How to keep data consistency between Machines using Hadoop?

Question

My Configuration:
Hadoop 2.2.0, two computer, one is master,another is node1.

I want to know following scene: If node1 is down by some reason, but I don't know node1 can't work, then I use hadoop command to put a file, such as:
$ hadoop fs -put graph.txtgraphin/graph.txt

I know graph.txt file will be put master machine, but node1 computer don't contains this file. After some time, node1 machine can repair, there will be inconsistent because of graph.txt file, how to realize consistency between master machine and node1 machine?

Answer 1

You can't achieve data consistency on your cluster configuration. To do this you need at least 3 data nodes and enabled replication with level 3 ( dfs.replication property in hdfs-site.xml).

Answer 2

It does not seem like your hadoop -put ... command will even complete - the master is not receiving the file at any point. It instructs the node1 to connect to the client, after asking the node1 if it is in a state that it can receive data to be written, which depends on several other daemons being available and several successful internal RPC calls. A file does not get written to the master for storage - the only thing the master has is a transaction log telling it that client A sent a request to -put file graph.txt file in the HDFS storage location hdfs://user/$submitterusername/graphin/ and whether that request succeeded. There is another set of processes that tell the node1 to tell master periodically what files or pieces of files it has, and that record is stored in another file, but nothing goes to master for storing.

So if node1 is down, you will know when you try and -put something, since its your only data node - if it was not able to receive something for storage, master would know there is no node alive to put anything in, and tell you "I cant put that in HDFS because there are no storage nodes that are alive" but in Java exceptions instead of plain language.