LIKE predicate pushdown feature

Impala pushes eligible SQL LIKE predicates on Iceberg tables to Iceberg expressions for file and partition pruning.

The Iceberg specification supports expression-based filtering through manifest statistics. Impala translates eligible LIKE predicates to Iceberg startsWith() and equal() expressions during query planning. This enables Iceberg to skip files and partitions that cannot contain matching rows.

You need not run this optimization with new SQL syntax or table properties. Impala applies it automatically when the LIKE pattern is eligible for pushdown.

Direct predicate pushdown

For patterns where wildcard behavior is straightforward, Impala translates the LIKE predicate directly to an Iceberg expression:


SQL `LIKE` pattern	Iceberg expression
`LIKE 'abc%'`	`startsWith('abc')`
`LIKE 'pre_fix%'`	`startsWith('pre')`
`LIKE 'exact'`	`equal('exact')`
`LIKE 'asd\%'`	`equal('asd%')`

Relaxed predicate pushdown

For complex patterns that contain literal content after a wildcard (for example, LIKE 'prefix%suffix' or LIKE 'd%d'), Impala uses a two-phase approach:

Impala pushes a relaxed prefix predicate to Iceberg (for example, startsWith('prefix')) for coarse-grained file and partition pruning.
Impala retains the full LIKE predicate and evaluates it on the surviving rows to return correct results.

Limitations

Impala cannot push down LIKE patterns that start with a wildcard, such as LIKE '%suffix' or LIKE '_prefix', because there is no leading literal prefix to translate.

For ALTER TABLE … DROP PARTITION and SHOW FILES IN … PARTITION, Impala supports simple prefix LIKE patterns (for example, s LIKE 'd%'). Patterns with literal content after a wildcard are rejected to prevent unintended partition operations. For example, DROP PARTITION (s LIKE 'd%d') returns an error similar to the following:

AnalysisException: Predicate 's LIKE 'd%d'' can only be partially converted to Iceberg expression

Impala example

-- Direct pushdown: prefix pattern on a partitioned column
SELECT count(*) FROM ice_part WHERE action LIKE 'c%';

-- Relaxed pushdown: literal content after wildcard
SELECT s FROM ice_tbl WHERE s LIKE 'test%value';