Restoring a snapshot into a new cluster - thoughts on replica placement

["Peer, Oded" <Od...@rsa.com>]

I read the documentation for restoring a snapshot into a new cluster.
It got me thinking about replica placement in that context. "NetworkTopologyStrategy places replicas in the same data center by walking the ring clockwise until reaching the first node in another rack."
It seems it is not enough to restore the token ranges on an equal-size cluster since you also need to restore the rack information.

Assume I have two 6-node clusters with three racks for each cluster represented by "lower" "UPPER" and "numeric".
In the first cluster the ring is represented by: a -> B -> 3 -> d -> E -> 6 ->
In the second cluster the ring uses the same token ranges and  is represented by: a -> b -> C -> D -> 5 -> 6 ->

In this case data restored from the first cluster matches the token distribution on the second cluster but does not match the expected replica placement.
Token t will reside on nodes a,B,3 on the first cluster but should reside on nodes a,C,5 on the second cluster.

Does this make sense? Did I miss something?

Re: Restoring a snapshot into a new cluster - thoughts on replica placement

[Robert Coli <rc...@eventbrite.com>]

On Wed, Dec 2, 2015 at 5:06 AM, Peer, Oded <Od...@rsa.com> wrote:

> It seems it is not enough to restore the token ranges on an equal-size
> cluster since you also need to restore the rack information.
>

Yep, if you're using a rack-aware snitch, that is correct. Because in that
case, rack determines replica placement.

=Rob