Migrating Data with gptransfer
A newer version of this documentation is available. Click here to view the most up-to-date release of the Greenplum 5.x documentation.
Migrating Data with gptransfer
This topic describes how to use the gptransfer utility to transfer data between databases.
The gptransfer migration utility transfers Greenplum Database metadata and data from one Greenplum database to another Greenplum database, allowing you to migrate the entire contents of a database, or just selected tables, to another database. The source and destination databases may be in the same or a different cluster. Data is transferred in parallel across all the segments, using the gpfdist data loading utility to attain the highest transfer rates.
gptransfer handles the setup and execution of the data transfer. Participating clusters must already exist, have network access between all hosts in both clusters, and have certificate-authenticated ssh access between all hosts in both clusters.
The interface includes options to transfer one or more full databases, or one or more database tables. A full database transfer includes the database schema, table data, indexes, views, roles, user-defined functions, and resource queues. Configuration files, including postgres.conf and pg_hba.conf, must be transferred manually by an administrator. Extensions installed in the database with gppkg, such as MADlib, must be installed in the destination database by an administrator.
See the Greenplum Database Utility Guide for complete syntax and usage information for the gptransfer utility.
- The gptransfer utility can only be used with Greenplum Database. Apache HAWQ is not supported as a source or destination.
- The source and destination Greenplum clusters must both be version 4.2 or higher.
- At least one Greenplum instance must include the gptransfer utility in its distribution. If neither the source or destination includes gptransfer, you must upgrade one of the clusters to use gptransfer.
- The gptransfer utility can be run from the cluster with the source or destination database.
- The number of segments in the destination cluster must be greater than or equal to the number of hosts in the source cluster. The number of segments in the destination may be smaller than the number of segments in the source, but the data will transfer at a slower rate.
- The segment hosts in both clusters must have network connectivity with each other.
- Every host in both clusters must be able to connect to every other host with certificate-authenticated SSH. You can use the gpssh_exkeys utility to exchange public keys between the hosts of both clusters.
What gptransfer Does
gptransfer uses writable and readable external tables, the Greenplum gpfdist parallel data-loading utility, and named pipes to transfer data from the source database to the destination database. Segments on the source cluster select from the source database table and insert into a writable external table. Segments in the destination cluster select from a readable external table and insert into the destination database table. The writable and readable external tables are backed by named pipes on the source cluster's segment hosts, and each named pipe has a gpfdist process serving the pipe's output to the readable external table on the destination segments.
- creates a writable external table in the source database
- creates a readable external table in the destination database
- creates named pipes and gpfdist processes on segment hosts in the source cluster
- executes a SELECT INTO statement in the source database to insert the source data into the writable external table
- executes a SELECT INTO statement in the destination database to insert the data from the readable external table into the destination table
- optionally validates the data by comparing row counts or MD5 hashes of the rows in the source and destination
- cleans up the external tables, named pipes, and gpfdist processes
Fast Mode and Slow Mode
gptransfer sets up data transfer using the gpfdist parallel file serving utility, which serves the data evenly to the destination segments. Running more gpfdist processes increases the parallelism and the data transfer rate. When the destination cluster has the same or a greater number of segments than the source cluster, gptransfer sets up one named pipe and one gpfdist process for each source segment. This is the configuration for optimal data transfer rates and is called fast mode. The following figure illustrates a setup on a segment host when the destination cluster has at least as many segments as the source cluster.
The configuration of the input end of the named pipes differs when there are fewer segments in the destination cluster than in the source cluster. gptransfer handles this alternative setup automatically. The difference in configuration means that transferring data into a destination cluster with fewer segments than the source cluster is not as fast as transferring into a destination cluster of the same or greater size. It is called slow mode because there are fewer gpfdist processes serving the data to the destination cluster, although the transfer is still quite fast with one gpfdist per segment host.
When the destination cluster is smaller than the source cluster, there is one named pipe per segment host and all segments on the host send their data through it. The segments on the source host write their data to a writable external web table connected to a gpfdist process on the input end of the named pipe. This consolidates the table data into a single named pipe. A gpfdist process on the output of the named pipe serves the consolidated data to the destination cluster. The following figure illustrates this configuration.
On the destination side, gptransfer defines a readable external table with the gpfdist server on the source host as input and selects from the readable external table into the destination table. The data is distributed evenly to all the segments in the destination cluster.
Batch Size and Sub-batch Size
The degree of parallelism of a gptransfer execution is determined by two command-line options: --batch-size and --sub-batch-size. The --batch-size option specifies the number of tables to transfer in a batch. The default batch size is 2, which means that two table transfers are in process at any time. The minimum batch size is 1 and the maximum is 10. The --sub-batch-size parameter specifies the maximum number of parallel sub-processes to start to do the work of transferring a table. The default is 25 and the maximum is 50. The product of the batch size and sub-batch size is the amount of parallelism. If set to the defaults, for example, gptransfer can perform 50 concurrent tasks. Each thread is a Python process and consumes memory, so setting these values too high can cause a Python Out of Memory error. For this reason, the batch sizes should be tuned for your environment.
Preparing Hosts for gptransfer
When you install a Greenplum Database cluster, you set up all the master and segment hosts so that the Greenplum Database administrative user (gpadmin) can connect with SSH from every host in the cluster to any other host in the cluster without providing a password. The gptransfer utility requires this capability between every host in the source and destination clusters. First, ensure that the clusters have network connectivity with each other. Then, prepare a hosts file containing a list of all the hosts in both clusters, and use the gpssh-exkeys utility to exchange keys. See the reference for gpssh-exkeys in the Greenplum Database Utility Guide.
The host map file is a text file that lists the segment hosts in the source cluster. It is used to enable communication between the hosts in Greenplum clusters. The file is specified on the gptransfer command line with the --source-map-file=host_map_file command option. It is a required option when using gptransfer to copy data between two separate Greenplum clusters.
host1_name,host1_ip_addr host2_name,host2_ipaddr ...The file uses IP addresses instead of host names to avoid any problems with name resolution between the clusters.
gptransfer transfers data from user databases only; the postgres, template0, and template1 databases cannot be transferred. Administrators must transfer configuration files manually and install extensions into the destination database with gppkg.
The destination cluster must have at least as many segments as the source cluster has segment hosts. Transferring data to a smaller cluster is not as fast as transferring data to a larger cluster.
Transferring small or empty tables can be unexpectedly slow. There is significant fixed overhead in setting up external tables and communications processes for parallel data loading between segments that occurs whether or not there is actual data to transfer. It can be more efficient to transfer the schema and smaller tables to the destination database using other methods, then use gptransfer with the -t option to transfer large tables.
Full Mode and Table Mode
[ERROR]:- gptransfer: error: --full option specified but tables exist on destination system
By default, gptransfer fails if you attempt to transfer a table that already exists in the destination database:
[INFO]:-Validating transfer table set... [CRITICAL]:- gptransfer failed. (Reason='Table database.schema.table exists in database database .') exiting...
Override this behavior with the --skip-existing, --truncate, or --drop options.
The following table shows the objects that are copied in full mode and table mode.
|Object||Full Mode||Table Mode|
The --full option and the --schema-only option can be used together if you want to copy databases in phases, for example, during scheduled periods of downtime or low activity. Run gptransfer --full --schema-only ... to create the databases on the destination cluster, but with no data. Then you can transfer the tables in stages during scheduled down times or periods of low activity. Be sure to include the --truncate or --drop option when you later transfer tables to prevent the transfer from failing because the table already exists at the destination.
The -x option enables table locking. An exclusive lock is placed on the source table until the copy and validation, if requested, are complete.
- count – Compares the row counts for the tables in the source and destination databases.
- md5 – Sorts tables on both source and destination, and then performs a row-by-row comparison of the MD5 hashes of the sorted rows.
If the database is accessible during the transfer, be sure to add the -x option to lock the table. Otherwise, the table could be modified during the transfer, causing validation to fail.
[WARNING]:-Some tables failed to transfer. A list of these tables [WARNING]:-has been written to the file failed_transfer_tables_20140808_101813.txt [WARNING]:-This file can be used with the -f option to continue
The failed transfers file is in the format required by the -f option, so you can use it to start a new gptransfer session to retry the failed transfers.
Be careful not to exceed host memory by specifying too much parallelism with the --batch-size and --sub-batch-size command line options. Too many sub-processes can exhaust memory, causing a Python Out of Memory error. Start with a smaller batch size and sub-batch size, and increase based on your experiences.
Transfer a database in stages. First, run gptransfer with the --schema-only and -d database options, then transfer the tables in phases. After running gptransfer with the --schema-only option, be sure to add the --truncate or --drop option to prevent a failure because a table already exists.
Be careful choosing gpfdist and external table parameters such as the delimiter for external table data and the maximum line length. For example, don't choose a delimiter that can appear within table data.
If you have many empty tables to transfer, consider a DDL script instead of gptransfer. The gptransfer overhead to set up each table for transfer is significant and not an efficient way to transfer empty tables.
gptransfer creates table indexes before transferring the data. This slows the data transfer since indexes are updated at the same time the data is inserted in the table. For large tables especially, consider dropping indexes before running gptransfer and recreating the indexes when the transfer is complete.