Part of #519 (Thrift vs SEA metadata parity).
Problem
When an empty-string identifier argument (catalog="", schema="", table="", foreign_schema="", …) is passed to a metadata op, the protocols diverge in outcome — the single biggest non-value bucket, 1,763 outcome diffs:
outcome: thrift="exception: HiveServer2Exception" → sea="Apache.Arrow.Adbc.QueryResult"
Trigger breakdown (which empty arg caused the Thrift-only throw):
catalog="" → 464schema_pattern="" → 224schema="" → 48foreign_schema="" → 36table="" / foreign_table="" → 20 each- plus combinations
Affected ops: list_columns (1,020), list_tables (336), get_primary_keys (213), get_cross_reference (194 — every cross-reference diff is this).
Expected
Empty-string vs null arguments must be handled consistently across protocols — either both throw or both return. They should not differ in outcome.
How JDBC does it
The JDBC SEA path explicitly treats an empty-string catalog as a sentinel and returns an empty result set (DatabricksMetadataSdkClient.listSchemas → returns 0 rows when catalog.isEmpty()); the Thrift path passes it through. ADBC should define one behavior (recommend: empty-string → empty result set, never an exception) and apply it on both paths before the RPC/SQL is issued.
Suggested fix
Normalize empty-string identifier args up front (in the shared layer, #519): map to the chosen defined behavior so the Thrift path stops raising HiveServer2Exception.
Evidence
diff-report-20260615-064733Z.json, where = step[0] / outcome.
Part of #519 (Thrift vs SEA metadata parity).
Problem
When an empty-string identifier argument (
catalog="",schema="",table="",foreign_schema="", …) is passed to a metadata op, the protocols diverge in outcome — the single biggest non-value bucket, 1,763 outcome diffs:Trigger breakdown (which empty arg caused the Thrift-only throw):
catalog=""→ 464schema_pattern=""→ 224schema=""→ 48foreign_schema=""→ 36table=""/foreign_table=""→ 20 eachAffected ops:
list_columns(1,020),list_tables(336),get_primary_keys(213),get_cross_reference(194 — every cross-reference diff is this).Expected
Empty-string vs null arguments must be handled consistently across protocols — either both throw or both return. They should not differ in outcome.
How JDBC does it
The JDBC SEA path explicitly treats an empty-string
catalogas a sentinel and returns an empty result set (DatabricksMetadataSdkClient.listSchemas→ returns 0 rows whencatalog.isEmpty()); the Thrift path passes it through. ADBC should define one behavior (recommend: empty-string → empty result set, never an exception) and apply it on both paths before the RPC/SQL is issued.Suggested fix
Normalize empty-string identifier args up front (in the shared layer, #519): map to the chosen defined behavior so the Thrift path stops raising
HiveServer2Exception.Evidence
diff-report-20260615-064733Z.json,where = step[0] / outcome.