[CALCITE-4263] Clarify RelNode#estimateRowCount behavior with regards to nulls

• Type: Sub-task
• Status: Open
• Priority: Blocker
• Resolution: Unresolved
• Affects Version/s: 1.25.0
• Component/s: core
• Assignee: Unassigned
• Reporter: Vladimir Sitnikov
• Created: 18/Sep/20 07:27
• Updated: 28/Feb/24 06:03

Description

RelNode#estimateRowCount implementation must return non-nullable value, and it often uses metadata which might return null

For instance,

class Correlate {
  @Override public double estimateRowCount(RelMetadataQuery mq) {
    double leftRowCount = mq.getRowCount(left); // <-- NPE here
    switch (joinType) {
    case SEMI:
    case ANTI:
      return leftRowCount;
    default:
      return leftRowCount * mq.getRowCount(right); // <-- NPE here
    }
  }

There should be documentation on the desired behavior in case metadata returns "unknown", and the code needs adjustment as lot of the implementation perform blind unboxing.

Issue Links

• relates to: CALCITE-4215 Ensure org.apache.calcite.schema.Statistic uses null vs emptyList appropriately