Regression in `30.0.0` write_parquet ignores all options

[sergiimk]

Describe the bug

After upgrading to DataFusion 30.0.0 I noticed that my carefully-crafted options for DataFrame::write_parquet are now being discarded

I see there was some major refactoring, but is this regression intentional? I don't see it called out in breaking changes of the release 29/30.

We use options to force dictionary and delta encoding on some columns. How can this be achieved now?

To Reproduce

No response

Expected behavior

• No regression in functionality would be preferred
• Release notes should call this out and suggest a workaround
• API probably should panic with unimplemented!() when options are specified instead of ignoring silently

Additional context

No response

[devinjdangelo]

Yes, this was an intentional regression. I agree raising an error on passing configuration settings would have been a good idea for this release. I am actively working on adding support for configuration back in (bulk of the work to reenable is in #7390).

The old write implementation is still in DataFusion 30.0.0 just not via the DataFrame methods directly. You can call the old write_parquet code via SessionContext::write_parquet https://github.com/apache/arrow-datafusion/blob/e0961d55479aab1c4f92eca817fcce4ec25d7c3e/datafusion/core/src/execution/context.rs#L1325-L1333

Apologies for the inconvenience in the current release and not calling out this particular breaking change more clearly.

[sergiimk]

@devinjdangelo thanks for explanation and your hard work! Will stick with older API for now.

I will close this issue and start watching #7390

[sergiimk]

Note: Switching to SessionContext::write_parquet didn't work right away, resulting in IoError(Os { code: 2, kind: NotFound, message: "No such file or directory" }).

The source of the error was this call to .canonicalize() that seems to expect that destination path exists. I worked around this by creating an empty directory before saving to Parquet.