mirror of
https://github.com/apache/doris.git
synced 2026-11-04 16:03:23 +00:00
Jerry Hu
699680e6f8
### What problem does this PR solve?
Issue Number: None
Problem Summary:
`JSON_SEARCH` builds the returned JSON path by concatenating each object
key as a bare member leg (`.key`), only double-quoting the leg when the
key contains whitespace. When the key itself contains a character that
`JsonbPath`'s parser treats as structural when unquoted — `.`, `[`, a
leading `*`, or an empty key — the emitted path does not parse back into
the same single leg, so feeding the result into
`JSON_EXTRACT`/`JSON_SEARCH` again resolves to a different value (or
fails to parse).
For example:
```sql
JSON_SEARCH('{"a.b":"hit","a":{"b":"miss"}}', 'one', 'hit')
-- returns $.a.b, which JSON_EXTRACT parses as a -> b and resolves to "miss" instead of "hit"
```
The fix makes `leg_info::to_string` double-quote a member leg whenever
the unquoted form would not read back as the same leg (mirroring
`JsonbPath`'s own unquoted-leg grammar: empty key, leading `*`, or
containing whitespace/`.`/`[`), and adds a few supporting changes so the
rest of the path code treats a quoted literal key consistently with a
parsed wildcard:
- A new `leg_info::is_wildcard` flag, set only by the parser for an
unquoted `*` / `[*]` leg, distinguishes a real wildcard from the literal
key `"*"`. `to_string` and `findValue` both switch to this flag instead
of comparing leg bytes, since a quoted `"*"` leg has the same bytes as a
wildcard leg.
- `$.""` (an explicit empty key) is accepted by the parser and can be
looked up; `ObjectVal::search` no longer treats an empty key length as
"not found", and the internal lookup distinguishes an explicit empty key
from an object entry that is stored by numeric key id (which also has
key length 0).
- `JsonbValue::contains` looks up a key-id-encoded right-hand key by id
instead of by name, for the same reason.
- `JSON_SET`/`JSON_INSERT` no longer write a member-style leg into an
array (matching MySQL); previously the leg's default array index (0)
would overwrite or append to the array.
- A pre-existing formatting issue where a wildcard array leg (`[*]`) was
printed back as `[0]` is fixed alongside, since it shares the same root
cause (no explicit wildcard marker).
### Release note
`JSON_SEARCH` now quotes a returned path segment whenever the key
contains `.`, `[`, starts with `*`, or is empty, so the result can be
parsed again by `JSON_EXTRACT`/`JSON_SEARCH`/`JSON_SET`/etc. As a side
effect, `$.""` is now a valid path for an explicit empty key, a quoted
`"*"` key is treated as a literal key instead of a wildcard, and
`JSON_SET`/`JSON_INSERT` no longer modify an array when the path's last
segment is a member (non-array-index) leg.
### Check List (For Author)
- Test
- [x] Unit Test
- [x] Regression test
- Behavior changed:
- [x] Yes. See Release note above.
- Does this need documentation?
- [ ] No.