ARROW-6187: [C++] Fallback to storage type when writing ExtensionType to Parquet

[jorisvandenbossche]

https://issues.apache.org/jira/browse/ARROW-6187

[pitrou]

It would be better to store the extension name and serialization in the Parquet metadata, no? Then you can recreate a true extension array when reading?

[jorisvandenbossche]

Yes, that was the idea to look into next. But I can probably directly try that in this PR.

[pitrou]

I think that would be nice, since that would validate the whole approach. The metadata keys are currently private in arrow/ipc/metadata_internal.cc (see kExtensionTypeKeyName and kExtensionMetadataKeyName), you can probably make them extern in a .h somewhere.

Perhaps you could instead use Parquet-specific metadata keys if Parquet has a well-known convention for them, though.

[jorisvandenbossche]

To a potential reviewer already looking at it, this is currently a very ugly implementation(just wanted to have something working now to pass the tests), will try to clean-up tomorrow.

[pitrou]

The implementation looks good on the principle, feel free to ping when this is ready for review.

[jorisvandenbossche]

This should be ready for review now.

[pitrou]

LGTM, thanks!

[pitrou]

AppVeyor build: https://ci.appveyor.com/project/jorisvandenbossche/arrow/builds/27629229