[CI][C++][Parquet] Failure on `parquet-reader-test` with simdjson 5.x

[tadeja]

Describe the bug, including details regarding any error messages, version, and platform.

Failing jobs: AMD64 Windows MinGW MINGW64 C++, AMD64 Windows MinGW CLANG64 C++, AMD64 Conda C++ AVX2 and
C++ Extra / AMD64 Ubuntu Meson

With simdjson 5.x failing on parquet-reader-test

[  FAILED  ] 4 tests, listed below:
[  FAILED  ] TestJSONWithLocalFile.JSONOutputWithStatistics
[  FAILED  ] TestJSONWithLocalFile.JSONOutput
[  FAILED  ] TestJSONWithLocalFile.JSONOutputFLBA
[  FAILED  ] TestJSONWithLocalFile.JSONOutputSortColumns
...
The following tests FAILED:
	 92 - parquet-reader-test (Failed)                      parquet-tests unittest

First MinGW job failure on main: https://github.com/apache/arrow/actions/runs/36565311423/job/109395674762#step:13:1885

Because of MSYS2 simdjson update to 5.0.1
mingw-w64-x86_64-simdjson-5.0.1-1-any downloading...
( -- ARROW_SIMDJSON_BUILD_VERSION: v4.6.4 later in the log is probably misleading )

Last successful MinGW job log on main still had
mingw-w64-x86_64-simdjson-4.6.11-1-any downloading...

Component(s)

Continuous Integration, C++

[aryan9948]

Hi @tadeja, I'd like to investigate this issue. I can look into the MinGW parquet-reader-test failures and whether the MSYS2 simdjson 5.0.1 update is causing the regression. Would you be okay with assigning this issue to me?

[kou]

You can just do it and share what you found here. We can collaborate with it. See also the findings in #51640.

[tadeja]

PR #50877 introduced simdjson::fractured_json_string and testing exact outputs in reader_test.cc

Now simdjson 5.0 changes FracturedJson (simdjson/simdjson#2870, simdjson/simdjson#2878 and simdjson/simdjson#2879 change whitespace, line breaks, key order in tables),
which already causes parquet-reader-test failures for MinGW jobs and soon other jobs (probably ARM64 macOS 14 C++ and AMD64 Conda C++ AVX2 when Docker layer cache gets invalidated.)

I have a workaround in #51640 to get MinGW two jobs from 5.0.1 to bundled 4.6.4 simdjson and passing CI for the time being.

Likely this needs a more permanent fix follow-up in parquet/reader_test.cc to accommodate for both test outputs, 4.6 and 5.0
(I have a working demo by 🤖 on my fork tadeja#2)

Fyi, @kou @Reranko05

[kou]

How about testing TestJSONWithLocalFile.JSONOutput* only with simdjson 5 and updating the expected JSONs only for simdjson 5?

[raulcd]

I think the approach @kou is proposing is reasonable. Either only test and validate the output (as I understand is slightly different in format) on simdjson 5 or update the check so it matches the expected output based on the simdjson version.
@tadeja I am marking this as 26.0.0 even though it does not feel like a blocker is good to have a green CI on the release.

[tadeja]

@kou, @raulcd so what would you say about both in #51661 - new expected output for >= 5 (else existing 4.x test output)

[tadeja]

Two more jobs fail on main:
C++ / AMD64 Conda C++ AVX2
C++ Extra / AMD64 Ubuntu Meson