Managing vnodes cluster and configuring rpc_address with EC2MultiRegionSnitch

[Alain RODRIGUEZ <ar...@gmail.com>]

Hi

We recently switched to vnodes on our production cluster. I have some
operational questions about managing a vnodes cluster :

- Can we add multiple nodes at once (The previous rule was no more than one
server per range at the same time...)
- Can we decommission multiple nodes at the same time (I used to do it when
not on the same part of the ring - taking replicas into account).
- Is there a way to distribute smartly heavy operations such as repair or
cleanup through nodes ? (Using a RF of 3, previously I just use to launch a
repair or cleanup on one node out of 3 in token order. With this I was sure
that at least 2 nodes out of 3 weren't performing a repair and were
available to answer queries fast enough, is there a way to reproduce this
behavior using vnodes ?)

If someone could point me to a "best practice in cassandra management with
vnodes" article / blog / tuto, It would be very cool.

Else, just a feed back on these points would be appreciated.

An other question :

How to configure rpc_address, using ec2MultiRegionSnitch, I mean should it
be public ip or private ip ? (I have private IP as listen_address and
public IP as broadcast_address. Seeds are public IPs)

We currently use 0.0.0.0 which seems to be incompatible with new CQLs
drivers.

Thanks for any insight on any of these points.

Alain

Re: Managing vnodes cluster and configuring rpc_address with EC2MultiRegionSnitch

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

On Tue, Nov 19, 2013 at 7:24 AM, Alain RODRIGUEZ <ar...@gmail.com>wrote:

> - Can we add multiple nodes at once (The previous rule was no more than
> one server per range at the same time...)
>

Vnodes are just like a lot of nodes. So no.


> - Can we decommission multiple nodes at the same time (I used to do it
> when not on the same part of the ring - taking replicas into account).
>

This becomes much more difficult/impossible with 256 virtual nodes per
hardware node.


> - Is there a way to distribute smartly heavy operations such as repair or
> cleanup through nodes ? (Using a RF of 3, previously I just use to launch a
> repair or cleanup on one node out of 3 in token order. With this I was sure
> that at least 2 nodes out of 3 weren't performing a repair and were
> available to answer queries fast enough, is there a way to reproduce this
> behavior using vnodes ?)
>

No, but that's a design feature of vnodes, you repair fewer ranges per node
and the repair is spread among more nodes. May or may not be a win in
practice, as I don't use vnodes I don't know!


> If someone could point me to a "best practice in cassandra management with
> vnodes" article / blog / tuto, It would be very cool.
>

I'll let you know when I deploy vnodes and then write it, lol. :)


> How to configure rpc_address, using ec2MultiRegionSnitch, I mean should it
> be public ip or private ip ? (I have private IP as listen_address and
> public IP as broadcast_address. Seeds are public IPs)
>

I don't know the answer, but are you sure you want to use
EC2MultiRegionSnitch?

https://issues.apache.org/jira/browse/CASSANDRA-3810

=Rob