Arrow and ADBC extensions and remaining limits¶
The scoped extensions below build on the supported Arrow and ADBC contracts. Trackers #300–#306 are closed as completed; their change records document validation and remaining boundaries. Map and Union have explicit stored contracts and partition limits described in the Arrow guide.
| Follow-up capability | Implemented scope and remaining boundary | Tracker |
|---|---|---|
| Incremental execution for more SQL plans | Bounded unordered DISTINCT is incremental with stream_required; general HASH discovery and other blocking operators keep documented fallback/rejection. |
#300 |
| Dictionary-encoded IPC files | Bounded file-wide dictionary construction/remapping is implemented; per-batch dictionaries remain available for IPC streams. | #301 |
| Explicit result layouts in partitioned reads | Non-dictionary layouts replay through versioned live/retained descriptors; dictionary partition output remains rejected. | #302 |
| Column-level statistics | Authoritative bounded NULL counts join row counts; distinct/min/max/width/size remain omitted. | #303 |
| Broader partition SQL coverage | Live byte-length functions expand the deterministic subset; regex SUBSTRING, recursion, generators and lateral/table functions remain excluded. | #304 |
| Nested extensions and additional explicit mappings | UUID/JSON normalization extends through bounded STRUCT nesting; unknown and other nested extension containers reject. | #305 |
| Cancellation during driver-side sorts | Query sorting checks every 256 comparisons and aborts before result publication. | #306 |
Transaction and constraint gaps remain in #113 and #114. No shared source snapshot, remote server-work termination or rollback guarantee is added by the support labels. Experimental GPU/device expansions have their own open trackers in the feature matrix.