Hi Susan, Andy, all,

there would be two datastore - "local" and "ephemeral".
Local implements and owns foo, instantiated as you describe below.  This is the 
golden copy.
Furthermore, for simplicity sake, assume that list foo is contained in a 
container, foos.

The ephemeral datastore implements a separate model "foo-ext".
This model declares a reference / "mount" to "foos" in the local datastore.
In addition, it defines some other data, that is "owned" by the ephemeral 
datastore.  (This data can have conditions/constraints to data that has been 
mounted from foo.)

So, in the ephemeral you might have

foo-ext
+-- M foos  --> /local/foos
+-- bar

An application/client that accesses ephemeral can access foo 42 via 
/foo-ext/foos/foo 42.  The fact that the ephemeral datastore turns around as 
needed, as the golden copy is owned by local, is transparent to the 
application.  If someone changes foo 42, e.g. because they access the local 
datastore directly, that will be reflected in the ephemeral datastore by virtue 
of the fact that it in effect acts as a "cache" to the data in the local 
datastore.

Now, there are a couple of items to consider:

-          If you want to have a list where some list elements reside on one 
datastore, others in another, you cannot do that directly.  What you need to do 
in that case is to introduce a partition (as part of the model that you 
implement in the ephemeral) that e.g. has a "my-foos" container containing the 
list elements it authoritatively owns, and a separate "cached-foos" container 
that "contains" the list elements from the other datastore.

-          Should you be able to have a remote data item contain a local one?  
I would argue that you should not allow for that, precisely to avoid the 
scenario in which the local object get "decapitated" because something happens 
to the containing objects that are owned by a different system.  If you want to 
attach a piece of ephemeral data to the remote data item, you should do so with 
other ways than containment, e.g. adding a leafref to the mounted/referenced 
remote object.

The distinction to other proposals is that with this proposal, the ephemeral 
datastore defines its own separate model (which indicates which data it "owns", 
vs which data it "caches"). Note that the model for the cached data does not 
need to be redefined; it is used as-is; the fact that there is another 
datastore which decides to incorporate that information is not something the 
local datastore needs to be aware of.

Does this answer your questions?
--- Alex


From: Susan Hares [mailto:[email protected]]
Sent: Friday, September 26, 2014 8:30 AM
To: Alexander Clemm (alex); 'Andy Bierman'; 'Martin Bjorklund'
Cc: 'Jeffrey Haas'; [email protected]; Eric Voit (evoit)
Subject: RE: [i2rs] Two thoughts on an ephemeral data store

Alexander:

To help me understand your proposal, let us move to the previous example:

Previous example:

  list foo {
    key id;
    leaf id { type int32; }
    leaf a { type int32; }
  }

local config:

   foo 42

In ephemeral config we now do SET /foo[id=42]/a  to 4711.  Thus, in
ephemeral we now have a single node (a) with value 4711.

Your model:

a)  The value 4711 - would be a "non-golden" copy.

b)  It would not be written through to the the "golden copy".

Questions:

a)  How would you specify a "write-through" copy?

b)  Can I2RS decide later it wants to write-through the golden copy?  Or are 
variables only "non-write-through" and "write-through"?

c)  If we delete foo 42, what happens?

Thank you,
Sue Hares

From: i2rs [mailto:[email protected]] On Behalf Of Alexander Clemm (alex)
Sent: Thursday, September 25, 2014 8:57 PM
To: Andy Bierman; Martin Bjorklund
Cc: Jeffrey Haas; [email protected]<mailto:[email protected]>; Eric Voit (evoit)
Subject: Re: [i2rs] Two thoughts on an ephemeral data store

FWIW, I would think the semantics should be kept simple.

The complexity here comes from the fact that there are dependencies between 
different data stores, and some objects that are part of one data store need to 
be reflected in a different data store.

It would seem this can be addressed with two fairly simple principles:

(a)    A datastore needs to have a clear way to reference objects in a 
different datastore, really have them incorporated into the same namespace.

(b)   It needs to be clear who owns the "golden copy" of an object.  I needs to 
be clear which objects are "authoritatively owned" by a datastore vs which ones 
are reference.  This is the datastore where the object is maintained, updated, 
created; this is where conditions and constraints are evaluated, etc.

Where an ephemeral datastore has dependencies on data in another datastore, it 
should incorporate these other objects "by reference".  The objects that are 
authoritatively owned by the ephemeral datastore can refer to those objects, 
have them referred to in conditions and constraints, and so on.  (This can also 
indicate which ephemeral objects are to be removed when an object in the other 
datastore they depend on is deleted, etc)

Changes to the non-ephemeral objects (e.g. the running datastore) have to be 
made to the "golden copy", i.e. the owning datastore.  One way to do that 
involves implementing a "write-through" operation, in which an update to an 
ephemeral copy of the object is realized by having the server of the ephemeral 
datastore turn around and make a corresponding request at the other datastore.

Very simple semantics.  I think this is preferrable to have different copies of 
the same object in different datastores, requiring "logical anding" (or other 
inter-datastore arithmetic) of different copies representing the same object to 
figure out what actual value is in effect, etc.

In the netmod WG, we have today posted a draft for what we refer to as 
requirements for a peer mount 
(http://www.ietf.org/internet-drafts/draft-voit-netmod-peer-mount-requirements-00.txt),
 motivating why a it would be useful to have a capability to "mount" subtrees 
from a remote datastore into a local datastore, and the requirements that such 
a capability needs to address.  While the original use case and motivation 
described there are somewhat different, it seems applicable to the discussion 
here.

--- Alex

From: i2rs [mailto:[email protected]] On Behalf Of Andy Bierman
Sent: Thursday, September 25, 2014 8:44 AM
To: Martin Bjorklund
Cc: Jeffrey Haas; [email protected]<mailto:[email protected]>
Subject: Re: [i2rs] Two thoughts on an ephemeral data store



On Thu, Sep 25, 2014 at 7:39 AM, Martin Bjorklund 
<[email protected]<mailto:[email protected]>> wrote:
Jeffrey Haas <[email protected]<mailto:[email protected]>> wrote:
> On Thu, Sep 25, 2014 at 12:29:35PM +0200, Martin Bjorklund wrote:
> > Jeffrey Haas <[email protected]<mailto:[email protected]>> wrote:
> > > In the proposed overlay model, presume that we have ephemeral data from a
> > > model that lives within an augmentation to a local config model.  In other
> > > words, the ephemeral nodes are children of the local config nodes.
> > >
> > > Presume, per discussion, that the local config lives in the "config" data
> > > store and that the ephemeral config - the augmenting nodes above - live in
> > > the ephemeral data store.
> > >
> > > If we delete the container in the local config that the epehemeral config 
> > > is
> > > augmenting, is there any expectation that such a deletion should carry
> > > through to the ephemeral config?
> > >
> > > Per the netmod interim discussion, probably not.
> >
> > My interpretation of the interim discussion is that the deletion
> > carries through.
>
> To be clear what I meant, consider:
>
> local config:      ephemeral:
> A                  A/B - B is introduced as an augmentation of A

I think there might be a terminology confusion here, so let's do a
simple example.

  list foo {
    key id;
    leaf id { type int32; }
    leaf a { type int32; }
  }

local config:

   foo 42

In ephemeral config we now do SET /foo[id=42]/a  to 4711.  Thus, in
ephemeral we now have a single node (a) with value 4711.

What happens if we in local config delete foo 42?

If /foo[id=42]/a is NOT deleted from the ephemeral config, what is now
presented to the internal apps?


Yes -- and what is presented to a client that retrieves the ephemeral config
in a GET request? IMO, coupling the datastores does not make sense.

Your example is 1 reason I prefer the "shadow shapshot" approach.
I think the local config and client that added the "foo" entry in the ephemeral 
datastore
are meant to have different priorities.  The entries are not coupled.
One wins and the other loses (main use-case is that ephemeral wins).

Editing "foo 42" in the local config just changes what will be installed as 
local config
when the device restarts (or the ephemeral state is removed).  It should not 
change
the injected I2RS state at all.  IMO it is really important that edits stay 
within a single
datastore.




/martin

Andy

_______________________________________________
i2rs mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/i2rs

Reply via email to