Support columnar storage layout for composite types (tuple/struct)

[zhanglei1949]

Background

Currently, composite types such as tuples or structs are stored as a single value/blob. This makes it inefficient to access individual fields, and prevents us from leveraging columnar storage benefits (e.g., better compression, vectorized scans, projection pushdown).

Proposal

Adapt the storage layer to support a columnar layout for composite types. The attributes of a tuple/struct should be physically stored as multiple sub-columns, one per field, instead of a single packed record.

Example

For a struct Person { name: string, age: int, address: Address { city: string, zip: int } }, the storage would expand into sub-columns:

• Person.name
• Person.age
• Person.address.city
• Person.address.zip

Requirements

• Support flat tuple/struct types — each field becomes a sub-column
• Support nested composite types — sub-columns can themselves be decomposed recursively
• Field-level access should read only the relevant sub-column(s) (projection pushdown)
• Schema evolution considerations (adding/removing fields)

Benefits

• Better compression ratio (homogeneous data per column)
• Faster scans when only a subset of fields is needed
• Natural fit for analytical / vectorized execution

Open Questions

• How to encode nested paths in the sub-column naming scheme?
• How to handle null/optional fields at each nesting level?
• Interaction with existing index structures?