Planning Greenplum System Expansion

Careful planning will help to ensure a successful Greenplum expansion project.

The topics in this section help to ensure that you are prepared to execute a system expansion.

System Expansion Checklist is a checklist you can use to prepare for and execute the system expansion process.
Planning New Hardware Platforms covers planning for acquiring and setting up the new hardware.
Planning New Segment Initialization provides information about planning to initialize new segment hosts with gpexpand.
Planning Table Redistribution provides information about planning the data redistribution after the new segment hosts have been initialized.

System Expansion Checklist

This checklist summarizes the tasks for a Greenplum Database system expansion.

Table 1. Greenplum Database System Expansion Checklist
Online Pre-Expansion Tasks * System is up and available
	Devise and execute a plan for ordering, building, and networking new hardware platforms.
	Devise a database expansion plan. Map the number of segments per host, schedule the offline period for testing performance and creating the expansion schema, and schedule the intervals for table redistribution.
	Perform a complete schema dump.
	Install Greenplum Database binaries on new hosts.
	Copy SSH keys to the new hosts (`gpssh-exkeys`).
	Validate the operating system environment of the new hardware (`gpcheck`).
	Validate disk I/O and memory bandwidth of the new hardware (`gpcheckperf`).
	Validate that the master data directory has no extremely large files in the `pg_log` or `gpperfmon/data` directories.
	Validate that there are no catalog issues (`gpcheckcat`).
	Prepare an expansion input file (`gpexpand`).
Offline Expansion Tasks * The system is locked and unavailable to all user activity during this process.
	Validate the operating system environment of the combined existing and new hardware (`gpcheck`).
	Validate disk I/O and memory bandwidth of the combined existing and new hardware (`gpcheckperf`).
	Initialize new segments into the array and create an expansion schema (`gpexpand-i` `input_file`).
Online Expansion and Table Redistribution * System is up and available
	Before you start table redistribution, stop any automated snapshot processes or other processes that consume disk space.
	Redistribute tables through the expanded system (`gpexpand`).
	Remove expansion schema (`gpexpand -c`).
	Run `analyze` to update distribution statistics. During the expansion, use `gpexpand -a`, and post-expansion, use `analyze`.

Planning New Hardware Platforms

A deliberate, thorough approach to deploying compatible hardware greatly minimizes risk to the expansion process.

Hardware resources and configurations for new segment hosts should match those of the existing hosts. Work with Greenplum Platform Engineering before making a hardware purchase to expand Greenplum Database.

The steps to plan and set up new hardware platforms vary for each deployment. Some considerations include how to:

Prepare the physical space for the new hardware; consider cooling, power supply, and other physical factors.
Determine the physical networking and cabling required to connect the new and existing hardware.
Map the existing IP address spaces and developing a networking plan for the expanded system.
Capture the system configuration (users, profiles, NICs, and so on) from existing hardware to use as a detailed list for ordering new hardware.
Create a custom build plan for deploying hardware with the desired configuration in the particular site and environment.

After selecting and adding new hardware to your network environment, ensure you perform the burn-in tasks described in Verifying OS Settings.

Planning New Segment Initialization

Expanding Greenplum Database requires a limited period of system down time. During this period, run gpexpand to initialize new segments into the array and create an expansion schema.

The time required depends on the number of schema objects in the Greenplum system and other factors related to hardware performance. In most environments, the initialization of new segments requires less than thirty minutes offline.

Note: After you begin initializing new segments, you can no longer restore the system using backup files created for the pre-expansion system. When initialization successfully completes, the expansion is committed and cannot be rolled back.

Planning Mirror Segments

If your existing array has mirror segments, the new segments must have mirroring configured. If there are no mirrors configured for existing segments, you cannot add mirrors to new hosts with the gpexpand utility.

For Greenplum Database arrays with mirror segments, ensure you add enough new host machines to accommodate new mirror segments. The number of new hosts required depends on your mirroring strategy:

Spread Mirroring — Add at least one more host to the array than the number of segments per host. The number of separate hosts must be greater than the number of segment instances per host to ensure even spreading.
Grouped Mirroring — Add at least two new hosts so the mirrors for the first host can reside on the second host, and the mirrors for the second host can reside on the first. For more information, see About Segment Mirroring

Increasing Segments Per Host

By default, new hosts are initialized with as many primary segments as existing hosts have. You can increase the segments per host or add new segments to existing hosts.

For example, if existing hosts currently have two segments per host, you can use gpexpand to initialize two additional segments on existing hosts for a total of four segments and four new segments on new hosts.

The interactive process for creating an expansion input file prompts for this option; the input file format allows you to specify new segment directories manually, also. For more information, see Creating an Input File for System Expansion.

About the Expansion Schema

At initialization, gpexpand creates an expansion schema. If you do not specify a database at initialization (gpexpand -D), the schema is created in the database indicated by the PGDATABASE environment variable.

The expansion schema stores metadata for each table in the system so its status can be tracked throughout the expansion process. The expansion schema consists of two tables and a view for tracking expansion operation progress:

gpexpand.status
gpexpand.status_detail
gpexpand.expansion_progress

Control expansion process aspects by modifying gpexpand.status_detail. For example, removing a record from this table prevents the system from expanding the table across new segments. Control the order in which tables are processed for redistribution by updating the rank value for a record. For more information, see Ranking Tables for Redistribution.

Planning Table Redistribution

Table redistribution is performed while the system is online. For many Greenplum systems, table redistribution completes in a single gpexpand session scheduled during a low-use period. Larger systems may require multiple sessions and setting the order of table redistribution to minimize performance impact. Pivotal recommends completing the table redistribution in one session if possible.

Important: To perform table redistribution, your segment hosts must have enough disk space to temporarily hold a copy of your largest table. All tables are unavailable for read and write operations during redistribution.

The performance impact of table redistribution depends on the size, storage type, and partitioning design of a table. Per table, redistributing a table with gpexpand takes as much time as a CREATE TABLE AS SELECT operation does. When redistributing a terabyte-scale fact table, the expansion utility can use much of the available system resources, with resulting impact on query performance or other database workloads.

Managing Redistribution in Large-Scale Greenplum Systems

You can manage the order in which tables are redistributed by adjusting their ranking. See Ranking Tables for Redistribution. Manipulating the redistribution order can help adjust for limited disk space and restore optimal query performance.

When planning the redistribution phase, consider the impact of the exclusive lock taken on each table during redistribution. User activity on a table can delay its redistribution. Tables are unavailable during redistribution.

Systems with Abundant Free Disk Space

In systems with abundant free disk space (required to store a copy of the largest table), you can focus on restoring optimum query performance as soon as possible by first redistributing important tables that queries use heavily. Assign high ranking to these tables, and schedule redistribution operations for times of low system usage. Run one redistribution process at a time until large or critical tables have been redistributed.

Systems with Limited Free Disk Space

If your existing hosts have limited disk space, you may prefer to first redistribute smaller tables (such as dimension tables) to clear space to store a copy of the largest table. Available disk space on the original segments increases as each table is redistributed across the expanded array. When enough free space exists on all segments to store a copy of the largest table, you can redistribute large or critical tables. Redistribution of large tables requires exclusive locks; schedule this procedure for off-peak hours.

Also consider the following:

Run multiple parallel redistribution processes during off-peak hours to maximize available system resources.
When running multiple processes, operate within the connection limits for your Greenplum system. For information about limiting concurrent connections, see Limiting Concurrent Connections.

Redistributing Append-Optimized and Compressed Tables

gpexpand redistributes append-optimized and compressed append-optimized tables at different rates than heap tables. The CPU capacity required to compress and decompress data tends to increase the impact on system performance. For similar-sized tables with similar data, you may find overall performance differences like the following:

Uncompressed append-optimized tables expand 10% faster than heap tables.
zlib-compressed append-optimized tables expand at a significantly slower rate than uncompressed append-optimized tables, potentially up to 80% slower.
Systems with data compression such as ZFS/LZJB take longer to redistribute.

Important: If your system nodes use data compression, use identical compression on new nodes to avoid disk space shortage.

Redistributing Tables with Primary Key Constraints

There is a time period during which primary key constraints cannot be enforced between the initialization of new segments and successful table redistribution. Duplicate data inserted into tables during this period prevents the expansion utility from redistributing the affected tables.

After a table is redistributed, the primary key constraint is properly enforced again. If an expansion process violates constraints, the expansion utility logs errors and displays warnings when it completes. To fix constraint violations, perform one of the following remedies:

Clean up duplicate data in the primary key columns, and re-run gpexpand.
Drop the primary key constraints, and re-run gpexpand.

Redistributing Tables with User-Defined Data Types

You cannot perform redistribution with the expansion utility on tables with dropped columns of user-defined data types. To redistribute tables with dropped columns of user-defined types, first re-create the table using CREATE TABLE AS SELECT. After this process removes the dropped columns, redistribute the table with gpexpand.

Redistributing Partitioned Tables

Because the expansion utility can process each individual partition on a large table, an efficient partition design reduces the performance impact of table redistribution. Only the child tables of a partitioned table are set to a random distribution policy. The read/write lock for redistribution applies to only one child table at a time.

Redistributing Indexed Tables

Because the gpexpand utility must re-index each indexed table after redistribution, a high level of indexing has a large performance impact. Systems with intensive indexing have significantly slower rates of table redistribution.