My first target will be sites with existing cells, and that's why I only
discuss volumes named "efs.*".   When administrators add an existing AFS
cell to an EFS domain, they have a one-time task of setting up the top level
mount point for "efs.root".    Then, EFS will manage that volume, and it's
contents, and nothing else.

That's why asserting a simple volume naming convention up front is so
important.    If EFS can assume that volume's named "efs.*" are all under
it's management, then things are pretty easy.  We may have to make that
prefix configurable of course, but that'll be easy.

Let me put it this way.  I would like to have EFS self-managing the AFS
volumes that belong to AFS for an EFS 3.* release.   I think complete, top
to bottom AFS cell management would be a 4.* feature.  IOW, it's not around
the corner.

I think what we need to really make this conversation meaningful would be
some real world models of existing multi-cell infrastructures.   I am still
very familiar with Morgan Stanley's basic architecture, but that's a special
case and to be honest, probably the easiest one for EFS adopt to, but I
don't want to go there.... :-|

The "mystery potential clients" that no one will talk about on this public
forum are the ones we should obviously model first, and I'll follow up
offline and see if I can get something to work with.

On Sat, Sep 25, 2010 at 2:38 PM, Steven Jenkins <[email protected]>wrote:

> On Sat, Sep 25, 2010 at 11:24 AM, Phillip Moore
> <[email protected]> wrote:
> >
> > I'm splitting up the topics here, just to avoid writing a novel in one
> > email....
> > Volume names: the more I think about this, the more I think they should
> be
> > managed per-domain (EFS domain, that is).   One thing EFS *must* assert
> for
> > a single EFS domain that manages multiple cells, is consistency in the
> > directory structure, in the namespace that is /efs.  This is core
> > functionality for EFS.
> > I think that if you make the volume names per-AFS cell, it will just
> > complicate the code, and make debugging and diagnosing problems with the
> > namespace more difficult.  It will also complicate the code and the
> database
> > significantly.
> > One facet of the EFS design I am not willing to change is the conceptual
> > idea that an EFS domain is a single, high level administrative domain.
>  EFS
> > is all about bring global consistency to a large enterprise, and I think
> > this is an example of enforcing that consistency.   I am NEVER going to
> > support per-cell, per-region, etc.  releasealiases, for example (I
> predict
> > an annual argument with the user community over this one).
> > Having said that, I'll certainly listen to counter arguments for allowing
> > the volume names to be cell-specific, but I personally think they should
> be
> > global.
> >
> Ah.  I was interpreting per domain vs per cell as a different
> argument.  I completely agree that a given volume name that is 'in'
> EFS should be consistent across all cells in the domain.
>
> There needs to be some thought, though, to how migration of an
> existing infrastructure would occur.  One could, for example, simply
> build out new cells for EFS, but it's worth thinking through how we
> might introduce EFS into an existing AFS cell.  That would probably
> argue for certain volumes (e.g., root.afs, root.cell) being managed
> specially for a given cell, and then certain mountpoints existing
> under that for the 'legacy' and 'EFS' namespaces.
>
> But as you say elsewhere, we can speculate forever on how integration
> might happen.  Better is talking with people who are considering using
> it and seeing what their needs actually are.
>
> Steven
> _______________________________________________
> EFS-dev mailing list
> [email protected]
> http://mailman.openefs.org/mailman/listinfo/efs-dev
>
_______________________________________________
EFS-dev mailing list
[email protected]
http://mailman.openefs.org/mailman/listinfo/efs-dev

Reply via email to