OutOfMemoryError in job commit / ParquetMetadataConverter

We're trying to write some 14B rows (about 3.6 TB in parquets) to parquet files. When our ETL job finishes, it throws this exception, and the status is "died in job commit".

2015-05-14 09:24:28,158 FATAL [CommitterEvent Processor #4] org.apache.hadoop.yarn.YarnUncaughtExceptionHandler: Thread Thread[CommitterEvent Processor #4,5,main] threw an Error. Shutting down now...
java.lang.OutOfMemoryError: GC overhead limit exceeded
at java.nio.ByteBuffer.wrap(ByteBuffer.java:373)
at java.nio.ByteBuffer.wrap(ByteBuffer.java:396)
at parquet.format.Statistics.setMin(Statistics.java:237)
at parquet.format.converter.ParquetMetadataConverter.toParquetStatistics(ParquetMetadataConverter.java:243)
at parquet.format.converter.ParquetMetadataConverter.addRowGroup(ParquetMetadataConverter.java:167)
at parquet.format.converter.ParquetMetadataConverter.toParquetMetadata(ParquetMetadataConverter.java:79)
at parquet.hadoop.ParquetFileWriter.serializeFooter(ParquetFileWriter.java:405)
at parquet.hadoop.ParquetFileWriter.writeMetadataFile(ParquetFileWriter.java:433)
at parquet.hadoop.ParquetFileWriter.writeMetadataFile(ParquetFileWriter.java:423)
at parquet.hadoop.ParquetOutputCommitter.writeMetaDataFile(ParquetOutputCommitter.java:58)
at parquet.hadoop.mapred.MapredParquetOutputCommitter.commitJob(MapredParquetOutputCommitter.java:43)
at org.apache.hadoop.mapred.OutputCommitter.commitJob(OutputCommitter.java:259)
at org.apache.hadoop.mapreduce.v2.app.commit.CommitterEventHandler$EventProcessor.handleJobCommit(CommitterEventHandler.java:253)
at org.apache.hadoop.mapreduce.v2.app.commit.CommitterEventHandler$EventProcessor.run(CommitterEventHandler.java:216)
at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
at java.lang.Thread.run(Thread.java:745)

This seems to have something to do with the _metadata file creation, as the parquet files are perfectly fine and usable. Also I'm not sure how to alleviate this (i.e. add more heap space) since the crash is outside the Map/Reduce tasks themselves but seems in the job/application controller itself.

Environment: CentOS, MapR,. Scalding
Reporter: hy5446

_{Note: This issue was originally created as PARQUET-282. Please see the migration documentation for further details.}

Start with parquet.hadoop.ParquetOutputCommitter.commitJob and ParquetFileWriter.writeMetadataFile, then trace ParquetMetadataConverter.toParquetMetadata and toParquetStatistics. Reproduce the failure with a large set of row groups and inspect the generated _metadata path; done means job commit completes without the reported OutOfMemoryError while the Parquet files remain usable.