Skip to content

Table version clause (AT(...), FOR SYSTEM_TIME AS OF, VERSION AS OF) is displayed after the alias and doesn't reparse #2804

Open

Description

@alpaylan

The parser reads the version clause before the alias and table hints, but Display writes it after them, so the output doesn't parse:

Snowflake:  SELECT * FROM my_table AT(OFFSET => -60*5) AS T WHERE T.flag = 'valid'
  -> displays as: SELECT * FROM my_table AS T AT(OFFSET => -60 * 5) WHERE T.flag = 'valid'
  -> reparse: Expected: identifier, found: ( at Line: 1, Column: 31

MsSql:      SELECT * FROM t FOR SYSTEM_TIME AS OF '2014-01-01' AS x WITH (NOLOCK)
  -> displays as: SELECT * FROM t AS x WITH (NOLOCK) FOR SYSTEM_TIME AS OF '2014-01-01'
  -> reparse: Expected: one of UPDATE or SHARE, found: SYSTEM_TIME at Line: 1, Column: 40

Databricks: SELECT * FROM events VERSION AS OF 123 AS e
  -> displays as: SELECT * FROM events AS e VERSION AS OF 123
  -> reparse: Expected: end of statement, found: VERSION at Line: 1, Column: 27

The Snowflake input is the example from https://docs.snowflake.com/en/sql-reference/constructs/at-before. SQL Server documents table_or_view_name [ FOR SYSTEM_TIME <system_time> ] [ [ AS ] table_alias ] ... [ WITH ( <table_hint> ... ) ] (https://learn.microsoft.com/en-us/sql/t-sql/queries/from-transact-sql), and Databricks puts the temporal spec in table_name, before table_alias (https://docs.databricks.com/aws/en/sql/language-manual/sql-ref-syntax-qry-select-table-reference).

BigQuery documents the opposite order, table_name [ as_alias ] [ FOR SYSTEM_TIME AS OF ... ] (https://cloud.google.com/bigquery/docs/reference/standard-sql/query-syntax#from_clause), and with BigQueryDialect SELECT * FROM t AS x FOR SYSTEM_TIME AS OF '2020-01-01' is rejected (Expected: one of UPDATE or SHARE, found: SYSTEM_TIME), while the other order parses and displays as the rejected form. So the right position seems to be dialect dependent.

Parser: maybe_parse_table_version runs before maybe_parse_table_alias in parse_table_factor. Display: the TableFactor::Table arm writes alias, index hints, WITH (...), then the version. The open PR #2486 mentions this in passing and leaves it for a separate change.

Tested on main ce12f1f (also 0.63.0; on 0.58.0 the Databricks input is a parse error).

Found with a property-based fuzzer (round trip check); I used AI assistance for triage and verified every case above.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions