It turned out that Impala hashes STRINGs differently than Hive.
Impala's implementation simply hashes the original byte array (e.g. a UTF-8 encoded string), while Hive hashes the UTF-16 encoded char array behind java strings. If the STRING is cast to BINARY in Hive (e.g. ds_hll_sketch(cast(s as binary)) ), then it is interoperable with Impala's current implementation.
I am not sure how to proceed - we could UTF-16 encode the strings in Impala before hashing, but this would be pretty slow, and I think that Hive actually could be also faster if it would hash UTF-8 arrays - as STRINGs are stored as org.apache.hadoop.io.Text, they are currently UTF-8 decoded to java string first and could be hashed directly without any conversion. This would break compatibility with existing Hive produced sketches though.