Discuss with expertTry it for free
© 2026 Direct Cursus Technology L.L.C.

Resource groups

Written by
Updated at July 6, 2026
View in Markdown

By using resource groups, users with the mdb_admin role can set quotas for the following resources allocated for query processing:

  • CPU computing resources
  • Amount of RAM
  • Number of concurrent transactions

A resource group may include several roles that will share their quotas.

Yandex MPP Analytics for PostgreSQL gets a transaction query from the user and matches its resource requirements with the unused quotas in the resource group. If the quotas are sufficient, the query execution begins immediately. If not, the query waits for other queries to be completed and to free up the resources. Queries are queued based on the first in, first out (FIFO) principle.

Clusters with Apache Cloudberry™ and Greenplum® DBMS’s have different resource group parameters, default resource groups, and RAM allocation.

Learn how to manage resource groups in this section.

Resource group parameters

Name

Description

CONCURRENCY

Maximum number of concurrent transactions a resource group can process. Both active and pending transactions are counted. The minimum value is 0 (no transactions processed); the maximum value is taken from the max_connections setting, 20 by default.

CPU_MAX_PERCENT

Maximum CPU percentage a resource group can use on a segment. The minimum value is 1; the maximum value is 100. The default value is -1 (not set). If you set CPU_MAX_PERCENT, the CPUSET parameter is ignored.

For more information, see Specifics of CPU resource allocation.

CPU_WEIGHT

Resource group priority for CPU allocation. The minimum value is 1; the maximum value is 500. The default value is 100.

CPUSET

Numbers of logical CPU cores allocated to a resource group. It is set for the master and segments as separate values or ranges. The default value is -1 (not set). If you set CPU_MAX_PERCENT, the CPUSET parameter is ignored.

For example, to allocate core 1 for the master, and cores 1, 3, 4 for segments, specify CPUSET='1;1,3-4'.

Warning

For stable cluster operation, do not allocate core 0.

IO_LIMIT

Maximum read and write throughput, as well as the maximum number of IOPS for a resource group. Set separately for each table space:

  • wbps: Maximum write throughput, in MBps.
  • rbps: Maximum read throughput, in MBps.
  • wiops: Maximum writes per second.
  • riops: Maximum reads per second.

For each IO_LIMIT parameter, the valid values range from 2 to 4294967295. The default value is -1 (not limited). You may skip some of the parameters. The parameters you skip take their defaults.

Here is an example: IO_LIMIT='pg_default:wbps=1000,rbps=1000,wiops=100,riops=100'

MEMORY_QUOTA

Amount of RAM, in MB, a resource group can use on a segment. The default value is -1 (not limited); the statement_mem setting will be used.

For more information, see Specifics of RAM allocation.

MIN_COST

Minimum cost of query plan sufficient to assign a query to a resource group. The default value is 0.

Name

Description

MEMORY_AUDITOR

Memory auditor used by a resource group. Yandex MPP Analytics for PostgreSQL uses vmtracker as the default value.

CONCURRENCY

Maximum number of concurrent transactions a resource group can process. Both active and pending transactions are counted. The default value is 20

CPU_RATE_LIMIT

CPU percentage a resource group can use on a segment. The minimum value is 1; the maximum value is 100. The value is automatically reset to -1 if you specify CPUSET.

CPUSET

Numbers of allocated CPU cores. The numbers or intervals are enclosed in quotes, e.g., '1,3-4'. Use the numbers that exist in the system and are not allocated to other resource groups. The value is automatically reset to -1 if you specify CPU_RATE_LIMIT.

Warning

For stable cluster operation, do not allocate core 0.

MEMORY_LIMIT

RAM percentage a resource group can use on a segment. The minimum value is 0 (default); the maximum value is 100. When the value is 0, the resource group can only use non-allocated RAM to run queries. The total of MEMORY_LIMIT values for all resource groups may not exceed 100.

MEMORY_SHARED_QUOTA

Percentage of MEMORY_LIMIT available to all resource group transactions whenever needed.

For more information, see Specifics of RAM allocation.

MEMORY_SPILL_RATIO

RAM usage threshold for transactions as a percentage of MEMORY_LIMIT. If exceeded, temporary files are created on the disk to process transactions. The minimum value is 0 (default); the maximum value is 100. If MEMORY_LIMIT equals 0, this parameter must also equal 0.

Default resource groups

When you create a cluster with an Apache Cloudberry™ DBMS, the following resource groups are created:

  • admin_group: Assigned to the roles with the SUPERUSER attribute.

  • system_group: Used by Apache Cloudberry™ system processes. You cannot assign this group manually.

  • mdb_admin_group: Group for administrators. By default, it is assigned to the administrator whose username and password were specified when creating the cluster.

    If you reach the resource limit for this group, the admin will not be able to connect to the cluster and terminate resource-hungry processes. This is why we recommend assigning mdb_admin_group only to users with the mdb_admin role.

  • default_group: Assigned to other roles if no resource group was set when creating them.

Resource group parameter values:

Parameter admin_group system_group mdb_admin_group default_group
CONCURRENCY 20 0 20 20
CPU_MAX_PERCENT -1 10 1 20
CPU_WEIGHT 100 100 100 100
CPUSET -1 -1 -1 -1
IO_LIMIT -1 -1 -1 -1
MEMORY_QUOTA -1 -1 -1 -1
MIN_COST 0 500 0 500

You cannot change parameter values for admin_group.

For more information about managing roles, see Users and roles and this Apache Cloudberry™ guide.

When you create a cluster with a Greenplum® DBMS DBMS, the following resource groups are created:

  • admin_group: Service group assigned to the monitor and gpadmin service roles.

  • mdb_admin_group: Group for administrators. By default, it is assigned to the administrator whose username and password were specified when creating the cluster.

    If the administrator and other users belong to the same resource group, and all its resources are used up, the administrator will not be able to connect to the database and interrupt resource-intensive processes manually. This is why we recommend assigning this resource group only to the users with the mdb_admin role.

  • default_group: Shared group. It is assigned to a newly created role unless it is assigned another explicit resource group.

Resource group parameter values:

Parameter admin_group mdb_admin_group default_group
MEMORY_AUDITOR vmtracker vmtracker vmtracker
CONCURRENCY 10 20 20
CPU_RATE_LIMIT 10 1 30
CPUSET -1 -1 -1
MEMORY_LIMIT 10 0 0
MEMORY_SHARED_QUOTA 80 80 80
MEMORY_SPILL_RATIO 0 0 0

You cannot change parameter values for admin_group.

For more information about managing roles, see Users and roles and this Greenplum® guide.

Specifics of RAM allocation

  • If there is a limit on RAM per resource group (the MEMORY_QUOTA parameter), RAM management takes place at the segment level.

    RAM per query is decided by the gp_resgroup_memory_query_fixed_mem setting. If it is not set, RAM per query is calculated using this formula: MEMORY_QUOTA ÷ CONCURRENCY.

    Read more about the gp_resgroup_memory_query_fixed_mem setting in this Greenplum® guide.

    For more information on additional criteria affecting RAM per query, see this Apache Cloudberry™ guide.

  • If there is no limit on RAM per resource group, RAM management takes place at the query level. RAM per query is decided by the statement_mem setting.

  • You can allocate a segment's RAM share for each resource group using the MEMORY_LIMIT parameter. This percentage is split into the shared quota and guaranteed quota:

    • The shared quota is set in the MEMORY_SHARED_QUOTA parameter and can be used by all the resource group transactions. Yandex MPP Analytics for PostgreSQL distributes the shared quota to transactions based on the first in, first out principle.

    • The guaranteed quota is allocated to each transaction and calculated by the following formula:

      (MEMORY_LIMIT − MEMORY_SHARED_QUOTA) ÷ CONCURRENCY

  • If the total of MEMORY_LIMIT values of all resource groups is below 100, Yandex MPP Analytics for PostgreSQL distributes unallocated RAM to any transactions from any resource groups based on the first in, first out principle.

    The transaction uses non-allocated memory when both conditions are met:

    • The guaranteed memory share for the transaction is used up.
    • The shared quota for the resource group is used up.

    For your cluster to run in a stable and efficient manner, leave 10% to 20% of memory unallocated.

Specifics of CPU resource allocation

You can allocate CPU resources using one of the following methods:

  • By setting a CPU percentage per resource group on a segment. In Apache Cloudberry™, the parameter used to set CPU percentage is CPU_MAX_PERCENT, while in Greenplum®, it is CPU_RATE_LIMIT.

    If a resource group is idle, its resources get re-allocated towards highly loaded groups. In Apache Cloudberry™, resource group priority (the CPU_WEIGHT parameter) is another factor affecting CPU resource allocation. As soon as the resource group stops idling, it gets its re-allocated CPU resources back.

  • By specifying the numbers of logical CPU cores you are allocating to the resource group (in the CPUSET parameter). These cores will be unavailable to other resource groups even if the resource group they are allocated to is idle. Therefore, we recommend minimizing the number of resource groups with allocated cores.

    Warning

    For stable cluster operation, do not allocate core 0.

CPU allocation methods may vary from one resource group to another.

For more on resource groups, see:

Greenplum® and Greenplum Database® are registered trademarks or trademarks of Broadcom Inc. in the United States and/or other countries.

Apache® and Apache Cloudberry™ are registered trademarks or trademarks of the Apache Software Foundation in the United States and/or other countries.

Previous
Backups
Next
Sharding