We have a lot of direct inserts in our database. So, we have a lot of mergeout operations.
But sometimes we observe strange behaviour of that operations.
Some mergeout operations consume CPU on one node for couple of hours. LA of that node increase and sometimes this node have a huge delay in Global Catalog operations with this node (transactions lock GC at this node with ~20 second delay). So, sometimes we have GCL timeout problem. This mergeout sessions can be seen in system_sessions table, but doesn't appear in tuple_mover_operations table. We can grep it in vertica.log
This looks like this:
Average ROS containers count for our projections is about 300-400.
Recently, we had similar problem, but it was solved by proper TM configuration: https://forum.vertica.com/discussion/238751/high-cpu-usage
So, is this normal behaviour? Has somebody such problems in mergeouts?
