Query queues

This topic describes how to manage query queues in StarRocks.

From v2.5, StarRocks supports query queues. With query queues enabled, StarRocks automatically queues the incoming queries when the concurrency threshold or resource limit is reached, thereby avoiding the overload deteriorating. Pending queries wait in a queue until there is enough compute resources available to begin execution.

The Query Queue feature has two versions:

• Query Queue v1: Triggers queuing based on query concurrency, BE memory usage, and BE CPU usage. The original query queue configurations and behaviors in this topic belong to v1. From v3.1.4 onwards, v1 supports setting query queues on the resource group level.
• Query Queue v2: Supported from v3.3 onwards. v2 estimates the BE resources consumed by each query, represents BE resources as logical slots, and queues and schedules queries based on the number of slots each query needs.

Query Queue v1

Query Queue v1 supports setting thresholds on CPU usage, memory usage, and query concurrency to trigger query queues.

Roadmap:

Version	Global query queue	Resource group-level query queue	Collective concurrency management	Dynamic concurrency adjustment
v2.5	✅	❌	❌	❌
v3.1.4	✅	✅	✅	✅

Enable Query Queue v1

Query queues are disabled by default. You can enable global or resource group-level query queues for INSERT loading, SELECT queries, and statistics queries by setting corresponding global session variables.

Enable global query queues

• Enable query queues for loading tasks:

SET GLOBAL enable_query_queue_load = true;

• Enable query queues for SELECT queries:

SET GLOBAL enable_query_queue_select = true;

• Enable query queues for statistics queries:

SET GLOBAL enable_query_queue_statistic = true;

Enable resource group-level query queues

From v3.1.4 onwards, StarRocks supports setting query queues on the resource group level.

To enable the resource group-level query queues, you also need to set enable_group_level_query_queue in addition to the global session variables mentioned above.

SET GLOBAL enable_group_level_query_queue = true;

Specify resource thresholds

Specify resource thresholds for global query queues

You can set the thresholds that trigger query queues via the following global session variables:

Variable	Default	Description
query_queue_concurrency_limit	0	The upper limit of concurrent queries on a BE. It takes effect only after being set greater than 0. Setting it to 0 indicates no limit is imposed.
query_queue_mem_used_pct_limit	0	The upper limit of memory usage percentage on a BE. It takes effect only after being set greater than 0. Setting it to 0 indicates no limit is imposed. Range: [0, 1]
query_queue_cpu_used_permille_limit	0	The upper limit of CPU usage permille (CPU usage * 1000) on a BE. It takes effect only after being set greater than 0. Setting it to 0 indicates no limit is imposed. Range: [0, 1000]

note

• After Query Queue v2 is enabled, query_queue_mem_used_pct_limit and query_queue_cpu_used_permille_limit are no longer supported for queue triggering.
• By default, BE reports resource usage to FE at one-second intervals. You can change this interval by setting the BE configuration item report_resource_usage_interval_ms.

Specify resource thresholds for resource group-level query queues

From v3.1.4 onwards, you can set individual concurrency limits (concurrency_limit) and CPU core limits (max_cpu_cores) when creating a resource group. When a query is initiated, if any of the resource consumptions exceed the resource threshold at either the global or resource group level, the query will be placed in queue until all resource consumptions are within the threshold.

Variable	Default	Description
concurrency_limit	0	The concurrency limit for the resource group on a single BE node. It takes effect only when it is set to greater than 0.
max_cpu_cores	0	The CPU core limit for this resource group on a single BE node. It takes effect only when it is set to greater than 0. Range: [0, avg_be_cpu_cores], where avg_be_cpu_cores represents the average number of CPU cores across all BE nodes.

You can use SHOW USAGE RESOURCE GROUPS to view the resource usage information for each resource group on each BE node, as described in View Resource Group Usage Information.

Manage query concurrency

When the number of running queries (num_running_queries) exceeds the global or resource group's concurrency_limit, incoming queries are placed in the queue. The way to obtain num_running_queries differs between versions < v3.1.4 and ≥ v3.1.4.

• In versions < v3.1.4, num_running_queries is reported by BEs at the interval specified in report_resource_usage_interval_ms. Therefore, there might be some delay in the identification of changes in num_running_queries. For example, if the num_running_queries reported by BEs at the moment does not exceed the global or resource group's concurrency_limit, but incoming queries arrive and exceed the concurrency_limit before the next report, these incoming queries will be executed without waiting in the queue.

• In versions ≥ v3.1.4, all running queries are collectively managed by the Leader FE. Each Follower FE notifies the Leader FE when initiating or finishing a query, allowing the StarRocks to handle scenarios where there is a sudden increase in queries exceeding the concurrency_limit.

Configure Query Queue v1

You can set the capacity of a query queue and the maximum timeout of queries in queues via the following global session variables:

Variable	Default	Description
query_queue_max_queued_queries	1024	The upper limit of queries in a queue. When this threshold is reached, incoming queries are rejected. It takes effect only after being set greater than 0.
query_queue_pending_timeout_second	300	The maximum timeout of a pending query in a queue. When this threshold is reached, the corresponding query is rejected. Unit: second.

Configure dynamic adjustment of query concurrency

Starting from version v3.1.4, for queries managed by the query queue and run by the Pipeline Engine, StarRocks can dynamically adjust the query concurrency pipeline_dop for incoming queries based on the current number of running queries num_running_queries, the number of fragments num_fragments, and the query concurrency pipeline_dop. This allows you to dynamically control query concurrency while minimizing scheduling overhead, ensuring optimal BE resource utilization. For more information about fragments and query concurrency pipeline_dop, see Query Management - Adjusting Query Concurrency.

For each query under a query queue, StarRocks maintains a concept of drivers, which represent the concurrent fragments of a query on a single BE. Its logical value num_drivers, which represents the total concurrency of all fragments of that query on a single BE, is equal to num_fragments * pipeline_dop. When a new query arrives, StarRocks adjusts the query concurrency pipeline_dop based on the following rules:

• The more the number of running drivers num_drivers exceeds the low water limit of concurrent drivers query_queue_driver_low_water, the lower the query concurrency pipeline_dop is adjusted to.
• StarRocks restrains the number of running drivers num_drivers below the high water limit of concurrent drivers for queries query_queue_driver_high_water.

You can configure the dynamic adjustment of query concurrency pipeline_dop using the following global session variables:

Variable	Default	Description
query_queue_driver_high_water	-1	The high water limit of concurrent drivers for a query. It takes effect only when it is set to a non-negative value. When set to 0, it is equivalent to avg_be_cpu_cores * 16, where avg_be_cpu_cores represents the average number of CPU cores across all BE nodes. When set to a value greater than 0, that value is used directly.
query_queue_driver_low_water	-1	The lower limit of concurrent drivers for queries. It takes effect only when it is set to a non-negative value. When set to 0, it is equivalent to avg_be_cpu_cores * 8. When set to a value greater than 0, that value is used directly.

Query Queue v2

From v3.3 onwards, StarRocks supports Query Queue v2. In Query Queue v2, query queues are no longer triggered based on fixed thresholds for query concurrency, BE memory usage, or BE CPU usage. Instead, it estimates the BE resources required by each query and queues and schedules queries based on logical slots. If there are not enough available slots, the query waits in the queue until enough slots are released.

Configure Query Queue v2

Query Queue v2 is enabled and tuned through FE configuration items. Changes to enable_query_queue_v2 require restarting FE nodes to take effect.

Configuration item	Default	Description
enable_query_queue_v2	false (v3.3 to v4.0) true (from v4.1 onwards)	Whether to enable Query Queue v2. When this item is set to true, StarRocks uses the v2 slot-based query scheduling mechanism.
query_queue_v2_concurrency_level	4	The logical concurrency level used by Query Queue v2 to calculate the total number of cluster slots. A larger value allows the system to admit more queries. This is a relative tuning parameter.
query_queue_slots_estimator_strategy	PBE	The slot estimation strategy used for queue-based queries. Valid values: PBE (parallelism-based, the default), MBE (memory-cost-based), and CBE (CPU-cost-based). PBE estimates a query's slots from scan parallelism, capped by the worker count: for OLAP tables it uses the number of scan ranges left after pruning, so only very small queries fall below the worker count; a connector/external scan is treated as a full-parallelism scan (the worker count) rather than a single-slot query. MBE estimates slots from the query's memory cost divided by query_queue_v2_mem_bytes_per_slot. CBE estimates slots from the plan CPU cost divided by query_queue_v2_cpu_costs_per_slot. MBE and CBE per-query slots are additionally capped by number_of_workers * max(1, pipeline_dop / 2). The legacy values MAX and MIN are still accepted for forward compatibility and are treated as the default estimator; any other value is rejected by configuration validation.
query_queue_v2_schedule_strategy	SWRR	The scheduling policy used by Query Queue V2 to order pending queries. Supported values (case-insensitive) are SWRR (Smooth Weighted Round Robin) — the default, suitable for mixed/hybrid workloads that need fair weighted sharing — and SJF (Short Job First + Aging) — prioritizes short jobs while using aging to avoid starvation. The value is parsed with case-insensitive enum lookup; an unrecognized value is logged as an error and the default policy is used. This configuration only affects behavior when Query Queue V2 is enabled and interacts with V2 sizing settings such as query_queue_v2_concurrency_level.
query_queue_v2_mem_bytes_per_slot	0	Per-slot memory target used by the memory-cost-based estimator (MBE). When query_queue_slots_estimator_strategy is MBE, the total slots are derived from the warehouse memory budget, and a query's slots are estimated from its total memory cost divided by this value, capped by number_of_workers * max(1, pipeline_dop / 2). If it is non-positive, Query Queue V2 uses the average worker memory per core.
query_queue_v2_cpu_costs_per_slot	1000000000	Per-slot CPU cost threshold used by the CPU-cost-based estimator (CBE) to estimate how many slots a query needs from its plan CPU cost. The scheduler computes slots as ceil(plan_cpu_costs / query_queue_v2_cpu_costs_per_slot) and clamps the result to the range [1, min(totalSlots, number_of_workers * max(1, pipeline_dop / 2))]. A non-positive value is normalized to 1. Increasing this value reduces slots allocated per query (favoring fewer, larger-slot queries); decreasing it increases slots per query.
query_queue_concurrency_limit	0	The upper limit of concurrent queries on a BE. It takes effect only after being set greater than 0. Setting it to 0 indicates no limit is imposed.

note

query_queue_mem_used_pct_limit and query_queue_cpu_used_permille_limit apply only to Query Queue v1. After Query Queue v2 is enabled, these parameters no longer take effect.

Resource slots

Query Queue v2 represents BE resources as logical slots:

• Total cluster slots: StarRocks sets a logical total number of slots for the entire cluster. This total is positively correlated with the number of BEs and BE CPU cores, and is also affected by query_queue_v2_concurrency_level.
• Slots required by a query: StarRocks estimates the number of slots required by each query. The estimation is based on factors such as statistics, query complexity, the number of fragments, estimated input and output data volumes of complex operators, and DOP.

Queuing logic

When the number of slots required by a query exceeds the current number of remaining slots, the query waits in the queue. Query Queue v2 preferentially satisfies queries that require fewer slots, allowing small queries to obtain resources first and avoiding head-of-line blocking where a large query at the head of the queue blocks later small queries.

The entire queuing logic is completed on FE, including setting the total number of cluster slots, estimating the number of slots required by a query, and deciding which query's slot requirement to satisfy first. Query Queue v2 does not schedule based on the actual resource usage of BEs.

Choose an estimation strategy

PBE

Parallel-based estimation (PBE) strategy is best for:

• Normal reporting queries
• Mixed point lookups and large queries
• Users who do not want to understand cost model details
• DBAs who want stable, simple, and explainable queueing behavior first

Expected behaviors with PBE include:

• Point lookups or queries that scan little data after pruning use fewer slots
• Queries that scan larger ranges use more slots
• Small queries are more likely to get execution resources during peak hours

The following example sets PBE as the strategy:

ADMIN SET FRONTEND CONFIG ("query_queue_slots_estimator_strategy" = "PBE");

MBE

Memory cost-based estimation (MBE) strategy is suitable for dealing with memory pressure, such as large joins, large aggregations, or high-cardinality aggregations.

The following example sets MBE as the strategy, and allocates 2 GB of memory to each slot:

ADMIN SET FRONTEND CONFIG ("query_queue_slots_estimator_strategy" = "MBE");
ADMIN SET FRONTEND CONFIG ("query_queue_v2_mem_bytes_per_slot" = "2147483648");

MBE divides the query’s total memory cost by this value to get query slots, and divides the warehouse memory budget by this value to get total slots.

Tune the MBE strategy in the following directions:

Symptom: Memory still gets saturated easily

• Adjustment: Decrease query_queue_v2_concurrency_level
• Effect: Directly lowers the MBE total memory budget

Symptom: Queue is too long but BE memory still has room

• Adjustment: Increase query_queue_v2_concurrency_level
• Effect: Directly raises the MBE total memory budget

Symptom: max_slots is very small and integer rounding is visible

• Adjustment: Decrease query_queue_v2_mem_bytes_per_slot
• Effect: Uses a finer memory slot granularity and reduces coarse rounding error

CBE

CPU cost-based estimation (CBE) strategy is suitable for dealing with memory pressure, such as compute-heavy SQL, complex expressions, or heavy CPU work after scanning.

The following example sets CBE as the strategy, and set the CPU cost threshold to 1000000000 for each slot:

ADMIN SET FRONTEND CONFIG ("query_queue_slots_estimator_strategy" = "CBE");
ADMIN SET FRONTEND CONFIG ("query_queue_v2_cpu_costs_per_slot" = "1000000000");

Symptom: CPU is often saturated

• Adjustment: Decrease query_queue_v2_cpu_costs_per_slot
• Effect: The same CPU cost for more slots, making concurrency more conservative

Symptom: Queries queue noticeably but CPU still has room

• Adjustment: Increase query_queue_v2_cpu_costs_per_slot
• Effect: The same CPU cost for fewer slots, making concurrency looser

Tune concurrency capacity

If you only want to increase or decrease overall concurrency, do not switch between PBE, MBE, and CBE first. Tune total slot capacity first:

ADMIN SET FRONTEND CONFIG ("query_queue_v2_concurrency_level" = "<value>");

Recommended process:

1. Start with default value 4.
2. Observe remain_slots, max_slots, query_pending_length, CPU, memory, and query latency.
3. If there is resource headroom but queries queue noticeably, gradually increase query_queue_v2_concurrency_level.
4. If resources are often saturated or queries interfere with each other heavily, gradually decrease query_queue_v2_concurrency_level.
5. Make small changes each time, such as 10% to 25%, and observe one business peak period before making another change.

Tuning Priority: Use query_queue_v2_concurrency_level to tune overall capacity first. Consider switching to MBE or PBE only after that. Do not change multiple parameters at the same time at the beginning, because it becomes hard to tell which parameter caused the effect.

Fallback Concurrency Cap

query_queue_concurrency_limit is a fallback concurrency cap and applies to PBE, MBE, and CBE. Query Queue V2 first estimates the slots required by a query with the current estimator and checks whether total slots are available. After that, it checks whether the current number of running queries has reached query_queue_concurrency_limit.

The default value 0 means unlimited. Set it only when you need an absolute cap on the number of concurrently running queries:

ALTER WAREHOUSE default_warehouse SET ("query_queue_concurrency_limit" = "8");

Use query_queue_v2_concurrency_level to tune resource capacity first. Use query_queue_concurrency_limit only when you need to explicitly limit the number of queries run at the same time.

Monitor query queues

You can view information related to query queues using the following methods.

SHOW PROC

You can check the number of running queries, and memory and CPU usages in BE nodes using SHOW PROC:

mysql> SHOW PROC '/backends'\G
*************************** 1. row ***************************
...
    NumRunningQueries: 0
           MemUsedPct: 0.79 %
           CpuUsedPct: 0.0 %

SHOW PROCESSLIST

You can check if a query is in a queue (when IsPending is true) using SHOW PROCESSLIST:

mysql> SHOW PROCESSLIST;
+------+------+---------------------+-------+---------+---------------------+------+-------+-------------------+-----------+
| Id   | User | Host                | Db    | Command | ConnectionStartTime | Time | State | Info              | IsPending |
+------+------+---------------------+-------+---------+---------------------+------+-------+-------------------+-----------+
|    2 | root | xxx.xx.xxx.xx:xxxxx |       | Query   | 2022-11-24 18:08:29 |    0 | OK    | SHOW PROCESSLIST  | false     |
+------+------+---------------------+-------+---------+---------------------+------+-------+-------------------+-----------+

FE audit log

You can check the FE audit log file fe.audit.log. The field PendingTimeMs indicates the time a query spent waiting in a queue, and its unit is milliseconds.

Monitoring metrics

You can obtain metrics of query queues in StarRocks using the Monitor and Alert feature. The following FE metrics are derived from the statistical data of each FE node.

Metric	Unit	Type	Description
starrocks_fe_query_queue_pending	Count	Instantaneous	The current number of queries in the queue.
starrocks_fe_query_queue_total	Count	Instantaneous	The total number of queries historically queued (including those currently running).
starrocks_fe_query_queue_timeout	Count	Instantaneous	The total number of queries that have timed out while in the queue.
starrocks_fe_resource_group_query_queue_total	Count	Instantaneous	The total number of queries historically queued in this resource group (including those currently running). The name label indicates the name of the resource group. This metric is supported from v3.1.4 onwards.
starrocks_fe_resource_group_query_queue_pending	Count	Instantaneous	The number of queries currently in the queue for this resource group. The name label indicates the name of the resource group. This metric is supported from v3.1.4 onwards.
starrocks_fe_resource_group_query_queue_timeout	Count	Instantaneous	The number of queries that have timed out while in the queue for this resource group. The name label indicates the name of the resource group. This metric is supported from v3.1.4 onwards.

SHOW RUNNING QUERIES

From v3.1.4 onwards, StarRocks supports the SQL statement SHOW RUNNING QUERIES, which is used to display queue information for each query. The meanings of each field are as follows:

• QueryId: The ID of the query.
• ResourceGroupId: The ID of the resource group that the query hit. When there is no hit on a user-defined resource group, it will be displayed as "-".
• StartTime: The start time of the query.
• PendingTimeout: The time when the PENDING query will time out in the queue.
• QueryTimeout: The time when the query times out.
• State: The queue state of the query, where "PENDING" indicates it is in the queue, and "RUNNING" indicates it is currently executing.
• Slots: The logical resource quantity requested by the query. In Query Queue v1, this value is usually 1. In Query Queue v2, this value is the estimated number of slots for the query.
• Frontend: The FE node that initiated the query.
• FeStartTime: The start time of the FE node that initiated the query.

Example:

MySQL [(none)]> SHOW RUNNING QUERIES;
+--------------------------------------+-----------------+---------------------+---------------------+---------------------+-----------+-------+---------------------------------+---------------------+
| QueryId                              | ResourceGroupId | StartTime           | PendingTimeout      | QueryTimeout        |   State   | Slots | Frontend                        | FeStartTime         |
+--------------------------------------+-----------------+---------------------+---------------------+---------------------+-----------+-------+---------------------------------+---------------------+
| a46f68c6-3b49-11ee-8b43-00163e10863a | -               | 2023-08-15 16:56:37 | 2023-08-15 17:01:37 | 2023-08-15 17:01:37 |  RUNNING  | 1     | 127.00.00.01_9010_1692069711535 | 2023-08-15 16:37:03 |
| a6935989-3b49-11ee-935a-00163e13bca3 | 12003           | 2023-08-15 16:56:40 | 2023-08-15 17:01:40 | 2023-08-15 17:01:40 |  RUNNING  | 1     | 127.00.00.02_9010_1692069658426 | 2023-08-15 16:37:03 |
| a7b5e137-3b49-11ee-8b43-00163e10863a | 12003           | 2023-08-15 16:56:42 | 2023-08-15 17:01:42 | 2023-08-15 17:01:42 |  PENDING  | 1     | 127.00.00.03_9010_1692069711535 | 2023-08-15 16:37:03 |
+--------------------------------------+-----------------+---------------------+---------------------+---------------------+-----------+-------+---------------------------------+---------------------+