Tuesday, January 17, 2012

HBase (Intra) Row Transactions

As said in a previous post "HBase ensures that all new versions created by single Put operation for a particular rowkey are either all seen by other clients or seen by none."

Indeed HBase can execute atomic Put operations and atomic Delete operations (as well as a few specialized operations like Increment and Append).

What HBase cannot currently do is to execute a grouping of different operations atomically. For example you cannot execute a Put and Delete operation atomically.

HBASE-3584 and HBASE-5203 change that. It is now possible to group multiple Puts and Deletes for the same row key together as a single atomic operation. The combined operation is atomic even when the executing regionserver fails half way through the operation.

The client facing API looks like this:

HTable t = ...; 
byte[] row = ...;
RowMutation arm = new RowMutation(row);
Put p = new Put(row);
Delete d = new Delete(now);

RowMutation implements the Row interface and can hence itself be part of a multi row batch operation:

HTable t = ...; 
byte[] row1, row2;
RowMutation arm1 = new RowMutation(row1);
RowMutation arm2 = new RowMutation(row2);
List<Row> rows = ...;

But note that this multi row batch is not atomic between different rows.

No comments:

Post a Comment