Hi @rpaidar Most, of the code that does this can be found in the "tablet" module. Iterators for all rowsets are assembled in Tablet::CaptureConsistentIterators() (https://github.com/cloudera/kudu/blob/master/src/kudu/tablet/tablet.cc#L1572) Then, for diskrowsets, the base data iterator is wrapped in a delta iterator. This process starts at DiskRowSet::NewRowIterator() (https://github.com/cloudera/kudu/blob/master/src/kudu/tablet/diskrowset.cc#L592) To get more insight on how the base is materialized and deltas are applied you can follow the different iterators that are created in DiskRowSet::NewRowIterator(). HTH David
... View more
That suggests that there is still at least one machine running impala that can't communicate with a kudu server. Are you sure you've fixed the problem for all machines? If there is at least one machine that still has the problem, that might be the reason why you're still seeing the error. A small table that has less partitions and/or smaller data that only goes into a few of them only writes to a subset of all kudu tservers, while a big table with more partitions and/or data that goes into all of them might have to write to all of them. If there is a single machine/resolution that is not "fixed" as per Adar suggestion and not used in the small table/data case might have to be used in the big table/data case.
... View more