We should add tests for most or all of the subtasks of IMPALA-4834 that exercise the expected OOM code path. I'll add some bullet points here to track what the coverage is and what's missing.
Aggregations with large var-len string expressions Top-N with large window Kudu scans: covered by kudu-scan-mem-usage.test Nested loop join: covered by single-node-nlj-exhaustive.test Many duplicate keys on build side of hash join Large number of NULLS on build side of NAAJ: covered by spilling-naaj-no-deny-reservation.test HDFS table partitioned insert: covered insert-mem-limit.test Large analytic window can't be spilled
- Queries processing large strings (may need multiple tests to cover different places). large_strings.test has some coverage
- Parquet files with large pages
Exchange uses a lot of memory