incubator-cassandra-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Clint Kelly <>
Subject Re: Meaning of "token" column in system.peers and system.local
Date Mon, 31 Mar 2014 15:06:49 GMT
Hi Theo,

Thanks for your response.  I understand what you are saying with
regard to the load balancing.  I posted my question to the DataStax
list and one of the folks there answered it.  I put his response below
(for anyone who may be curious):

Sylvain Lebresne

4:03 AM (4 hours ago)

to java-driver-us.
The system tables are a bit specific in the sense that they are local
to the node that coordinate the query. And by default the java driver
round robin the queries over the node of the cluster. The result is
that more likely than not, your two system queries (on system.local
and system.peers) do not reach the same coordinator, hence what you

It's possible to enforce that both query goes to the same coordinator
by mean of modifying/providing a custom load balancing policy. You
could for instance write a wrapper Statement class, that allow to
specify which node is supposed to be contacted, and then write a
custom load balancing policy that recognize this wrapper class and
force the user provided host if there is one (and say fallback on
another load balancing policy otherwise). Or, simpler but somewhat
less flexible, if all you want is to have 2 requests go to the same
coordinator (which is enough to get all tokens of a cluster really),
then you can make sure to use TokenAwarePolicy (a good idea anyway),
and make sure both query have the same "routing key" (whatever it is
is not all that important, you can use an empty ByteBuffer), see

Note that I would agree that what's suggested above is slightly
involved and could be supported more "natively" by the driver. And I
do plan on exposing the cluster tokens more simply in particular
(probably directly from the Host object, it's just "a todo not yet
done". And I'll probably add the load balancing stuff + Statement
wrapper I describe above, because that's probably somewhat generally
useful for debugging for instance.  Still, it's possible to do
currently, just a bit more involved than is probably necessary.


On Mon, Mar 31, 2014 at 3:30 AM, Theo Hultberg <> wrote:
> your assumption about 256 tokens per node is correct.
> as for you second question, it seems to me like most of your assumptions are
> correct, but I'm not sure I understand them correctly. hopefully someone
> else can answer this better. tokens are a property of the cluster and not
> the keyspace. the first replica of any token will be the same for all
> keyspaces, but with different replication factors the other replicas will
> differ.
> when you query the system.local and system.peers tables you must make sure
> that you don't connect to other nodes. I think the inconsistency you think
> you found is because the first and second queries went to different nodes.
> the java driver will connect to all nodes and load balance requests by
> default.
> T#
> On Mon, Mar 31, 2014 at 4:06 AM, Clint Kelly <> wrote:
>> BTW one other thing that I have not been able to debug today that maybe
>> someone can help me with:
>> I am using a three-node Cassandra cluster with Vagrant.  The nodes in my
>> cluster are,, and
>> If I use cqlsh to connect to, I see unique sets of tokens
>> when I run the following three commands:
>> select tokens from system.local
>> select tokens from system.peers where peer=
>> select tokens from system.peers where peer=
>> This is what I expect.  However, when I tried making an application with
>> the Java driver that does the following:
>> Create a Session by connecting to
>> From that session, "select tokens from system.local"
>> From that session, "select tokens, peer from system.peers"
>> Now I get the exact-same set of tokens from system.local and from the row
>> in system.peers in which peer=
>> Anyone have any idea why this would happen?  I'm not sure how to debug
>> this.  I see the following log from the Java driver:
>> 14/03/30 19:05:24 DEBUG com.datastax.driver.core.Cluster: Starting new
>> cluster with contact points [/]
>> 14/03/30 19:05:24 INFO com.datastax.driver.core.Cluster: New Cassandra
>> host / added
>> 14/03/30 19:05:24 INFO com.datastax.driver.core.Cluster: New Cassandra
>> host / added
>> I'm running Cassandra 2.0.6 in the virtual machine and I built my
>> application with version 2.0.1 of the driver.
>> Best regards,
>> Clint
>> On Sun, Mar 30, 2014 at 4:51 PM, Clint Kelly <>
>> wrote:
>>> Hi all,
>>> I am working on a Hadoop InputFormat implementation that uses only the
>>> native protocol Java driver and not the Thrift API.  I am currently trying
>>> to replicate some of the behavior of
>>> Cassandra.client.describe_ring(myKeyspace) from the Thrift API.  I would
>>> like to do the following:
>>> Get a list of all of the token ranges for a cluster
>>> For every token range, determine the replica nodes on which the data in
>>> the token range resides
>>> Estimate the number of rows for every range of tokens
>>> Groups ranges of tokens on common replica nodes such that we can create a
>>> set of input splits for Hadoop with total estimated line counts that are
>>> reasonably close to the requested split size
>>> Last week I received some much-appreciated help on this list that pointed
>>> me to using the system.peers table to get the list of token ranges for the
>>> cluster and the corresponding hosts.  Today I created a three-node C*
>>> cluster in Vagrant ( and
>>> tried inspecting some of the system tables.  I have a couple of questions
>>> now:
>>> 1. How many total unique tokens should I expect to see in my cluster?  If
>>> I have three nodes, and each node has a cassandra.yaml with num_tokens =
>>> 256, then should I expect a total of 256*3 = 768 distinct vnodes?
>>> 2. How does the creation of vnodes and their assignment to nodes relate
>>> to the replication factor for a given keyspace?  I never thought about this
>>> until today, and I tried to reread the documentation on virtual nodes,
>>> replication in Cassandra, etc., and now I am sadly still confused.  Here is
>>> what I think I understand.  :)
>>> Given a row with a partition key, any client request for an operation on
>>> that row will go to a coordinator node in the cluster.
>>> The coordinator node will compute the token value for the row and from
>>> that determine a set of replica nodes for that token.
>>> One of the replica nodes I assume is the node that "owns" the vnode with
>>> the token range that encompasses the token
>>> The identity of the "owner" of this virtual node is a cross-keyspace
>>> property
>>> And the other replicas were originally chosen based on the
>>> replica-placement strategy
>>> And therefore the other replicas will be different for each keyspace
>>> (because replication factors and replica-placement strategy are properties
>>> of a keyspace)
>>> 3. What do the values in the "token" column in system.peers and
>>> system.local refer to then?
>>> Since these tables appear to be global, and not per-keyspace properties,
>>> I assume that they don't have any information about replication in them, is
>>> that correct?
>>> If I have three nodes in my cluster, 256 vnodes per node, and I'm using
>>> the Murmur3 partitioner, should I then expect to see the values of "tokens"
>>> in system.peers and system.local be 768 evenly-distributed values between
>>> -2^63 and 2^63?
>>> 4. Is there any other way, without using Thift, to get as much
>>> information as possible about what nodes contain replicas of data for all of
>>> the token ranges in a given cluster?
>>> I really appreciate any help, thanks!
>>> Best regards,
>>> Clint

View raw message