it
> needs to be done?  Today I'd have a use for at least 6
> variations, although I guess you'd double that with the
> suggested overlap of testing/staging instances.

So, in other words, you want a service that provides custom
tagging, revisioning and/or date-based retrieval.  That
includes your wish for a dynamic delta'ing repository and
real-time RPM generation.

Again, I don't think you understand what delta'ing systems
like CVS does compared to just a "HTTP accessable" repository
trees.  A world of difference!

> With a little thought about the process,

Again, I don't think you understand how delta'ing systems
like CVS work, and the massive mindshift in the "back-end"
that is required.  It's no "little thought about the
process."

> yum updates could be made to be repeatable without extra
> work, network traffic or any other overhead.

That's utter BS!  Delta'ing is the _worst_ overhead!
It works fine for textual data of a few MBs, but when you
start rebuilding tens of MBs, you're going to kill your
server after just a few clients!

> I just don't see why this is not considered desirable.

I _never_ said it wasn't desirable.
I just said it is _not_ feasible.

I have maintained version control systems for _large_
engineering components in my time -- everything from models
to IC schematics.  No matter how much you "break down" the
files into smaller files, you still put a _lot_ of data
around.

And that means I either have a massive Sun (and now Opteron)
box that does major I/O with a massive amount of memory for
"server-side" delta'ing assembly/disassembly, or a have the
same for NFS performance for "client-side" delta'ing
assembly/disassembly.

And that's _before_ we even get to the point of the "added
delay" that users will see using the "YUM" or whatever
client-side tool.  It will take signficantly longer to
resolve things -- regardless of who does it.

Although client-side resolution will be a crapload slower
over the Internet than server.  Which means these "servers"
will need to be "intelligent" and take a crapload more load
just for the resolution (even before we get to the actual
package delta'ing/services) than just "dumbly" serving files
up via HTTP.

Sorry, I've just built too many TB-sized engineering revision
control repositories to even listen to this thread any
longer.  Revision control exponentially increases the load
over just serving files whole.  That's tolerable on small
textual files, but intolerable on larger binaries (no matter
how small you break down things) -- especially when tens of
clients are hitting the server.



-- 
Bryan J. Smith                | Sent from Yahoo Mail
mailto:b.j.smith at ieee.org     |  (please excuse any
http://thebs413.blogspot.com/ |   missing headers)