Hi Robert,

forgive me if I’m wrong, but so far as I understand Flink uses nearly the same 
model as HDFS (not at all). Means the master receives an action and distribute 
that to the workers (more or less ;)) 
HDFS as example uses not an push mechanism, the DN clients fetch the token from 
the NN when they need them. Could that be a solution, too?

https://svn.apache.org/repos/asf/hadoop/common/branches/branch-1/src/hdfs/org/apache/hadoop/hdfs/tools/DelegationTokenFetcher.java
 
<https://svn.apache.org/repos/asf/hadoop/common/branches/branch-1/src/hdfs/org/apache/hadoop/hdfs/tools/DelegationTokenFetcher.java>

MapReduce gets the token from the JT with getDelegationToken()

http://hadoop.apache.org/docs/r2.5.2/api/org/apache/hadoop/mapreduce/Cluster.html#getDelegationToken(org.apache.hadoop.io.Text)
 
<http://hadoop.apache.org/docs/r2.5.2/api/org/apache/hadoop/mapreduce/Cluster.html#getDelegationToken(org.apache.hadoop.io.Text)>


BR,
 Alexander 


> On 12 Feb 2015, at 15:28, Robert Metzger <[email protected]> wrote:
> 
> Hi,
> 
> I'm a committer at the Apache Flink project.
> One of our users asked for adding support for reading from a secured HDFS 
> cluster.
> 
> Flink has a master-worker model. Since its not really feasible for users to 
> login with their kerberos credentials on all workers, I wanted to acquire the 
> security token on the master and send it to all workers.
> For that, I wrote the following code to get the tokens in to a byte array:
> 
> UserGroupInformation.setConfiguration(hdConf);
> Credentials credentials = new Credentials();
> UserGroupInformation currUsr = UserGroupInformation.getCurrentUser();
> 
> Collection<Token<? extends TokenIdentifier>> usrTok = currUsr.getTokens();
> for(Token<? extends TokenIdentifier> token : usrTok) {
>    final Text id = new Text(token.getIdentifier());
>    credentials.addToken(id, token);
> }
> DataOutputBuffer dob = new DataOutputBuffer();
> credentials.writeTokenStorageToStream(dob);
> dob.flush();
> However, the collection currUsr.getTokens() is empty, hence the output buffer 
> doesn't contain much data.
> I suspect that I didn't fully understand the Hadoop security concepts yet.
> It would be great if somebody from the list could clarify how to properly 
> acquire the tokens.
> 
> Also, I was wondering if there is any document describing how the 
> UserGroupInformation class is working (when is it loading the credentials, 
> does it only work for Kerberos, ...)
> 
> Best,
> Robert
> 

Reply via email to