Overview

Updated on 2024-12-13 GMT+08:00

View PDF

Doris Multi-Tenancy

Doris multi-tenancy is built on top of the Workload Group Resource soft limit of the kernel. Workloads are managed by group to ensure flexible allocation and control of memory and CPU resources. Each tenant role of a user can have many workload groups. You can manage CPU, memory, concurrency, and queues with this model shown in Figure 1.

NOTE:

(This feature is available for MRS 3.3.1 or later only.)
Multiple tenants can be created and managed only on FusionInsight Manager when Kerberos authentication is enabled for the cluster (the cluster is in security mode).
In MRS 3.5.0 and later versions: The default memory usage of the normal tenant is 90%, and the number of concurrent requests is 10,000.

Figure 1 Doris tenant model
Click to enlarge

After a user is associated with a tenant, all query tasks submitted by the user are added to the workload group. You can limit the percentage of CPU and memory resources on BE nodes for a single query and configure a soft memory limit for the workload group.

When cluster resources are insufficient, the system automatically stops several query tasks that occupy memory the most in the group. If the resources used by a workload group exceed the preset limit, multiple workloads share idle resources in the cluster and use resources more than the limit. This ensures stable execution of query tasks. In addition, the query queuing function is introduced to relieve system pressure in heavy-load scenarios. When creating a workload group, you can set the maximum number of concurrent queries to queue the exceeding queries.

Associations Between Doris Tenant Roles and Users

On the FusionInsight Manager service configuration and tenant management pages, you can create tenants, associate tenants with services, configure tenant resources, and associate tenants with users. The following figure shows the association between roles and users on the Manager and Doris sides.

Click to enlarge

A user, a tenant role, and a workload group are in one-on-one relationship. A user can have multiple permission roles.

Table 1 lists the Doris tenant resource configurations supported by the current version.

**Table 1** tenant resource configuration
Configuration	Value Range	Description	Remarks
CPU Quota Usage	1 to 100	Weight of CPU resources that can be used by a tenant	The value specifies a relative ratio that is valid during resource competition. For example, if this value for tenant A is 10 and that for tenant B is 20, the CPU resources can be used by the query tasks of tenant A is one-third of total resources, that is, 10/(10 + 20). If tenant C starts query tasks and its CPU quota usage is 30, a CPU quota of tenant A is one-sixth of total resources, that is, 10/(10 + 20 + 30).
Memory Quota	MRS 3.3.1: 1%~70% MRS 3.5.0 and later versions: 1%~90%	Maximum proportion of memory that can be used by a tenant	The available memory of a tenant is calculated as follows: Physical memory x mem_limit x Memory quota. The upper limit of the memory quota is 70% (90% for MRS 3.5.0 and later versions). The default Doris tenant normal occupies 70% (90% for MRS 3.5.0 and later versions) of memory.
Concurrency	MRS 3.3.1: 1~2147483647 MRS 3.5.0 and later versions: 1~10000	Maximum number of concurrent query tasks that a tenant can run	This parameter specifies the maximum number of tasks on each FE node. For example, if the number of concurrent SQL statements is set to 1 and the Doris has three FE nodes, the maximum number of SQL statements that can be executed in a cluster is 3.
Queue Length	MRS 3.3.1: 0~2147483647 MRS 3.5.0 and later versions: 0~1000	Maximum number of waiting query tasks	Excessive SQL statements are queued. When the queue is full, newly submitted queries are rejected.
Waiting Duration	MRS 3.3.1: 0~2147483647 MRS 3.5.0 and later versions: 0~86400000	Maximum waiting duration of a tenant query task	If the query waiting duration exceeds the value of this parameter, the query is rejected. The unit is millisecond.
Soft Memory Limit	On Off	Whether a tenant can use more memory resources than the limit	If this function is disabled, the system immediately cancels the tasks that occupy the most memory in the tenant groups when detecting that the memory usage of the tenant exceeds the upper limit. If this function is enabled and the cluster has idle memory resources, the tenant can use the system memory more than the limit. The tasks that occupy the most memory in the tenant groups are canceled only when cluster resources are insufficient.