Hi,
I'm currently setting up a Hadoop cluster on EC2, and everything works just fine when accessing the cluster from inside EC2, but as soon as I try to do something like upload a file from an external client, I get timeout errors like:

12/08/23 12:06:16 ERROR hdfs.DFSClient: Failed to close file /user/some_file._COPYING_
java.net.SocketTimeoutException: 65000 millis timeout while waiting for channel to be ready for connect. ch : java.nio.channels.SocketChannel[connection-pending remote=/10.123.x.x:50010]

What's clearly happening is my NameNode is resolving my DataNode's IPs to their internal EC2 values instead of their external values, and then sending along the internal IP to my external client, which is obviously unable to reach those. I'm thinking this must be a common problem. How do other people deal with it? Is there a way to just force my name node to send along my DataNode's hostname instead of IP, so that the hostname can be resolved properly from whatever box will be sending files?

Eli

I've seen a bunch of people with this exact same question all over Google with no answers. I know people have successful non-temporary clusters in EC2. Is there really no one that's needed to deal with having EC2 expose external addresses instead of internal addresses before? This seems like it should be a common thing.

On Aug 23, 2012, at 12:34 PM, igor Finkelshteyn wrote:

> Hi,
> I'm currently setting up a Hadoop cluster on EC2, and everything works just fine when accessing the cluster from inside EC2, but as soon as I try to do something like upload a file from an external client, I get timeout errors like:
>
> 12/08/23 12:06:16 ERROR hdfs.DFSClient: Failed to close file /user/some_file._COPYING_
> java.net.SocketTimeoutException: 65000 millis timeout while waiting for channel to be ready for connect. ch : java.nio.channels.SocketChannel[connection-pending remote=/10.123.x.x:50010]
>
> What's clearly happening is my NameNode is resolving my DataNode's IPs to their internal EC2 values instead of their external values, and then sending along the internal IP to my external client, which is obviously unable to reach those. I'm thinking this must be a common problem. How do other people deal with it? Is there a way to just force my name node to send along my DataNode's hostname instead of IP, so that the hostname can be resolved properly from whatever box will be sending files?
>
> Eli

Hi Igor,

Amazon offers a service where you can have a VPN gateway on your network
that leads directly back to the network where youre instances are at.  So
that 10.123.x.x subnet would be connected off of the VPN gateway on your
network and you'd set up your routers/routing to push traffic for that
subnet at the gateway.

On Thu, Aug 23, 2012 at 12:34 PM, igor Finkelshteyn <[EMAIL PROTECTED]>wrote:

> Hi,
> I'm currently setting up a Hadoop cluster on EC2, and everything works
> just fine when accessing the cluster from inside EC2, but as soon as I try
> to do something like upload a file from an external client, I get timeout
> errors like:
>
> 12/08/23 12:06:16 ERROR hdfs.DFSClient: Failed to close file
> /user/some_file._COPYING_
> java.net.SocketTimeoutException: 65000 millis timeout while waiting for
> channel to be ready for connect. ch :
> java.nio.channels.SocketChannel[connection-pending remote=/10.123.x.x:50010]
>
> What's clearly happening is my NameNode is resolving my DataNode's IPs to
> their internal EC2 values instead of their external values, and then
> sending along the internal IP to my external client, which is obviously
> unable to reach those. I'm thinking this must be a common problem. How do
> other people deal with it? Is there a way to just force my name node to
> send along my DataNode's hostname instead of IP, so that the hostname can
> be resolved properly from whatever box will be sending files?
>
> Eli

That would work, but wouldn't a much simpler solution just be to force the machines in the cluster to always pass around their external FQDNs, since those will properly resolve to the internal or external IP depending on what machine is asking? Is there no way to just do that?
On Aug 23, 2012, at 8:02 PM, Aaron Eng wrote:

> Hi Igor,
>
> Amazon offers a service where you can have a VPN gateway on your network that leads directly back to the network where youre instances are at.  So that 10.123.x.x subnet would be connected off of the VPN gateway on your network and you'd set up your routers/routing to push traffic for that subnet at the gateway.  
>
> On Thu, Aug 23, 2012 at 12:34 PM, igor Finkelshteyn <[EMAIL PROTECTED]> wrote:
> Hi,
> I'm currently setting up a Hadoop cluster on EC2, and everything works just fine when accessing the cluster from inside EC2, but as soon as I try to do something like upload a file from an external client, I get timeout errors like:
>
> 12/08/23 12:06:16 ERROR hdfs.DFSClient: Failed to close file /user/some_file._COPYING_
> java.net.SocketTimeoutException: 65000 millis timeout while waiting for channel to be ready for connect. ch : java.nio.channels.SocketChannel[connection-pending remote=/10.123.x.x:50010]
>
> What's clearly happening is my NameNode is resolving my DataNode's IPs to their internal EC2 values instead of their external values, and then sending along the internal IP to my external client, which is obviously unable to reach those. I'm thinking this must be a common problem. How do other people deal with it? Is there a way to just force my name node to send along my DataNode's hostname instead of IP, so that the hostname can be resolved properly from whatever box will be sending files?
>
> Eli
>