Datanodes are reading all the time without corresponding HDFS I/O activity after 5.4.5 upgrade

[#1]

Earlier this week I upgraded to CDH 5.4.5 from 5.4.4. Since then I can see that the datanode process is constantly reading 2M/s per drive in each host (with a number of writes maybe one order of magnitude smaller), but there is no corresponding HDFS I/O activity (just the usual activity, ~40k/s) 

It seems as if nobody is actually reading anything, but the HDFS process is doing something on its own.

Any ideas about what could be causing this? How could I find out / diagnose what's happening?

Thanks!

[#2]

Nevermind, it was a zombie spark job.

[#3]

I am happy to see you killed off the zombies. 🙂

[#4 · Accepted solution]

After some more research we found that it wasn't a zombie spark job that was causing the resource usage. It was the HDFS blockscanner. Apparently the default configuration changed with the upgrade and it started running right after we restarted upon upgrading. We had never seen it running before and hence the mistery.

[#5]

That makes sense but totally invalidates my Zombie sign. 🙂

Feel free to mark your last comment as the solution in case it can help others in the future.