Add metadata field describing the application that wrote the file

[asfimport]

Although the Parquet library should behave the same regardless of what application uses it, occasionally serious interoperability bugs are introduced in specific applications. For example, data written by a specific application may be unnecessarily adjusted or the calculated statistics may be invalid (both actual problems).

Unfortunately, currently it is not possible to recognize Parquet files affected by application problems because the metadata does not contain any information about the application using the Parquet library. (The name and version number of the Parquet library is recorded, but that only has limited use, because apart from Impala, the most widespread Parquet writers all use the same Java library.)

To allow creating workarounds for future known issues, we should introduce new metadata fields that applications can populate. The simplest approach is to have one field for the application name and another for its version number. A more sophisticated approach suggested by @julienledem could also reference a list of earlier issues that are known to be fixed in the application that wrote the Parquet file.

Reporter: Zoltan Ivanfi / @zivanfi

Related issues:

• Add tags to "created by" metadata in the file footer (duplicates)

_{Note: This issue was originally created as PARQUET-899. Please see the migration documentation for further details.}

[asfimport]

Nándor Kollár / @nandorKollar:
@zivanfi isn't org.apache.parquet.hadoop.metadata.FileMetaData#createdBy used for this purpose? Also, this class has a Map<String, String> keyValueMetaData variable too, applications can store arbitrary metadata there (like version number).

[asfimport]

Zoltan Ivanfi / @zivanfi:
The createdBy metadata stores the name and version number of the Parquet library, not the higher-level component that consumes the Parquet API. The freeform key-value metadata can be used for storing any information, but it won't be standard. I think we should add a new, dedicated field (or rather two separate ones for identifying the writer and its version, respectively) to the specification in parquet-format and the language bindings as well. By the latter I mean API functions that higher-level components can call to set this information. Another alternative would be to add this new parameter to some already existing function that components have to call when they write Parquet files. In this case we can even deprecate the older versions that do not accept this information.

[asfimport]

Zoltan Ivanfi / @zivanfi:
Quoting from the commit for PARQUET-352:

WriteSupport now has a getName getter method that is added to the footer
if it returns a non-null string as writer.model.name. This is intended
to help identify files written by object models incorrectly.

So writer.model.name is already there for this purpose, albeit undocumented.