Iceberg table writer writes space to _x20 in column names in the Parquet file

[yingsu00]

Using the IcebergQueryRunner, create table and insert would change the space in the column name to "_x20". For example, "two words" would be written as "two_x20words" in the Parquet file. This doesn't happen in Hive tables, where "two words" would just be written as "two words"

Your Environment

N/A

Expected Behavior

Iceberg table writer write space as is. This is in alignment with Hive.

Current Behavior

Iceberg table writer replaces space " " to "_x20". This is not desired.

Possible Solution

Steps to Reproduce

Start IcebergQueryRunner (Java). In presto cli, run the following

create table space2 ("two words" int) with (format='parquet');
insert into space2 values (1), (2), (3);

Then put a break point at com/facebook/presto/hive/parquet/ParquetPageSourceFactory.java line 207
MessageType fileSchema = fileMetaData.getSchema();

Repeat the same steps using HiveQueryRunner, and you will see the name is different.

Screenshots (if appropriate)

Context

[karteekmurthys]

The Iceberg module sanitzes the column names if it has special characters. Check here:
org/apache/iceberg/parquet/TypeToMessageType.java

    for (NestedField field : schema.columns()) {
      builder.addField(field(field));
    }

    return builder.named(AvroSchemaUtil.makeCompatibleName(name));
  }

Further up the stack it reaches here in org/apache/iceberg/avro/AvroSchemaUtil.java:

private static String sanitize(char character) {
    if (Character.isDigit(character)) {
      return "_" + character;
    }
    return "_x" + Integer.toHexString(character).toUpperCase();
  }

It is coming from dependent java libraries.

[hantangwangd]

Seems Iceberg will sanitize all the characters that aren't allowed in Avro field names, including space " ". Referring to apache/iceberg#216 (comment)

[yingsu00]

So this is expected behavior. Closing for now. We will support reading such columns through name-id mapping in Velox. See facebookincubator/velox#10085