Friday, December 27, 2013

How HDFS protects your data

Often we get questions how HDFS protects data and what the mechanisms are to prevent data corruption. Eric Sammer explain this en detail in Hadoop Operations.

Additional to the points below, you can also have a second cluster to sync the files, simply to prevent human being failures, like deleting a subset of data. If you have enough space in your cluster, enabling the trash per core-site.xml and setting to a higher value then a day helps too.

<description>Number of minutes after which the checkpoint
gets deleted. If zero, the trash feature is disabled. 1440 means 1 day

<description>Number of minutes between trash checkpoints.
Should be smaller or equal to fs.trash.interval.
Every time the checkpointer runs it creates a new checkpoint
out of current and removes checkpoints created more than
fs.trash.interval minutes ago.

HDFS is designed to protect data in different ways to minimize the risk of data loss with a valuable write speed. This enables in some circumstances HDFS as a NAS replacement for large files with the possibility to quickly access the stored data. The illustration below simplify the data flow:

HDFS has per default the following mechanisms implemented:
  1. Data written to files in HDFS is split up into chunks (usually 128MB in size). Each chunk (called a block) is replicated three times, by default, to three different machines. Multiple copies of a block are never written to the same machine. Replication level is configurable per file. 
  2. HDFS actively monitors the number of available replicas of each block, compared to the intended replication level. If, for some reason, a disk or node in the cluster should become unavailable, the filesystem will repair the missing block(s) by creating new replicas from the remaining copies of the data.
  3. HDFS can be (and normally is) configured to place block replicas across multiple racks of machines to protect against catastrophic failure of an entire rack or its constituent network infrastructure. This is called RackAwareness and should reflect your topology.
  4. Each block has an associated checksum computed on write, which is verified on all subsequent reads. Additionally, to protect against "bit rot" of files (and their blocks) that are not regularly read, the filesystem automatically verifies all checksums of all blocks on a regular basis. Should any checksum not verify correctly, HDFS will automatically detect this, discard the bad block, and create a new replica of the block.
  5. Filesystem metadata - information about object ownership, permissions, replication level, path, and so on - is served by a highly available pair of machines (i.e. namenodes) in CDH4. Updates to metadata are maintained in a traditional write-ahead transaction log that guarantees durability of changes to metadata information. The transaction log can be written to multiple physical disks and, in a highly available configuration, is written to multiple machines.
  6. HDFS block replicas are written in a synchronous, in-line replication pipeline. That is, when a client application receives a successful response from the cluster that a write was successful, it is true that at least a configurable minimum number of replicas are also complete. This eliminates the potential failure case of asynchronous replication where a client could complete a write to a node, receive a successful response, only for that one node to fail before it's able to replicate to another node.
  7. HDFS is fully instrumented with metric collection and reporting so monitoring systems (such as Cloudera Manager) can generate alerts when faults are detected. Metrics related to data integrity include unresponsive nodes in the cluster, failed disks, missing blocks, corrupt blocks, under-replicated blocks, and so on. Cloudera Manager has extensive HDFS-specific monitoring configured out of the box.
  8. HDFS supports directory-level filesystem quotas to protect against accidental denial of service attacks that could otherwise cause critical applications to fail to write data to the cluster.
  9. All higher level data storage and processing systems in CDH (MapReduce, HBase, Hive, Pig, Impala) use HDFS as their underlying storage substrate and, as a result, have the same data protection guarantees described above.