[IMPALA-10393] Iceberg field id-based column resolution fails in ASAN builds

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: Impala 4.0.0
• Labels: broken-build impala-iceberg
• Epic Color: ghx-label-4
• Assignee: Zoltán Borók-Nagy
• Reporter: Zoltán Borók-Nagy
• Created: 14/Dec/20 14:19
• Updated: 17/Dec/20 12:00
• Resolved: 17/Dec/20 12:00

Description

For MAP types field id resolution indexes the top-level columns via the current 'table_idx - 1' at:

https://github.com/apache/impala/blob/4c0bdbada0bc0eeb0435e1ea647573566f0cddbd/be/src/exec/parquet/parquet-metadata-utils.cc#L769-L771

In this case table_idx is either SchemaPathConstants::MAP_KEY or SchemaPathConstants::MAP_VALUE which are 0 and 1 respectively. Hence 'table_idx - 1' can be -1 which is not a valid index for a vector, hence we get an ASAN error. Even if the 'table_idx - 1' is zero we get a wrong field id.

Note that at this point in the schema resolution we have successfully found a MAP type with a matching field id, therefore it is safe to resolve the child via the value of 'table_idx' (which is the position of the child, MAP_KEY or MAP_VALUE).