Thomas, you have a legitimate request and concern.   First, there is no perfectly fool-proof solution because the resource consumption is somewhat dependent on what happens at runtime, and not all memory consumption is tracked by Impala (but must is). We are constantly making improvements in this area though.   1. I'd recommend fixing the num_scanner_threads for your queries. A different number of scanner threads can result in different memory consumption from run to run (and dependent on what else is going on in the system at the time). 2. The operators of a query do not run one-by-one. Some of them run concurrently (e.g. join builds may execute concurrently). So just looking at the highest peak in the exec summary is not enough. Taking the sum of the peaks over all operators is a safer bet, but tends to overestimate the actual consumption.   Hope this helps! ... View more

I just saw this thread after commenting on the Jira. Would "conv()" be a suitable workaround here?   select conv('100010', 2, 10); +-----------------------+ | conv('100010', 2, 10) | +-----------------------+ | 34 | +-----------------------+ Fetched 1 row(s) in 0.24s   More information on conv() can be found in the Impala documentation.    Edit: To make things complete, the Jira is IMPALA-4968. ... View more