Eli Collins created HDFS-3712:
---------------------------------
Summary: Support distcp from secure to insecure clusters
Key: HDFS-3712
URL: https://issues.apache.org/jira/browse/HDFS-3712
Project: Hadoop HDFS
Issue Type: Bug
Reporter: Eli Collins
Users currently can't distcp from secure to insecure clusters.
Relevant background from ATM:
There's no plumbing to make the HFTP client use AuthenticatedURL in the case
security is enabled. This means that even though you have the servlet filter
correctly configured on the server, the client doesn't know how to properly
authenticate to that filter.
The crux of the issue is that security is enabled globally instead of per-file
system. The trick of using HFTP as the source FS works when the source is
insecure, but not the source is secure.
Normal cp with two hdfs:// URL can be made to work. There is indeed logic in
o.a.h.ipc.Client to fall back to using simple authentication if your client
config has security enabled (hadoop.security.authentication set to "kerberos")
and the server responds with a response for simple authentication. Thing is,
there are at least 3 bugs with this that I bumped into. All three can be worked
around.
1) If your client config has security enabled you *must* have a valid Kerberos
TGT, even if you're interacting with an insecure cluster. The hadoop client
unfortunately tries to read the local ticket cache before it tries to connect
to the server, and so doesn't know that it won't need Kerberos credentials.
2) Even though the destination NN is insecure, it has to have a Kerberos
principal created for it. You don't need a keytab, and you don't need to change
any settings on the destination NN. The principal just needs to exist in the
principal database. This is again because the hadoop client will, before
connecting to the remote NN, try to get a service ticket for the hdfs/f.q.d.n
principal for the remote NN. If this fails, it won't even get to the part where
it tries to connect to the insecure NN and falls back to simple auth.
3) Once you get through problems 1 and 2, you will try to connect to the
remote, insecure NN. This will work, but the reported principal name of your
user will include a realm that the remote NN doesn't know about. You will
either need to change the default_realm setting in /etc/krb5.conf on the
insecure NN to be the same as the secure NN, or you will need to add some
custom hadoop.security.auth_to_local mappings on the insecure NN so it knows
how to translate this long principal name into a short name.
Even with all these changes, distcp still won't work since the first thing it
tries to do when submitting the job is to get a delegation token for all the
involved NNs, which won't work since the insecure NN isn't running a DT secret
manager. I haven't been able to figure out a way around this, except to make a
custom distcp which doesn't necessarily do this.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
