The " Connection reset by peer" message should have just been printed by the client (e.g. Hue, impala-shell, etc.). If the impalad itself crashed it should have printed some exit message or other logs that indicated why the impalad crashed (logs should be in impalad.INFO). If the impalad logs don't have anything useful, then it should have written either a minidump or core dump which can be parsed to see what caused it to crash. How big are your input files? Is it possible that with only one input file, only a single scan range was used to read the table, and that with more input files, multiple scan ranges were spawned? If that is the case, it sounds like there might be a race condition somewhere .
... View more
