I reported IMPALA-4725 last week but it seems like it has not been triaged yet. I wanted to bring some more attention to this issue (and possible suggestions for workarounds) since it has a heavy impact on us.
To summarize it seems like Impala mixes-up values in arrays of structs which to me seems like a fundamental problem in the parquet reader. Alternatively the values gets mixed-up when presented as a result.
Either way, I would very much appreciated an initiated persons view on this issue.
We are running Impala that is bundled with CDH 5.8.3