EC2中的Cassandra群集通过私有IP与自身通信。 我该如何与外界沟通？

Question

I've got a three machine cluster hosted in AWS.

• Each machine has its listen address defined as its local IP.
• Each machine has it's broadcast address listed as its local IP.
• Each machine has its RPC address listed as its local IP.
• Each machine has it's "seeds" property set as a list of the three local IPs.

The cluster starts up correctly and it communicates with itself just fine. When I use Spring Data to connect to this cluster I set the contactPoints property as a list of the three pubic IP addresses of the nodes. I can successfully connect to one node however the Cassandra Java Driver throws exceptions about trying to connect to the other two nodes via their private IPs. I suspect that that node is giving the driver internal IPs to use instead of the public IPs I gave it via contactPoints.

1. My first guess was to change the RPC address of each one to their public IP but that did not work.
2. My second guess was to change the broadcast address of each node to it's public IP but that did not work.
3. My third guess was to do both #1 and #2 and that did not work.

Is there something still missing? Was one of those supposed to work and I just messed it up?

Thanks

Answer 1

The cluster uses the broadcast_address to announce it to the other instances that they should attempt to communicate with the node over this address. By no means the cluster knows neither that the clients will be connecting to it using some sort of NAT translation (what if there is a mix of local and external clients), nor its external IP address (because it is an infrastructure, not cluster, property).

So it is a driver's responsibility to find out how to connect to the cluster. Some drivers accept just a finite list of IPs they should connect to and only connect to those.

But the Java driver has a policy called AddressTranslater that you can implement to perform translation from internal IP to external - for example by using AWS API. http://www.datastax.com/drivers/java/2.0/com/datastax/driver/core/policies/AddressTranslater.html

Here is a link to an implementation (I haven't used it myself): http://docs.hazelcast.org/docs/3.3/javadoc/com/hazelcast/client/spi/impl/AwsAddressTranslator.html

Answer 2

The broadcast_address is the address broadcast to other Cassandra nodes. It is good to set both listen_address and broadcast_address to the node's local IP. As you are going to access Cassandra via the node's public IP, you should set rpc_address and broadcast_rpc_address to the node's public IP.

But for security reason, it is better to not expose Cassandra to the public world. If your application also runs in AWS, it is better to let Cassandra listen on the local IP only, and have the application access the local IP.