[VL] bug of parquet read

[FelixYBW]

Describe the bug
We saw this bug from another customer again. let's track it here.

To Reproduce
only appear in customer env.

[PHILO-HE]

This issue can be reproduced by reading a parquet file whose string column has long min/max string value. A reproducible extreme case is the column has a string containing 1000 repeated 'z' characters as the max value.

Velox community has opened a similar issue: facebookincubator/velox#3882
And there is a PR already in the community: facebookincubator/velox#4108
Our test shows the PR works for fixing this issue.

[FelixYBW]

This issue can be reproduced by reading a parquet file whose string column has large min/max string value. A reproducible extreme case is the column has a string containing 1000 repeated 'z' characters as the max value.

Velox community has opened a similar issue: facebookincubator/velox#3882 And there is a PR already in the community: facebookincubator/velox#4108 Our test shows the PR works for fixing this issue.

Do we have a PR in Gluten?

[PHILO-HE]

This issue can be reproduced by reading a parquet file whose string column has large min/max string value. A reproducible extreme case is the column has a string containing 1000 repeated 'z' characters as the max value.
Velox community has opened a similar issue: facebookincubator/velox#3882 And there is a PR already in the community: facebookincubator/velox#4108 Our test shows the PR works for fixing this issue.

Do we have a PR in Gluten?

The fix is only required on velox code side. Even though the velox community patch works for solving this issue, it is still under development to avoid causing any other possible issues I guess. We will create a PR on our side once community PR is in good shape.

[FelixYBW]

@PHILO-HE can we have a temp PR to pick up Velox's PR, so customer can try?

[PHILO-HE]

@PHILO-HE can we have a temp PR to pick up Velox's PR, so customer can try?

Yes, I just filed a PR to port upstream patches: oap-project/velox#145. On my side, a simple test shows the PR can work well.