You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@cassandra.apache.org by S G <sg...@gmail.com> on 2021/09/13 17:57:48 UTC
How many coordinator nodes in the cluster at a given time?
Hello,
As mentioned in https://cassandra.apache.org/_/cassandra-basics.html, can
someone clarify how many coordinator nodes are there in the cluster at any
given time?
1. Is there a dedicated *single* coordinator node in the whole cluster
that changes from time to time?
2. Or any node becomes a coordinator node when the client-request lands
on it?
I am assuming it is the latter because the former will make the coordinator
node a SPOF.
As an example, if we have a 100 node cluster and there are no hot
partitions etc, then is it fair to say that at any given time, generally
speaking, each node will be acting as a coordinator node for an equal
number of requests?
So for each request, there is a single coordinator node.
But for all requests at a given time, the coordinator nodes are equally
distributed among the cluster.
Assumption: The client is intelligent enough and not sending traffic to
just a single node.
Thanks in advance for clarifying !
Re: How many coordinator nodes in the cluster at a given time?
Posted by Jeff Jirsa <jj...@gmail.com>.
On Mon, Sep 13, 2021 at 10:58 AM S G <sg...@gmail.com> wrote:
> Hello,
>
> As mentioned in https://cassandra.apache.org/_/cassandra-basics.html, can
> someone clarify how many coordinator nodes are there in the cluster at any
> given time?
>
> 1. Is there a dedicated *single* coordinator node in the whole cluster
> that changes from time to time?
> 2. Or any node becomes a coordinator node when the client-request
> lands on it?
>
>
> I am assuming it is the latter because the former will make the
> coordinator node a SPOF.
>
Every host with binary/native enabled can act as a coordinator.
>
> As an example, if we have a 100 node cluster and there are no hot
> partitions etc, then is it fair to say that at any given time, generally
> speaking, each node will be acting as a coordinator node for an equal
> number of requests?
>
Assuming your clients are using token-aware load balancing/routing, yes.
> So for each request, there is a single coordinator node.
>
Yes (modulo retries)
> But for all requests at a given time, the coordinator nodes are equally
> distributed among the cluster.
>
>
This is the typical behavior, yes.
> Assumption: The client is intelligent enough and not sending traffic to
> just a single node.
>
> Thanks in advance for clarifying !
>
There's an asterisk here that you CAN do an advanced model where you
DONT enable native/binary on every host, and instead only do it on a subset
of hosts, and disable storage on those hosts (-Dcassandra.join_ring=false)
and then you build effectively a stateless coordination tier in front of
the storage hosts. That's relatively rare/advanced, but strictly possible.
- Jeff