Thanks Marcelo, that makes a lot of sense to me now. Do you know if there are any plans to expand Kerberos auth to the executors? The current executor behaviour is quite curious - you can see in the trace information that it consumes the jaas conf file and keytab (indeed they're required - it will fail with file not found if you put in something bogus) but then goes off and tries to use the simple auth local user credentials. I hacked in something nasty in the MS jdbc driver as an experiment - nulled out the creds it receives - it then loads up correctly from the keytab and everything works.
The UserGroupInformation thing seemed promising (don't mind distributing keytab and understand --keytab will not work). I tried something simpleminded: val remoteUgi = UserGroupInformation.loginUserFromKeytabAndReturnUGI(..., ...) remoteUgi.doAs(... df.write.jdbc(..., ...) ) but this just replicated the same behaviour - works on the driver, fails on the executors. So guessing you meant something more fine-grained when you say "manage the Kerberos login in your code that runs in executors"? If there's any example code you can point to I'd be grateful (couldn't find the Kafka example). On Tue, Oct 16, 2018 at 3:32 AM Marcelo Vanzin <van...@cloudera.com> wrote: > Spark only does Kerberos authentication on the driver. For executors it > currently only supports Hadoop's delegation tokens for Kerberos. > > To use something that does not support delegation tokens you have to > manually manage the Kerberos login in your code that runs in executors, > which might be tricky. It means distributing the keytab yourself (not with > Spark's --keytab argument) and calling into the UserGroupInformation API > directly. > > I don't have any examples of that, though, maybe someone does. (We have a > similar example for Kafka on our blog somewhere, but not sure how far that > will get you with MS SQL.) > > > On Mon, Oct 15, 2018 at 12:04 AM Foster Langbein < > foster.langb...@riskfrontiers.com> wrote: > >> Has anyone gotten spark to write to SQL server using Kerberos >> authentication with Microsoft's JDBC driver? I'm having limited success, >> though in theory it should work. >> >> I'm using a YARN-mode 4-node Spark 2.3.0 cluster and trying to write a >> simple table to SQL Server 2016. I can get it to work if I use SQL server >> credentials, however this is not an option in my application. I need to >> use windows authentication - so-called integratedSecurity - and in >> particular I want to use a keytab file. >> >> The solution half works - the spark driver creates a table on SQL server >> - so I'm pretty confident the Kerberos implementation/credentials etc are >> setup correctly and valid. However the executors then fail to write any >> data to the table with an exception: >> "java.security.PrivilegedActionException: >> GSSException: No valid credentials provided (Mechanism level: Failed to >> find any Kerberos tgt)" >> >> After much tracing/debugging it seems executors are behaving differently >> to the spark driver and ignoring the specification to use the credentials >> supplied in the keytab and instead trying to use the default spark cluster >> user. I simply haven't been able to force them to use what's in the keytab >> after trying many. many variations. >> >> Very grateful if anyone has any help/suggestions/ideas on how to get this >> to work. >> >> >> -- >> >> >> >> *Dr Foster Langbein* | Chief Technology Officer | Risk Frontiers >> >> Level 2, 100 Christie St, St Leonards, NSW, 2065 >> >> >> Telephone: +61 2 8459 9777 >> >> Email: foster.langb...@riskfrontiers.com | Website: www.riskfrontiers.com >> >> >> >> >> *Risk Modelling | Risk Management | Resilience | Disaster Management >> | Social Research Australia | New Zealand | Asia Pacific* >> >> >> > > > -- > Marcelo >