Running a parallel direct path load v16

Suggest edits

You can further improve the performance of a direct path load by distributing the loading process over two or more sessions running concurrently. Each session runs a direct path load into the same table.

Since the same table is loaded from multiple sessions, the input records to load into the table must be divided among several data files. This way, each EDB*Loader session uses its own data file, and the same record isn't loaded into the table more than once.

Restrictions

The target table of a parallel direct path load is under the same restrictions as a direct path load run in a single session.

The restrictions on the target table of a direct path load are:

  • Triggers aren't permitted.
  • Check constraints aren't permitted.
  • Foreign key constraints on the target table referencing another table aren't permitted.
  • Foreign key constraints on other tables referencing the target table aren't permitted.
  • You must not partition the table.
  • Rules can exist on the target table, but they aren't executed.

In addition, you must specify the APPEND clause in the control file used by each EDB*Loader session.

Running a parallel direct path load

To run a parallel direct path load, run EDB*Loader in a separate session for each participant of the parallel direct path load. You must include the DIRECT=TRUE and PARALLEL=TRUE parameters when invoking each such EDB*Loader session.

Each EDB*Loader session runs as an independent transaction. Aborting and rolling back changes of one of the parallel sessions doesn't affect the loading done by the other parallel sessions.

Note

In a parallel direct path load, each EDB*Loader session reserves a fixed number of blocks in the target table using turns. Some of the blocks in the last allocated chunk might not be used, and those blocks remain uninitialized. A later use of the VACUUM command on the target table might show warnings about these uninitialized blocks, such as the following:

WARNING: relation "emp" page 98264 is uninitialized --- fixing

WARNING: relation "emp" page 98265 is uninitialized --- fixing

WARNING: relation "emp" page 98266 is uninitialized --- fixing

This behavior is expected and doesn't indicate data corruption.

Indexes on the target table aren't updated during a parallel direct path load. They are therefore marked as invalid after the load is complete. You must use the REINDEX command to rebuild the indexes.

This example shows the use of a parallel direct path load on the emp table.

Note

If you attempt a parallel direct path load on the sample emp table provided with EDB Postgres Advanced Server, you must first remove the triggers and constraints referencing the emp table. In addition, the primary key column, empno, was expanded from NUMBER(4) to NUMBER in this example to allow for inserting more rows.

This is the control file used in the first session:

LOAD DATA
  INFILE    '/home/user/loader/emp_parallel_1.dat'
  APPEND
  INTO TABLE emp
    FIELDS TERMINATED BY ',' OPTIONALLY ENCLOSED BY '"'
    TRAILING NULLCOLS
  (
    empno,
    ename,
    job,
    mgr,
    hiredate,
    sal,
    deptno,
    comm
  )

You must specify the APPEND clause in the control file for a parallel direct path load.

This example invokes EDB*Loader in the first session. You must specify the DIRECT=TRUE and PARALLEL=TRUE parameters.

$ /usr/edb/as16/bin/edbldr -d edb USERID=enterprisedb/password
CONTROL=emp_parallel_1.ctl DIRECT=TRUE PARALLEL=TRUE
WARNING: index maintenance will be skipped with PARALLEL load
EDB*Loader: Copyright (c) 2007-2021, EnterpriseDB Corporation.

The control file used for the second session appears as follows. It's the same as the one used in the first session, but it uses a different data file.

LOAD DATA
  INFILE '/home/user/loader/emp_parallel_2.dat'
  APPEND
  INTO TABLE emp
    FIELDS TERMINATED BY ',' OPTIONALLY ENCLOSED BY '"'
    TRAILING NULLCOLS
  (
    empno,
    ename,
    job,
    mgr,
    hiredate,
    sal,
    deptno,
    comm
  )

This control file is used in a second session:

$ /usr/edb/as16/bin/edbldr -d edb USERID=enterprisedb/password
CONTROL=emp_parallel_2.ctl DIRECT=TRUE PARALLEL=TRUE
WARNING: index maintenance will be skipped with PARALLEL load
EDB*Loader: Copyright (c) 2007-2021, EnterpriseDB Corporation.

EDB*Loader displays a message in each session when the load operation completes:

Successfully loaded (10000) records

This query shows that the index on the emp table was marked INVALID:

SELECT index_name, status FROM user_indexes WHERE table_name = 'EMP';
Output
  index_name | status
-------------+---------
   EMP_PK    | INVALID
 (1 row)
Note

user_indexes is the view of indexes compatible with Oracle databases owned by the current user.

Queries on the emp table don't use the index unless you rebuild it using the REINDEX command:

REINDEX INDEX emp_pk;

A later query on user_indexes shows that the index is now marked as VALID:

SELECT index_name, status FROM user_indexes WHERE table_name = 'EMP';
Output
  index_name | status
-------------+--------
   EMP_PK    | VALID
 (1 row)
← Prev

Running a direct path load

Next →

Performing remote loading