advitrocks9 opened a new pull request, #51108:
URL: https://github.com/apache/arrow/pull/51108
### Rationale for this change
`{"a": [null, 1]}` parses to `[[1]]`. Nothing raises, and
`validate(full=True)` is the only thing
that notices:
```
Offset invariant failure: offset for slot 1 out of bounds: 2 > 1
```
Debug builds abort on a DCHECK instead, at `parser.cc:540` if the second
element is null and
`parser.cc:1151` if it is typed.
### What changes are included in this PR?
`RawBuilderSet::AppendNull` takes the child builder by value. Builders of
`Kind::kNull` have no
arena, their null count is packed into `BuilderPtr::index`, so the increment
updates the parent
but is lost to the caller. The next null then re-increments the same stale
value, which leaves an
all-null list as `NullArray(1)` and a promoted list with its leading nulls
missing.
It now takes a `BuilderPtr*`, matching `MakeBuilder` in the same class. The
regression came in
with `b7054c21aab` and has shipped since 0.14.0.
### Are these changes tested?
`BlockParser.NullsInList` covers both arms. `arrow-json-test` passes 261/261
in Debug and
Release. Reverting `parser.cc` with the test in place aborts in Debug and
fails both columns in
Release.
### Are there any user-facing changes?
Those inputs parse correctly now. This is not inference-only: an explicit
`list(null())` schema
reaches the same arm under both `Error` and `Ignore`.
**This PR contains a "Critical Fix".** Well-formed JSON silently loses list
elements in release
builds and aborts on a DCHECK in debug builds.
* GitHub Issue: #47417
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
