PARQUET_ARRAY_RESOLUTION Query Option (Impala 2.9 or higher only)

The PARQUET_ARRAY_RESOLUTION query option controls the behavior of the indexed-based resolution for nested arrays in Parquet.

In Parquet, you can represent an array using a 2-level or 3-level representation. The modern, standard representation is 3-level. The legacy 2-level scheme is supported for compatibility with older Parquet files. However, there is no reliable metadata within Parquet files to indicate which encoding was used. It is even possible to have mixed encodings within the same file if there are multiple arrays. The PARQUET_ARRAY_RESOLUTION option controls the process of resolution that is to match every column/field reference from a query to a column in the Parquet file.

The supported values for the query option are:

• THREE_LEVEL: Assumes arrays are encoded with the 3-level representation, and does not attempt the 2-level resolution.
• TWO_LEVEL: Assumes arrays are encoded with the 2-level representation, and does not attempt the 3-level resolution.
• TWO_LEVEL_THEN_THREE_LEVEL: First tries to resolve assuming a 2-level representation, and if unsuccessful, tries a 3-level representation.

All of the above options resolve arrays encoded with a single level.

A failure to resolve a column/field reference in a query with a given array resolution policy does not necessarily result in a warning or error returned by the query. A mismatch might be treated like a missing column (returns NULL values), and it is not possible to reliably distinguish the 'bad resolution' and 'legitimately missing column' cases.

The name-based policy generally does not have the problem of ambiguous array representations. You specify to use the name-based policy by setting the PARQUET_FALLBACK_SCHEMA_RESOLUTION query option to NAME.

Type: Enum of TWO_LEVEL, TWO_LEVEL_THEN_THREE_LEVEL, and THREE_LEVEL

Default: THREE_LEVEL

Added in: Impala 2.9.0

Examples:

EXAMPLE A: The following Parquet schema of a file can be interpreted as a 2-level or 3-level:

ParquetSchemaExampleA {
  optional group single_element_groups (LIST) {
    repeated group single_element_group {
      required int64 count;
    }
  }
}

The following table schema corresponds to a 2-level interpretation:

CREATE TABLE t (col1 array<struct<f1: bigint>>) STORED AS PARQUET;

Successful query with a 2-level interpretation:

SET PARQUET_ARRAY_RESOLUTION=TWO_LEVEL;
SELECT ITEM.f1 FROM t.col1;

The following table schema corresponds to a 3-level interpretation:

CREATE TABLE t (col1 array<bigint>) STORED AS PARQUET;

Successful query with a 3-level interpretation:

SET PARQUET_ARRAY_RESOLUTION=THREE_LEVEL;
SELECT ITEM FROM t.col1

EXAMPLE B: The following Parquet schema of a file can be only be successfully interpreted as a 2-level:

ParquetSchemaExampleB {
  required group list_of_ints (LIST) {
    repeated int32 list_of_ints_tuple;
  }
}

The following table schema corresponds to a 2-level interpretation:

CREATE TABLE t (col1 array<int>) STORED AS PARQUET;

Successful query with a 2-level interpretation:

SET PARQUET_ARRAY_RESOLUTION=TWO_LEVEL;
SELECT ITEM FROM t.col1

Unsuccessful query with a 3-level interpretation. The query returns NULLs as if the column was missing in the file:

SET PARQUET_ARRAY_RESOLUTION=THREE_LEVEL;
SELECT ITEM FROM t.col1