My point was that this is a problem you need to solve in GeoMesa and it
doesn't matter how you go about doing it :). A proxy around Connector
would be one way to do this, if that's what you're exposing to your users.
Like I said earlier, you should be ensuring that any call which may
result in an RPC is wrapped in a doAs().
On 6/13/19 4:41 PM, Emilio Lahr-Vivaz wrote:
Hello,
GeoMesa is a library (providing a GeoTools data store backed by
Accumulo, among other things), so there isn't a single entry point. We
could try to wrap every method call, but that would likely be complex
(the GeoTools API has a lot of surface area).
Would it make sense to create a delegate proxy for the Accumulo
connection instance, and wrap all its methods with a doAs()? Sometimes
we create additional threads, or create a scanner but then pass it off
before reading it - would we need to wrap e.g. every call to next() on a
scanner, or would wrapping the connector call to create it be enough?
Thanks,
Emilio
On 6/13/19 4:09 PM, Josh Elser wrote:
Yes. Anything in which you're interacting with Accumulo (read-as,
anything that's going to execute an RPC to talk to the Master or a
TabletServer) needs to be wrapped in a doAs().
What's often easiest is to wrap your "entry point" in a doAs().
For example, if you had some class with a main:
public class MyGeoMesa {
public MyGeoMesa() {}
public void do() {...}
public static void main(String[] args) {
(new MyGeoMesa()).do();
}
}
You could turn that `do()` call into:
UserGroupInformation user =
UserGroupInformation.loginFromKeytabAndReturnUGI(..):
user.doAs(new PrivilegedExecution<Void>() {
public Void run() {
(new MyGeoMesa()).do();
return null;
}
});
On 6/13/19 3:38 PM, James Srinivasan wrote:
Thanks, I think that helps. If I'm no longer updating the Hadoop user,
do I have to use doAs for any operation that touches Accumulo? I fear
there are lots...
James
On Thu, 13 Jun 2019 at 20:21, Josh Elser <[email protected]> wrote:
Hi James,
A thread calling checkTGTAndReloginFromKeytab() is still what you want
for renewals. Just make sure you wrap that in a UGI.doAs() for the user
whose ticket you want to renew.
In general, you just want to wrap any Accumulo-related calls in a
doAs()
to avoid fallback onto the "loginUser" semantics that UGI has. Being
explicit with a doAs() for the user you want to run some code as is the
trick.
Does that help?
On 6/13/19 3:10 PM, James Srinivasan wrote:
Hi all,
I'm finally getting around to fixing up some deprecation issues with
our use of Kerberos with Accumulo and GeoMesa
(https://github.com/locationtech/geomesa/). Because I didn't know any
better at the time, I used the KerberosToken ctor specifying that the
Hadoop user should be replaced. Combined with a thread to periodically
renew the ticket (calling
UserGroupInformation.getCurrentUser.checkTGTAndReloginFromKeytab()),
this has worked nicely for us.
However, there are some unfortunate side effects of updating the
Hadoop user - for instance, subsequent HDFS operations use the new
user, who may not have the same permissions as the original user in a
Zeppelin-type notebook environment. Plus the replaceCurrentUser param
is deprecated and removed in Accumulo 2.0. So I'm keen on not
replacing the Hadoop user, but how do I handle ticket renewal?
Thanks very much,
James
