Hadoop file_fdw is a slightly modified version of PostgreSQL 9.3's file_fdw module. This module extends on the original file_fdw to make it compatible with CitusDB's native Hadoop integration. The changes are still in Beta form.
To get started on using file_fdw on Hadoop, please see our documentation page on http://citusdata.com/docs/sql-on-hadoop! If you'd like to simply use file_fdw on distributed files instead, please check out http://citusdata.com/docs/foreign-data. Finally, you can find a performance evaluation of foreign data wrappers on our blog post.
The easiest way to get started with using file_fdw on Hadoop is going over the documentation pages. To build from scratch on POSIX compliant systems (Linux and OS X), you need to include the pg_config directory path in your make command. This path is typically the same as your CitusDB installation's bin/ directory path. For example:
PATH=/opt/citusdb/3.0/bin/:$PATH make USE_PGXS=1
sudo PATH=/opt/citusdb/3.0/bin/:$PATH make USE_PGXS=1 install
This module makes three changes to the original file_fdw that are still in Beta form. First, we add another option to file_fdw to accept hdfs_directory_path as an argument. Users creating the distributed foreign table on the master node use this option to associate the table with an HDFS directory path.
Second, file_fdw skips a malformed line and emits a debug message instead of erroring out on it. This is similar to Apache Hive's behavior, but isn't safe. In the future, we may error out if the number of malformed lines exceeds a certain threshold.
Third, CitusDB currently associates one HDFS block with one foreign table, and executes the entire SQL query locally on that block. If bytes for the last record in an HDFS block spill over to the next one, we currently don't fetch those bytes and instead skip the last record.
For all types of questions and comments about CitusDB's Hadoop integration or our changes to this wrapper, please contact us at http://citusdata.com/contact