Planning Greenplum System Expansion
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
Online Pre-Expansion Tasks* System is up and available
|Devise and execute a plan for ordering, building, and networking new hardware platforms, or provisioning cloud resources.|
|Devise a database expansion plan. Map the number of segments per host, schedule the downtime 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 disk I/O and memory bandwidth of the new hardware or cloud resources (gpcheckperf).|
|Validate that the master data directory has no extremely large files in the pg_log or gpperfmon/data directories.|
Offline Pre-Expansion Tasks* The system is unavailable to all user activity during this process.
|Validate that there are no catalog issues (gpcheckcat).|
|Validate disk I/O and memory bandwidth of the combined existing and new hardware or cloud resources (gpcheckperf).|
Online Segment Instance Initialization* System is up and available
|Prepare an expansion input file (gpexpand).|
|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.
Back Up Databases* System is up and available
|Back up databases using the gpbackup utility. Backups you created before you began the system expansion cannot be restored to the newly expanded system because the gprestore utility can only restore backups to a Greenplum Database system with the same number of segments.|
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.
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 expand-nodes.html#topic19.
Planning New Segment Initialization
Expanding Greenplum Database can be performed when the system is up and available. Run gpexpand to initialize new segment instances 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.
Planning Mirror Segments
If your existing system 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 more information about segment mirroring, see About Segment Mirroring.
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. You can specify this type of mirroring during system initialization or when you enable segment mirroring for an existing system.
- 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. This is the default type of mirroring if you enable segment mirroring during system initialization.
- Block Mirroring — Adding one or more blocks of host systems. For example add a block of four or eight hosts. Block mirroring is a custom mirroring configuration. For more information about block mirroring, see 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 initialize four new segments on new hosts.
The interactive process for creating an expansion input file prompts for this option; you can also specify new segment directories manually in the input configuration file. For more information, see Creating an Input File for System Expansion.
About the Expansion Schema
At initialization, the gpexpand utility creates an expansion schema named gpexpand in the postgres database.
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:
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. Complete the table redistribution in one session if possible.
The performance impact of table redistribution depends on the size, storage type, and partitioning design of a table. For any given table, redistributing it with gpexpand takes as much time as a CREATE TABLE AS SELECT operation would. When redistributing a terabyte-scale fact table, the expansion utility can use much of the available system resources, which could affect query performance or other database workloads.
Managing Redistribution in Large-Scale Greenplum Systems
When planning the redistribution phase, consider the impact of the ACCESS EXCLUSIVE lock taken on each table, and the table data redistribution method. User activity on a table can delay its redistribution, but also tables are unavailable for user activity during redistribution.
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 for high-priority queries sooner.
Table Redistribution Methods
- rebuild - Create a new table, copy all the data from the old to the new table, and replace the old table. This is the default. The rebuild method is similar to creating a new table with a CREATE TABLE AS SELECT command. During data redistribution, an ACCESS EXCLUSIVE lock is acquired on the table.
- move - Scan all the data and perform an UPDATE operation to move rows as needed to different segment instances. During data redistribution, an ACCESS EXCLUSIVE lock is acquired on the table. In general, this method requires less disk space, however, it creates obsolete table rows and might require a VACUUM operation on the table after the data redistribution. Also, this method updates indexes one row at a time, which can be much slower than rebuilding the index with the CREATE INDEX command.
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.
- 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
- Uncompressed append-optimized tables expand 10% faster than heap tables.
- Append-optimized tables that are defined to use data compression 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.
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.