Small file problem

Memory overhead

• File/Directory/Block is represented as object in namenode's memory and occupies 150 bytes. 10.000.000 small files ⇒ (10.000.000 blocks * replication factor + 10.000.000 file inodes ) * 150 bytes ⇒ 3 Gb
• Namenode is limited by main memory

• HAR: Hadoop archive; Layered file system on top of HDFS; Used for archiving (slow read)
• Sequence file: File name as key, contents as the value
• Consolidator
• HBase: Stores data in indexed SequenceFiles (HBase)
• Spark compaction: https://github.com/KeithSSmith/spark-compaction
• Filecrush: https://github.com/asdaraujo/filecrush