Migrating Data from Greenplum 4.3 or 5
Migrating Data from Greenplum 4.3 or 5
You can migrate data from Greenplum Database 4.3 or 5 to Greenplum 6 using the standard backup and restore procedures, gpbackup and gprestore, or by using gpcopy.
This topic identifies known issues you may encounter when moving data from Greenplum 4.3 to Greenplum 6. You can work around these problems by making needed changes to your Greenplum 4.3 databases so that you can create backups that can be restored successfully to Greenplum 6.
Preparing the Greenplum 6 Cluster
- Install and initialize a new Greenplum Database 6 cluster using the version 6
gpinitsystem utility. Note: gprestore only supports restoring data to a cluster that has an identical number of hosts and an identical number of segments per host, with each segment having the same content_id as the segment in the original cluster. Use gpcopy if you need to migrate data to a different-sized Greenplum 6 cluster.Note: Set the Greenplum Database 6 timezone to a value that is compatible with your host systems. Setting the Greenplum Database timezone prevents Greenplum Database from selecting a timezone each time the cluster is restarted. See Configuring Timezone and Localization Settings for more information.
- Install the latest release of the Greenplum Backup and Restore utilities, available to download from Pivotal Network.
- If you intend to install Greenplum Database 6 on the same hardware as your 4.3 system,
you will need enough disk space to accommodate over five times the original data set (two
full copies of the primary and mirror data sets, plus the original backup data in ASCII
format) in order to migrate data with gpbackup and
gprestore. Keep in mind that the ASCII backup data will require more
disk space than the original data, which may be stored in compressed binary format.
Offline backup solutions such as Dell EMC Data Domain can reduce the required disk space
on each host.
If you want to migrate your data on the same hardware but do not have enough free disk space on your host systems, gpcopy provides the --truncate-source-after option to truncate each source table after copying the table to the destination cluster and validating that the copy succeeded. This reduces the amount of free space needed to migrate clusters that reside on the same hardware. See Migrating Data with gpcopy for more information.
- Install any external modules used in your Greenplum 4.3 system in the Greenplum 6 system before you restore the backup, for example MADlib or PostGIS. If versions of the external modules are not compatible, you may need to exclude tables that reference them when restoring the Greenplum 4.3 backup to Greenplum 6.
- When restoring language-based user-defined functions, the shared object file must be in the location specified in the CREATE FUNCTION SQL command and must have been recompiled on the Greenplum 6 system. This applies to user-defined functions, user-defined types, and any other objects that use custom functions, such as aggregates created with the CREATE AGGREGATE command.
- Greenplum 6 provides resource groups, an alternative to managing resources using resource queues. Setting the gp_resource_manager server configuration parameter to queue or group selects the resource management scheme the Greenplum Database system will use. The default is queue, so no action is required when you move from Greenplum version 4.3 to version 6. To more easily transition from resource queues to resource groups, you can set resource groups to allocate and manage memory in a way that is similar to resource queue memory management. To select this feature, set the MEMORY_LIMIT and MEMORY_SPILL_RATIO attributes of your resource groups to 0. See Using Resource Groups for information about enabling and configuring resource groups.
Preparing Greenplum 4.3 and 5 Databases for Backup
Following are some issues that are known to cause errors when restoring a Greenplum 4.3 or 5 backup to Greenplum 6. Keep a list of any changes you make to the Greenplum 4.3 or 5 database to enable migration so that you can fix them in Greenplum 6 after restoring the backup.
- If you have configured PXF in your Greenplum Database 5 installation, review Migrating PXF from Greenplum 5 to plan for the PXF migration.
- Greenplum Database version 6 removes support for the gphdfs protocol. If you have created external tables that use gphdfs, remove the external table definitions and (optionally) recreate them to use Pivotal Extension Framework (PXF) before you migrate the data to Greenplum 6. Refer to Migrating gphdfs External Tables to PXF in the PXF documentation for the migration procedure.
- References to catalog tables or their attributes can cause a restore to fail due to
catalog changes from Greenplum 4.3 or 5 to Greenplum 6. Here are some catalog issues to be
aware of when migrating to Greenplum 6:
- In the pg_class system table, the reltoastidxid column has been removed.
- In the pg_stat_replication system table, the procpid column is renamed to pid.
- In the pg_stat_activity system table, the procpid column is renamed to pid. The current_query column is replaced by two columns state: the state of the backend, and query: the last run query, or currently running query if state is active.
- In the gp_distribution_policy system table, the attrnums column is renamed to distkey and its data type is changed to int2vector. A backend function pg_get_table_distributedby() was added to get the distribution policy for a table as a string.
- In the session_level_memory_consumption system view, the __gp_localid and __gp_masterid columns are removed.
- Filespaces are removed in Greenplum 6. The pg_filespace and pg_filespace_entry system tables are removed. Any reference to pg_filespace or pg_filespace_entry will fail in Greenplum 6.
- Restoring a Greenplum 4 backup can fail due to lack of dependency checking in Greenplum 4 catalog tables. For example, restoring a UDF can fail if it references a custom data type that is created later in the backup file.
- The int4_avg_accum() function signature changed in Greenplum 6 from int4_avg_accum(bytea, integer) to int4_avg_accum(bigint, integer). This function is the state transition function (sfunc) called when calculating the average of a series of 4-byte integers. If you have created a custom aggregate in a previous Greenplum release that called the built-in int4_avg_accum() function, you will need to revise your aggregate for the new signature.
- The string_agg(expression) function has been removed from Greenplum 6. The function concatenates text values into a string. You can replace the single argument function with the function string_agg(expression, delimiter) and specify an empty string as the delimiter, for example string_agg(txt_col1, '').
- gpbackup saves the distribution policy and distribution key for each table in the backup so that data can be restored to the same segment. If a table's distribution key in the Greenplum 4.3 or 5 database is incompatible with Greenplum 6, gprestore cannot restore the table to the correct segment in the Greenplum 6 database. This can happen if the distribution key in the older Greenplum release has columns with data types not allowed in Greenplum 6 distribution keys, or if the data representation for data types has changed or is insufficient for Greenplum 6 to generate the same hash value for a distribution key. You should correct these kinds of problems by altering distribution keys in the tables before you back up the Greenplum database.
- Greenplum 6 requires primary keys and unique index keys to match a table's distribution
key. The leaf partitions of partitioned tables must have the same distribution policy as
the root partition. These known issues should be corrected in the source Greenplum
database before you back up the database:
- If the primary key is different than the distribution key for a table, alter the table to either remove the primary key or change the primary key to match the distribution key.
- If the key columns for a unique index are not a subset of the distribution key columns, before you back up the source database, drop the index and, optionally, recreate it with a compatible key.
- If a partitioned table in the source database has a DISTRIBUTED BY distribution policy, but has leaf partitions that are DISTRIBUTED RANDOMLY, alter the leaf tables to match the root table distribution policy before you back up the source database.
- Columns of type abstime, reltime,
tinterval, money, or anyarray are not
supported as distribution keys in Greenplum 6.
If you have tables distributed on columns of type abstime, reltime, tinterval, money, or anyarray, use the ALTER TABLE command to set the distribution to RANDOM before you back up the database. After the data is restored, you can set a new distribution policy.
- Greenplum 6 no longer automatically converts from the deprecated timestamp format YYYYMMDDHH24MISS. The format could not be parsed unambiguously in previous Greenplum Database releases. You can still specify the YYYYMMDDHH24MISS format in conversion functions such as to_timestamp and to_char for compatibility with other database systems. You can use input formats for converting text to date or timestamp values to avoid unexpected results or query execution failures.
- Creating a table using the CREATE TABLE AS command in Greenplum 4.3 or 5 could create a table with a duplicate distribution key. The gpbackup utility saves the table to the backup using a CREATE TABLE command that lists the duplicate keys in the DISTRIBUTED BY clause. Restoring this backup will cause a duplicate distribution key error. The CREATE TABLE AS command was fixed in Greenplum 5.10 to disallow duplicate distribution keys.
- Only Boolean operators can use Boolean negators. In Greenplum Database 4.3 and 5 it was
possible to create a non-Boolean operator that specifies a Boolean negator function. For
example, this CREATE OPERATOR command creates an integer
@@ operator with a Boolean
CREATE OPERATOR public.@@ ( PROCEDURE = int4pl, LEFTARG = integer, RIGHTARG = integer, NEGATOR = OPERATOR(public.!!) );
If you restore a backup containing an operator like this to a Greenplum 6 system, gprestore produces an error: ERROR: only boolean operators can have negators (SQLSTATE 42P13).
- In Greenplum Database 4.3 and 5, the undocumented server configuration parameter allow_system_table_mods could have a value of none, ddl, dml, or all. In Greenplum 6, this parameter has changed to a Boolean value, with a default value of false. If there are any references to this parameter in the source database, remove them to prevent errors during the restore.
Backing Up and Restoring a Database
First use gpbackup to create a --metadata-only backup from the source Greenplum database and restore it to the Greenplum 6 system. This helps find any additional problems that are not identified in Preparing Greenplum 4.3 and 5 Databases for Backup. Refer to the Greenplum Backup and Restore documentation for syntax and examples for the gpbackup and gprestore utilities.
Review the gprestore log file for error messages and correct any remaining problems in the source Greenplum database.
When you are able to restore a metadata backup successfully, create the full backup and then restore it to the Greenplum 6 system, or use gpcopy to transfer the data. If needed, use the gpbackup or gprestore filter options to omit schemas or tables that cannot be restored without error.
If you use gpcopy to migrate data, initiate the gpcopy operation from the Greenplum 4.3 or 5 cluster. See Migrating Data with gpcopy for more information.
After restoring, you can redistribute these tables with the gp_use_legacy_hashops parameter set to off so that the tables use the new Greenplum 6 hash operators. See Working With Hash Operator Classes in Greenplum 6 for more information and examples.
Completing the Migration
Migrate any tables you skipped during the restore using other methods, for example using the COPY TO command to create an external file and then loading the data from the external file into Greenplum 6 with the COPY FROM command.
Recreate any objects you dropped in the Greenplum 4.3 or 5 database to enable migration, such as external tables, indexes, user-defined functions, or user-defined aggregates.
Here are some additional items to consider to complete your migration to Greenplum 6.
- If you are migrating from Greenplum Database 4.3.27 or an earlier 4.3.x release and have configured PgBouncer in your Greenplum Database installation, you must migrate to the new PgBouncer when you upgrade Greenplum Database. Refer to Migrating PgBouncer for specific migration instructions.
- Greenplum Database 5 and 6 remove automatic casts between the text type and other data types. After you migrate from Greenplum Database version 4.3 to version 6, this changed behavior may impact existing applications and queries. Refer to About Implicit Text Casting in Greenplum Database for information, including a discussion about supported and unsupported workarounds.
- After migrating data you may need to modify SQL scripts, administration scripts, and user-defined functions as necessary to account for changes in Greenplum Database version 6. Review the Pivotal Greenplum 6.0.0 Release Notes for features and changes that may necessitate post-migration tasks.
- To use the new Greenplum 6 default hash operator classes, use the command ALTER TABLE <table> SET DISTRIBUTED BY (<key>) to redistribute tables restored from Greenplum 4.3 or 5 backups. The gp_use_legacy_hashops parameter must be set to off when your execute the command. See Working With Hash Operator Classes in Greenplum 6 for more information about hash operator classes.
Working With Hash Operator Classes in Greenplum 6
Greenplum 6 has new jump consistent hash operators that map distribution keys for distributed tables to the segments. The new hash operators enable faster database expansion because they don't require redistributing rows unless they map to a different segment. The hash operators used in Greenplum 4.3 and 5 are present in Greenplum 6 as non-default legacy hash operator classes. For example, for integer columns, the new hash operator class is named int_ops and the legacy operator class is named cdbhash_int_ops.
This example creates a table using the legacy hash operator class cdbhash_int_ops.
test=# SET gp_use_legacy_hashops=on; SET test=# CREATE TABLE t1 ( c1 integer, c2 integer, p integer ) DISTRIBUTED BY (c1); CREATE TABLE test=# \d+ t1 Table "public.t1" Column | Type | Modifiers | Storage | Stats target | Description --------+---------+-----------+---------+--------------+------------- c1 | integer | | plain | | c2 | integer | | plain | | p | integer | | plain | | Distributed by: (c1)
Notice that the distribution key is c1. If the gp_use_legacy_hashops parameter is on and the operator class is a legacy operator class, the operator class name is not shown. However, if gp_use_legacy_hashops is off, the legacy operator class name is reported with the distribution key.
test=# SET gp_use_legacy_hashops=off; SET test=# \d+ t1 Table "public.t1" Column | Type | Modifiers | Storage | Stats target | Description --------+---------+-----------+---------+--------------+------------- c1 | integer | | plain | | c2 | integer | | plain | | p | integer | | plain | | Distributed by: (c1 cdbhash_int4_ops)
The operator class name is reported only when it does not match the setting of the gp_use_legacy_hashops parameter.
test=# SHOW gp_use_legacy_hashops; gp_use_legacy_hashops ----------------------- off (1 row) test=# ALTER TABLE t1 SET DISTRIBUTED BY (c1); ALTER TABLE test=# \d+ t1 Table "public.t1" Column | Type | Modifiers | Storage | Stats target | Description --------+---------+-----------+---------+--------------+------------- c1 | integer | | plain | | c2 | integer | | plain | | p | integer | | plain | | Distributed by: (c1)
To verify the default jump consistent operator class has been used, set gp_use_legacy_hashops to on before you show the table definition.
test=# SET gp_use_legacy_hashops=on; SET test=# \d+ t1 Table "public.t1" Column | Type | Modifiers | Storage | Stats target | Description --------+---------+-----------+---------+--------------+------------- c1 | integer | | plain | | c2 | integer | | plain | | p | integer | | plain | | Distributed by: (c1 int4_ops)