supabase / etl / 30000487710

74%

Build Type

push

github

Committed by web-flow

Commit Message

Fix array dimension prefix parsing and add codec fuzz targets (#929)

* fix(etl): Handle array dimension prefixes and reject multidim arrays

The text array parser treated unquoted braces in the array body as
element content and required the literal to start with '{'. Both break
on legal Postgres output:

- Arrays with a lower bound other than 1 render with an explicit
  dimensions prefix, e.g. '[0:1]={7,8}', and were spuriously rejected
  with 'Array input missing braces', erroring the table.
- Multidimensional values for string-typed arrays were silently
  corrupted: '{{NULL},{NULL}}' parsed as two non-null '{NULL}' strings.
  Integer-typed arrays failed loudly on the same input by accident of
  element parsing.

The parser now validates and skips a one-group dimensions prefix
(bounds carry no value information for a one-dimensional array) and
rejects nested unquoted braces and multi-group prefixes with a clear
error, since ArrayCell cannot represent multiple dimensions.

Found by two new value roundtrip properties: shifted-lower-bound int8[]
literals must roundtrip, and rendered 2-D text[] values must be
rejected loudly (pinning reject-not-corrupt until real support exists).
Minimal counterexamples were '[0:0]={NULL}' and '{{NULL},{NULL}}'.

* test(etl): Add fuzz targets for the text codec, COPY rows, and numeric

Extends the fuzz crate from one target (bytea hex) to four:

- parse_text_cell: drives parse_cell_from_postgres_text across every
  dedicated codec branch (scalar and array) plus the generic fallbacks,
  with the type picked by a selector byte.
- parse_copy_row: drives parse_table_row_from_postgres_copy_bytes with
  a schema derived from leading input bytes, covering COPY escape
  handling and row termination.
- numeric_text_roundtrip: checks parse/print/parse stability of
  PgNumeric text values, so the target catches logic bugs (unstable
  renderings) as well as panics.

The fuzzing facade in etl::fuzzing owns type-table selection and t... (continued)

Coverage Stats

82 of 116 new or added lines in 2 files covered. (70.69%)

26 existing lines in 2 files now uncovered.

43970 of 59412 relevant lines covered (74.01%)

8137.57 hits per line