Resource management in Azure SQL Database

APPLIES TO: Azure SQL Database

This article provides an overview of resource management in Azure SQL Database. It provides information on what happens when resource limits are reached, and describes resource governance mechanisms that are used to enforce these limits.

For specific resource limits per pricing tier (also known as service objective) for single databases, refer to either DTU-based single database resource limits or vCore-based single database resource limits. For elastic pool resource limits, refer to either DTU-based elastic pool resource limits or vCore-based elastic pool resource limits.

Tip

For Azure SQL Managed Instance limits, see resource limits for managed instances.

For Azure Synapse Analytics dedicated SQL pool limits, see capacity limits and memory and concurrency limits.

Logical server limits

Resource Limit
Databases per logical server 5000
Default number of logical servers per subscription in a region 20
Max number of logical servers per subscription in a region 250
DTU / eDTU quota per logical server 54,000
vCore quota per logical server 540
Max elastic pools per logical server Limited by number of DTUs or vCores. For example, if each pool is 1000 DTUs, then a server can support 54 pools.

Important

As the number of databases approaches the limit per logical server, the following can occur:

  • Increasing latency in running queries against the master database. This includes views of resource utilization statistics such as sys.resource_stats.
  • Increasing latency in management operations and rendering portal viewpoints that involve enumerating databases in the server.

Note

To obtain more DTU/eDTU quota, vCore quota, or more logical servers than the default number, submit a new support request in the Azure portal. For more information, see Request quota increases for Azure SQL Database.

What happens when resource limits are reached

Compute CPU

When database compute CPU utilization becomes high, query latency increases, and queries can even time out. Under these conditions, queries may be queued by the service and are provided resources for execution as resources become free. When encountering high compute utilization, mitigation options include:

Storage

When database space used reaches the maximum data size limit, database inserts and updates that increase data size fail and clients receive an error message. SELECT and DELETE statements remain unaffected.

In Premium and Business Critical service tiers, clients also receive an error message if combined storage consumption by data, transaction log, and tempdb exceeds maximum local storage size. For more information, see Storage space governance.

When encountering high space utilization, mitigation options include:

  • Increase maximum data size of the database or elastic pool, or scale up to a service objective with a higher maximum data size limit. See Scale single database resources and Scale elastic pool resources.
  • If the database is in an elastic pool, then alternatively the database can be moved outside of the pool so that its storage space isn't shared with other databases.
  • Shrink a database to reclaim unused space. In elastic pools, shrinking a database provides more storage for other databases in the pool. For more information, see Manage file space in Azure SQL Database.
  • Check if high space utilization is due to a spike in the size of Persistent Version Store (PVS). PVS is a part of each database, and is used to implement Accelerated Database Recovery. To determine current PVS size, see PVS troubleshooting. A common reason for large PVS size is a transaction that is open for a long time (hours), preventing cleanup of older versions in PVS.
  • For large databases in Premium and Business Critical service tiers, you may receive an out-of-space error even though used space in the database is below its maximum data size limit. This may happen if tempdb or transaction log consume a large amount of storage toward the maximum local storage limit. Fail over the database or elastic pool to reset tempdb to its initial smaller size, or shrink transaction log to reduce local storage consumption.

Sessions and workers (requests)

The maximum numbers of sessions and workers are determined by the service tier and compute size. New requests are rejected when session or worker limits are reached, and clients receive an error message. While the number of connections available can be controlled by the application, the number of concurrent workers is often harder to estimate and control. This is especially true during peak load periods when database resource limits are reached and workers pile up due to longer running queries, large blocking chains, or excessive query parallelism.

When encountering high session or worker utilization, mitigation options include:

Memory

Unlike other resources (CPU, workers, storage), reaching the memory limit does not negatively impact query performance, and does not cause errors and failures. As described in detail in Memory Management Architecture Guide, the database engine often uses all available memory, by design. Memory is used primarily for caching data, to avoid slower storage access. Thus, higher memory utilization usually improves query performance due to faster reads from memory, rather than slower reads from storage.

After database engine startup, as the workload starts reading data from storage, the database engine aggressively caches data in memory. After this initial ramp-up period, it is common and expected to see the avg_memory_usage_percent and avg_instance_memory_percent columns in sys.dm_db_resource_stats to be close or equal to 100%, particularly for databases that are not idle, and do not fully fit in memory.

Besides the data cache, memory is used in other components of the database engine. When there is demand for memory and all available memory has been used by the data cache, the database engine will dynamically reduce data cache size to make memory available to other components, and will dynamically grow data cache when other components release memory.

In rare cases, a sufficiently demanding workload may cause an insufficient memory condition, leading to out-of-memory errors. This can happen at any level of memory utilization between 0% and 100%. This is more likely to occur on smaller compute sizes that have proportionally smaller memory limits, and/or with workloads using more memory for query processing, such as in dense elastic pools.

When encountering out-of-memory errors, mitigation options include:

Solution Description
Reduce the size of memory grants For more information about memory grants, see the Understanding SQL Server memory grant blog post. A common solution for avoiding excessively large memory grants is keeping statistics up to date. This results in more accurate estimates of memory consumption by the query engine, avoiding unnecessarily large memory grants.

In databases using compatibility level 140 and above, the database engine may automatically adjust memory grant size using Batch mode memory grant feedback. In databases using compatibility level 150 and above, the database engine similarly uses Row mode memory grant feedback, for more common row mode queries. This built-in functionality helps avoid out-of-memory errors due to unnecessarily large memory grants.
Reduce the size of query plan cache The database engine caches query plans in memory, to avoid compiling a query plan for every query execution. To avoid query plan cache bloat caused by caching plans that are only used once, enable the OPTIMIZE_FOR_AD_HOC_WORKLOADS database-scoped configuration.
Reduce the size of lock memory The database engine uses memory for locks. When possible, avoid large transactions that may acquire a large number of locks and cause high lock memory consumption.

Resource consumption by user workloads and internal processes

Azure SQL Database requires compute resources to implement core service features such as high availability and disaster recovery, database backup and restore, monitoring, Query Store, Automatic tuning, etc. The system sets aside a certain limited portion of the overall resources for these internal processes using resource governance mechanisms, making the remainder of resources available for user workloads. At times when internal processes aren't using compute resources, the system makes them available to user workloads.

Total CPU and memory consumption by user workloads and internal processes is reported in the sys.dm_db_resource_stats and sys.resource_stats views, in avg_instance_cpu_percent and avg_instance_memory_percent columns. This data is also reported via the sqlserver_process_core_percent and sqlserver_process_memory_percent Azure Monitor metrics, for single databases and elastic pools at the pool level.

CPU and memory consumption by user workloads in each database is reported in the sys.dm_db_resource_stats and sys.resource_stats views, in avg_cpu_percent and avg_memory_usage_percent columns. For elastic pools, pool-level resource consumption is reported in the sys.elastic_pool_resource_stats view. User workload CPU consumption is also reported via the cpu_percent Azure Monitor metric, for single databases and elastic pools at the pool level.

A more detailed breakdown of recent resource consumption by user workloads and internal processes is reported in the sys.dm_resource_governor_resource_pools_history_ex and sys.dm_resource_governor_workload_groups_history_ex views. For details on resource pools and workload groups referenced in these views, see Resource governance. These views report on resource utilization by user workloads and specific internal processes in the associated resource pools and workload groups.

In the context of performance monitoring and troubleshooting, it's important to consider both user CPU consumption (avg_cpu_percent, cpu_percent), and total CPU consumption by user workloads and internal processes (avg_instance_cpu_percent,sqlserver_process_core_percent).

User CPU consumption is calculated as a percentage of the user workload limits in each service objective. User CPU utilization at 100% indicates that the user workload has reached the limit of the service objective. However, when total CPU consumption reaches the 70-100% range, it's possible to see user workload throughput flattening out and query latency increasing, even if reported user CPU consumption remains significantly below 100%. This is more likely to occur when using smaller service objectives with a moderate allocation of compute resources, but relatively intense user workloads, such as in dense elastic pools. This can also occur with smaller service objectives when internal processes temporarily require additional resources, for example when creating a new replica of the database, or backing up the database.

When total CPU consumption is high, mitigation options are the same as noted in the Compute CPU section, and include service objective increase and/or user workload optimization.

Resource governance

To enforce resource limits, Azure SQL Database uses a resource governance implementation that is based on SQL Server Resource Governor, modified and extended to run in the cloud. In SQL Database, multiple resource pools and workload groups, with resource limits set at both pool and group levels, provide a balanced Database-as-a-Service. User workload and internal workloads are classified into separate resource pools and workload groups. User workload on the primary and readable secondary replicas, including geo-replicas, is classified into the SloSharedPool1 resource pool and UserPrimaryGroup.DBId[N] workload groups, where [N] stands for the database ID value. In addition, there are multiple resource pools and workload groups for various internal workloads.

In addition to using Resource Governor to govern resources within the database engine, Azure SQL Database also uses Windows Job Objects for process level resource governance, and Windows File Server Resource Manager (FSRM) for storage quota management.

Azure SQL Database resource governance is hierarchical in nature. From top to bottom, limits are enforced at the OS level and at the storage volume level using operating system resource governance mechanisms and Resource Governor, then at the resource pool level using Resource Governor, and then at the workload group level using Resource Governor. Resource governance limits in effect for the current database or elastic pool are reported in the sys.dm_user_db_resource_governance view.

Data IO governance

Data IO governance is a process in Azure SQL Database used to limit both read and write physical IO against data files of a database. IOPS limits are set for each service level to minimize the "noisy neighbor" effect, to provide resource allocation fairness in a multi-tenant service, and to stay within the capabilities of the underlying hardware and storage.

For single databases, workload group limits are applied to all storage IO against the database. For elastic pools, workload group limits apply to each database in the pool. Additionally, the resource pool limit additionally applies to the cumulative IO of the elastic pool. Tempdb IO is subject to workload group limits, with the exception of Basic, Standard, and General Purpose service tier, where higher tempdb IO limits apply. In general, resource pool limits may not be achievable by the workload against a database (either single or pooled), because workload group limits are lower than resource pool limits and limit IOPS/throughput sooner. However, pool limits may be reached by the combined workload against multiple databases in the same pool.

For example, if a query generates 1000 IOPS without any IO resource governance, but the workload group maximum IOPS limit is set to 900 IOPS, the query won't be able to generate more than 900 IOPS. However, if the resource pool maximum IOPS limit is set to 1500 IOPS, and the total IO from all workload groups associated with the resource pool exceeds 1500 IOPS, then the IO of the same query may be reduced below the workgroup limit of 900 IOPS.

The IOPS and throughput max values returned by the sys.dm_user_db_resource_governance view act as limits/caps, not as guarantees. Further, resource governance doesn't guarantee any specific storage latency. The best achievable latency, IOPS, and throughput for a given user workload depend not only on IO resource governance limits, but also on the mix of IO sizes used, and on the capabilities of the underlying storage. SQL Database uses IOs that vary in size between 512 KB and 4 MB. For the purposes of enforcing IOPS limits, every IO is accounted regardless of its size, with the exception of databases with data files in Azure Storage. In that case, IOs larger than 256 KB are accounted as multiple 256-KB IOs, to align with Azure Storage IO accounting.

For Basic, Standard, and General Purpose databases, which use data files in Azure Storage, the primary_group_max_io value may not be achievable if a database doesn't have enough data files to cumulatively provide this number of IOPS, or if data isn't distributed evenly across files, or if the performance tier of underlying blobs limits IOPS/throughput below the resource governance limits. Similarly, with small log IOs generated by frequent transaction commits, the primary_max_log_rate value may not be achievable by a workload due to the IOPS limit on the underlying Azure Storage blob. For databases using Azure Premium Storage, Azure SQL Database uses sufficiently large storage blobs to obtain needed IOPS/throughput, regardless of database size. For larger databases, multiple data files are created to increase total IOPS/throughput capacity.

Resource utilization values such as avg_data_io_percent and avg_log_write_percent, reported in the sys.dm_db_resource_stats, sys.resource_stats, and sys.elastic_pool_resource_stats views, are calculated as percentages of maximum resource governance limits. Therefore, when factors other than resource governance limit IOPS/throughput, it's possible to see IOPS/throughput flattening out and latencies increasing as the workload increases, even though reported resource utilization remains below 100%.

To determine read and write IOPS, throughput, and latency per database file, use the sys.dm_io_virtual_file_stats() function. This function surfaces all IO against the database, including background IO that isn't accounted towards avg_data_io_percent, but uses IOPS and throughput of the underlying storage, and can impact observed storage latency. The function reports additional latency that may be introduced by IO resource governance for reads and writes, in the io_stall_queued_read_ms and io_stall_queued_write_ms columns respectively.

Transaction log rate governance

Transaction log rate governance is a process in Azure SQL Database used to limit high ingestion rates for workloads such as bulk insert, SELECT INTO, and index builds. These limits are tracked and enforced at the subsecond level to the rate of log record generation, limiting throughput regardless of how many IOs may be issued against data files. Transaction log generation rates currently scale linearly up to a point that is hardware-dependent and service tier-dependent.

Log rates are set such that they can be achieved and sustained in a variety of scenarios, while the overall system can maintain its functionality with minimized impact to the user load. Log rate governance ensures that transaction log backups stay within published recoverability SLAs. This governance also prevents an excessive backlog on secondary replicas, that could otherwise lead to longer than expected downtime during failovers.

The actual physical IOs to transaction log files are not governed or limited. As log records are generated, each operation is evaluated and assessed for whether it should be delayed in order to maintain a maximum desired log rate (MB/s per second). The delays aren't added when the log records are flushed to storage, rather log rate governance is applied during log rate generation itself.

The actual log generation rates imposed at run time may also be influenced by feedback mechanisms, temporarily reducing the allowable log rates so the system can stabilize. Log file space management, avoiding running into out of log space conditions and data replication mechanisms can temporarily decrease the overall system limits.

Log rate governor traffic shaping is surfaced via the following wait types (exposed in the sys.dm_exec_requests and sys.dm_os_wait_stats views):

Wait Type Notes
LOG_RATE_GOVERNOR Database limiting
POOL_LOG_RATE_GOVERNOR Pool limiting
INSTANCE_LOG_RATE_GOVERNOR Instance level limiting
HADR_THROTTLE_LOG_RATE_SEND_RECV_QUEUE_SIZE Feedback control, availability group physical replication in Premium/Business Critical not keeping up
HADR_THROTTLE_LOG_RATE_LOG_SIZE Feedback control, limiting rates to avoid an out of log space condition
HADR_THROTTLE_LOG_RATE_MISMATCHED_SLO Geo-replication feedback control, limiting log rate to avoid high data latency and unavailability of geo-secondaries

When encountering a log rate limit that is hampering desired scalability, consider the following options:

  • Scale up to a higher service level in order to get the maximum log rate of a service tier, or switch to a different service tier. The Hyperscale service tier provides 100 MB/s log rate regardless of chosen service level.
  • If data being loaded is transient, such as staging data in an ETL process, it can be loaded into tempdb (which is minimally logged).
  • For analytic scenarios, load into a clustered columnstore table, or a table with indexes that use data compression. This reduces the required log rate. This technique does increase CPU utilization and is only applicable to data sets that benefit from clustered columnstore indexes or data compression.

Storage space governance

In Premium and Business Critical service tiers, customer data including data files, transaction log files, and tempdb files is stored on the local SSD storage of the machine hosting the database or elastic pool. Local SSD storage provides high IOPS and throughput, and low IO latency. In addition to customer data, local storage is used for the operating system, management software, monitoring data and logs, and other files necessary for system operation.

The size of local storage is finite and depends on hardware capabilities, which determine the maximum local storage limit, or local storage set aside for customer data. This limit is set to maximize customer data storage, while ensuring safe and reliable system operation. To find the maximum local storage value for each service objective, see resource limits documentation for single databases and elastic pools.

You can also find this value, and the amount of local storage currently used by a given database or elastic pool, using the following query:

SELECT server_name, database_name, slo_name, user_data_directory_space_quota_mb, user_data_directory_space_usage_mb
FROM sys.dm_user_db_resource_governance
WHERE database_id = DB_ID();
Column Description
server_name Logical server name
database_name Database name
slo_name Service objective name, including hardware generation
user_data_directory_space_quota_mb Maximum local storage, in MB
user_data_directory_space_usage_mb Current local storage consumption by data files, transaction log files, and tempdb files, in MB. Updated every five minutes.

This query should be executed in the user database, not in the master database. For elastic pools, the query can be executed in any database in the pool. Reported values apply to the entire pool.

Important

In Premium and Business Critical service tiers, if the workload attempts to increase combined local storage consumption by data files, transaction log files, and tempdb files over the maximum local storage limit, an out-of-space error will occur.

As databases are created, deleted, and increase or decrease in size, local storage consumption on a machine fluctuates over time. If the system detects that available local storage on a machine is low, and a database or an elastic pool is at risk of running out of space, it will move the database or elastic pool to a different machine with sufficient local storage available.

This move occurs in an online fashion, similarly to a database scaling operation, and has a similar impact, including a short (seconds) failover at the end of the operation. This failover terminates open connections and rolls back transactions, potentially impacting applications using the database at that time.

Because all data is copied to local storage volumes on different machines, moving larger databases may require a substantial amount of time. During that time, if local space consumption by the database or elastic pool, or by the tempdb database grows rapidly, the risk of running out of space increases. The system initiates database movement in a balanced fashion to minimize out-of-space errors while avoiding unnecessary failovers.

Note

Database movement due to insufficient local storage only occurs in the Premium or Business Critical service tiers. It does not occur in the Hyperscale, General Purpose, Standard, and Basic service tiers, because in those tiers data files are not stored in local storage.

Next steps