fix: support nullable map key schemas and reject null key data - #176

Purpose

No Linked issue.

This PR restores nullable MAP-key schema compatibility that was removed by alibaba/paimon-cpp#226.

The changes:

• Accept nullable MAP key declarations when parsing existing Paimon schemas.
• Normalize parsed MAP key fields to non-nullable Arrow fields because Apache Arrow does not support nullable MAP keys.
• Remove the general schema validation requiring MAP keys to be explicitly declared as NOT NULL.
• Continue rejecting nullable MAP keys for the shared-shredding storage layout.
• Validate Arrow arrays at the ORC and Parquet format-reader boundaries so data containing an actual null MAP key returns Status::Invalid instead of triggering an Arrow abort.
• Restore nullable MAP key declarations in existing test schemas.

Tests

Added or updated coverage for:

• Parsing a nullable MAP key declaration and normalizing it for Arrow.
• Creating table schemas containing nullable MAP key declarations.
• Rejecting nullable MAP keys when shared-shredding is enabled.
• Reading the nullable-MAP-key ORC fixture with read.batch.size=1:
  • The first two rows are read and verified successfully.
  • The third row returns Map array keys array should have no nulls.

API and Format

Schema compatibility is relaxed so existing schemas may declare nullable MAP keys. An error is returned only when the physical data actually contains a null MAP key. Shared-shredding continues to require non-nullable MAP keys.

Documentation

Updated the data type and schema documentation to describe nullable MAP key compatibility and read-time validation behavior.

Generative AI tooling

Generated-by: OpenAI Codex (GPT-5)