Serverless compute tier for Azure SQL Database

Applies to: Azure SQL Database

Serverless is a compute tier for single databases in Azure SQL Database that automatically scales compute based on workload demand and bills for the amount of compute used per second. The serverless compute tier also automatically pauses databases during inactive periods when only storage is billed and automatically resumes databases when activity returns. The serverless compute tier is available in the General Purpose service tier and the Hyperscale service tier.

Overview

A compute autoscaling range and an auto-pause delay are important parameters for the serverless compute tier. The configuration of these parameters shapes the database performance experience and compute cost.

Performance configuration

• The minimum vCores and maximum vCores are configurable parameters that define the range of compute capacity available for the database. Memory and IO limits are proportional to the vCore range specified.
• The auto-pause delay is a configurable parameter that defines the period of time the database must be inactive before it is automatically paused. The database is automatically resumed when the next sign in or other activity occurs. Alternatively, automatic pausing can be disabled.

Cost

• The cost for a serverless database is the summation of the compute cost and storage cost.
• When compute usage is between the minimum and maximum limits configured, the compute cost is based on vCore and memory used.
• When compute usage is below the minimum limits configured, the compute cost is based on the minimum vCores and minimum memory configured.
• When the database is paused, the compute cost is zero and only storage costs are incurred.
• The storage cost is determined in the same way as in the provisioned compute tier.

For more cost details, see Billing.

Scenarios

Serverless is price-performance optimized for single databases with intermittent, unpredictable usage patterns that can afford some delay in compute warm-up after idle usage periods. In contrast, the provisioned compute tier is price-performance optimized for single databases or multiple databases in elastic pools with higher average usage that cannot afford any delay in compute warm-up.

Scenarios well suited for serverless compute

• Single databases with intermittent, unpredictable usage patterns interspersed with periods of inactivity, and lower average compute utilization over time.
• Single databases in the provisioned compute tier that are frequently rescaled and customers who prefer to delegate compute rescaling to the service.
• New single databases without usage history where compute sizing is difficult or not possible to estimate before deployment in an Azure SQL Database.

Scenarios well suited for provisioned compute

• Single databases with more regular, predictable usage patterns and higher average compute utilization over time.
• Databases that cannot tolerate performance trade-offs resulting from more frequent memory trimming or delays in resuming from a paused state.
• Multiple databases with intermittent, unpredictable usage patterns that can be consolidated into elastic pools for better price-performance optimization.

Compare compute tiers

The following table summarizes distinctions between the serverless compute tier and the provisioned compute tier:

Purchasing model and service tier

The following table describes serverless support based on purchasing model, service tiers, and hardware:

Autoscaling

Scaling responsiveness

Serverless databases are run on a machine with sufficient capacity to satisfy resource demand without interruption for any amount of compute requested, within limits set by the maximum vCores value. Occasionally, load balancing automatically occurs if the machine is unable to satisfy resource demand within a few minutes. For example, if the resource demand is 4 vCores, but only 2 vCores are available, then it can take up to a few minutes to load balance before 4 vCores are provided. The database remains online during load balancing except for a brief period at the end of the operation when connections are dropped.

Memory management

In both the General Purpose and Hyperscale service tiers, memory for serverless databases is reclaimed more frequently than for provisioned compute databases. This behavior is important to control costs in serverless and can impact performance.

Cache reclamation

Unlike provisioned compute databases, memory from the SQL cache is reclaimed from a serverless database when CPU or active cache utilization is low.

• Active cache utilization is considered low when the total size of the most recently used cache entries falls below a threshold, for a period of time.
• When cache reclamation is triggered, the target cache size is reduced incrementally to a fraction of its previous size and reclaiming only continues if usage remains low.
• When cache reclamation occurs, the policy for selecting cache entries to evict is the same selection policy as for provisioned compute databases when memory pressure is high.
• The cache size is never reduced below the minimum memory limit as defined by minimum vCores.

In both serverless and provisioned compute databases, cache entries can be evicted if all available memory is used.

When CPU utilization is low, active cache utilization can remain high depending on the usage pattern and prevent memory reclamation. Also, there can be other delays after user activity stops before memory reclamation occurs due to periodic background processes responding to prior user activity. For example, delete operations and Query Store cleanup tasks generate ghost records that are marked for deletion, but are not physically deleted until the ghost cleanup process runs. Ghost cleanup might involve reading data pages into cache.

Cache hydration

The SQL memory cache grows as data is fetched from disk in the same way and with the same speed as for provisioned databases. When the database is busy, the cache is allowed to grow unconstrained while there is available memory.

Disk cache management

In the Hyperscale service tier for both serverless and provisioned compute tiers, each compute replica uses a Resilient Buffer Pool Extension (RBPEX) cache, which stores data pages on local SSD to improve IO performance. However, in the serverless compute tier for Hyperscale, the RBPEX cache for each compute replica automatically grows and shrinks in response to increasing and decreasing workload demand. The maximum size the RBPEX cache can grow to is three times the maximum memory configured for the database. For details on maximum memory and RBPEX auto-scaling limits in serverless, see serverless Hyperscale resource limits.

Auto-pause and auto-resume

Currently, serverless auto-pausing and auto-resuming are only supported in the General Purpose tier.

Auto-pause

Auto-pausing is triggered if all of the following conditions are true during the auto-pause delay:

• Number of sessions = 0
• CPU = 0 for user workload running in the user resource pool

An option is provided to disable auto-pausing if desired.

The following features do not support auto-pausing, but do support auto-scaling. If any of the following features are used, then auto-pausing must be disabled and the database remains online regardless of the duration of database inactivity:

• Geo-replication (active geo-replication and failover groups).
• Long-term backup retention (LTR).
• The sync database used in SQL Data Sync. Unlike sync databases, hub and member databases support auto-pausing.
• DNS alias created for the logical server containing a serverless database.
• Elastic Jobs, Auto-pause enabled serverless database is not supported as a Job Database. Serverless databases targeted by elastic jobs do support auto-pausing. Job connections will resume a database.

Auto-pausing is temporarily prevented during the deployment of some service updates, which require the database be online. In such cases, auto-pausing becomes allowed again once the service update completes.

Auto-pause troubleshooting

If auto-pausing is enabled and features that block auto-pausing are not used, but a database does not auto-pause after the delay period, then application or user sessions might be preventing auto-pausing.

To see if there are any application or user sessions currently connected to the database, connect to the database using any client tool, and execute the following query:

SELECT session_id,
       host_name,
       program_name,
       client_interface_name,
       login_name,
       status,
       login_time,
       last_request_start_time,
       last_request_end_time
FROM sys.dm_exec_sessions AS s
INNER JOIN sys.dm_resource_governor_workload_groups AS wg
ON s.group_id = wg.group_id
WHERE s.session_id <> @@SPID
      AND
      (
          (
          wg.name like 'UserPrimaryGroup.DB%'
          AND
          TRY_CAST(RIGHT(wg.name, LEN(wg.name) - LEN('UserPrimaryGroup.DB') - 2) AS int) = DB_ID()
          )
      OR
      wg.name = 'DACGroup'
      );

• If the result set is nonempty, it indicates that there are sessions currently preventing auto-pausing.
• If the result set is empty, it is still possible that sessions were open, possibly for a short time, at some point earlier during the auto-pause delay period. To check for activity during the delay period, you can use Auditing for Azure SQL Database and Azure Synapse Analytics and examine audit data for the relevant period.

Auto-resume

Auto-resuming is triggered if any of the following conditions are true at any time:

Monitoring, management, or other solutions performing any of these operations trigger auto-resuming. Auto-resuming is also triggered during the deployment of some service updates that require the database be online.

Connectivity

If a serverless database is paused, the first connection attempt resumes the database and returns an error stating that the database is unavailable with error code 40613. Once the database is resumed, the login can be retried to establish connectivity. Database clients following connection retry logic recommendations should not need to be modified. For connection retry logic options and recommendations, see:

• Connection retry logic in SqlClient
• Connection retry logic in SQL Database using Entity Framework Core
• Connection retry logic in SQL Database using Entity Framework 6
• Connection retry logic in SQL Database using ADO.NET

Latency

The latency to auto-resume and auto-pause a serverless database is generally order of 1 minute to auto-resume and 1-10 minutes after the expiration of the delay period to auto-pause.

Customer managed transparent data encryption (BYOK)

Key deletion or revocation

If using customer managed transparent data encryption (BYOK) and the serverless database is auto-paused when key deletion or revocation occurs, then the database remains in the auto-paused state. In this case, after the database is next resumed, the database becomes inaccessible within approximately 10 minutes. Once the database becomes inaccessible, the recovery process is the same as for provisioned compute databases. If the serverless database is online when key deletion or revocation occurs, then the database also becomes inaccessible within approximately 10 minutes in the same way as with provisioned compute databases.

Key rotation

If using customer-managed transparent data encryption (BYOK), and serverless auto-pausing is enabled, the database is auto-resumed whenever keys are rotated. The database will then be auto-paused when auto-pausing conditions are satisfied.

Create a new serverless database

Creating a new database or moving an existing database into a serverless compute tier follows the same pattern as creating a new database in provisioned compute tier and involves the following two steps:

1. Specify the service objective. The service objective prescribes the service tier, hardware configuration, and maximum vCores. For service objective options, see serverless resource limits

2. Optionally, specify the minimum vCores and auto-pause delay to change their default values. The following table shows the available values for these parameters.

   Parameter	Value choices	Default value
   Minimum vCores	Depends on maximum vCores configured - see resource limits.	0.5 vCores
   Auto-pause delay	Minimum: 15 minutes Maximum: 10,080 minutes (seven days) Increments: 1 minute Disable auto-pause: -1	60 minutes

The following examples create a new database in the serverless compute tier.

Use Azure portal

See Quickstart: Create a single database in Azure SQL Database using the Azure portal.

Use PowerShell

New-AzSqlDatabase -ResourceGroupName $resourceGroupName -ServerName $serverName -DatabaseName $databaseName `
  -Edition GeneralPurpose -ComputeModel Serverless -ComputeGeneration Gen5 `
  -MinVcore 0.5 -MaxVcore 2 -AutoPauseDelayInMinutes 720

New-AzSqlDatabase -ResourceGroupName $resourceGroupName -ServerName $serverName -DatabaseName $databaseName ` 
  -Edition Hyperscale -ComputeModel Serverless -ComputeGeneration Gen5 ` 
  -MinVcore 0.5 -MaxVcore 2

New-AzSqlDatabase -ResourceGroupName $resourceGroupName -ServerName $serverName -DatabaseName $databaseName `
  -Edition Hyperscale -ComputeModel Serverless -ComputeGeneration Gen5 `
  -MinVcore 0.5 -MaxVcore 2 `
  -HighAvailabilityReplicaCount 1

Use Azure CLI

az sql db create -g $resourceGroupName -s $serverName -n $databaseName `
  -e GeneralPurpose --compute-model Serverless -f Gen5 `
  --min-capacity 0.5 -c 2 --auto-pause-delay 720

az sql db create -g $resourceGroupName -s $serverName -n $databaseName ` 
  -e Hyperscale --compute-model Serverless -f Gen5 `
  --min-capacity 0.5 -c 2 

az sql db create -g $resourceGroupName -s $serverName -n $databaseName `
  -e Hyperscale --compute-model Serverless -f Gen5 `
  --min-capacity 0.5 -c 2 `
  --ha-replicas 1 --backup-storage-redundancy Zone

Use Transact-SQL (T-SQL)

When using T-SQL to create a new serverless database, default values are applied for the minimum vCores and auto-pause delay. Their values can later be changed from the Azure portal or via API including PowerShell, Azure CLI, and REST.

For details, see CREATE DATABASE.

CREATE DATABASE testdb
( EDITION = 'GeneralPurpose', SERVICE_OBJECTIVE = 'GP_S_Gen5_1' ) ;

CREATE DATABASE testdb
( EDITION = 'Hyperscale', SERVICE_OBJECTIVE = 'HS_S_Gen5_2') ; 

Move a database between compute tiers or service tiers

A database can be moved between the provisioned compute tier and serverless compute tier.

A serverless database can also be moved from the General Purpose service tier to the Hyperscale service tier. Review Manage Hyperscale databases to learn more.

When moving a database between compute tiers, specify the compute model parameter as either Serverless or Provisioned when using PowerShell or Azure CLI, or the SERVICE_OBJECTIVE when using T-SQL. Review resource limits to identify the appropriate service objective.

The following examples move an existing database from provisioned compute to serverless.

Use PowerShell

Set-AzSqlDatabase -ResourceGroupName $resourceGroupName -ServerName $serverName -DatabaseName $databaseName `
  -Edition GeneralPurpose -ComputeModel Serverless -ComputeGeneration Gen5 `
  -MinVcore 1 -MaxVcore 4 -AutoPauseDelayInMinutes 1440

Set-AzSqlDatabase -ResourceGroupName $resourceGroupName -ServerName $serverName -DatabaseName $databaseName ` 
  -Edition Hyperscale -ComputeModel Serverless -ComputeGeneration Gen5 ` 
  -MinVcore 1 -MaxVcore 4 

Use Azure CLI

az sql db update -g $resourceGroupName -s $serverName -n $databaseName `
  --edition GeneralPurpose --compute-model Serverless --family Gen5 `
  --min-capacity 1 --capacity 4 --auto-pause-delay 1440

az sql db update -g $resourceGroupName -s $serverName -n $databaseName ` 
  --edition Hyperscale --compute-model Serverless --family Gen5 `
  --min-capacity 1 --capacity 4 

Use Transact-SQL (T-SQL)

When using T-SQL to move a database between compute tiers, default values are applied for the minimum vCores and auto-pause delay. Their values can subsequently be changed from the Azure portal or via API including PowerShell, Azure CLI, and REST. For more information, see ALTER DATABASE.

ALTER DATABASE testdb 
MODIFY ( SERVICE_OBJECTIVE = 'GP_S_Gen5_1') ;

ALTER DATABASE testdb  
MODIFY ( SERVICE_OBJECTIVE = 'HS_S_Gen5_2') ; 

Modify serverless configuration

Use PowerShell

Use Set-AzSqlDatabase to modify the maximum or minimum vCores, and autopause delay. Use the MaxVcore, MinVcore, and AutoPauseDelayInMinutes arguments. Serverless auto-pausing is not currently supported in the Hyperscale tier, so the auto-pause delay argument is only applicable to the General Purpose tier.

Use Azure CLI

Use az sql db update to modify the maximum or minimum vCores, and autopause delay. Use the capacity, min-capacity, and auto-pause-delay arguments. Serverless auto-pausing is not currently supported in the Hyperscale tier, so the auto-pause delay argument is only applicable to the General Purpose tier.

Monitor

Resources used and billed

The resources of a serverless database include the app package, SQL instance, and user resource pool entities.

App package

The app package is the outer most resource management boundary for a database, regardless of whether the database is in a serverless or provisioned compute tier. The app package contains the SQL instance and external services like Full-text Search that all together scope all user and system resources used by a database in SQL Database. The SQL instance generally dominates the overall resource utilization across the app package.

User resource pool

The user resource pool is an inner resource management boundary for a database, regardless of whether the database is in a serverless or provisioned compute tier. The user resource pool scopes CPU and IO for user workload generated by DDL (CREATE and ALTER) and DML (INSERT, UPDATE, DELETE, and MERGE, and SELECT) queries. These queries generally represent the most substantial proportion of utilization within the app package.

Metrics

The following table includes metrics for monitoring the resource usage of the app package and user resource pool of a serverless database, including any geo-replicas:

Pause and resume status

In the case of a serverless database with auto-pausing enabled, the status it reports includes the following values:

Use Azure portal

In the Azure portal, the database status is displayed in the overview page of the database and in the overview page of its server. Also in the Azure portal, the history of pause and resume events of a serverless database can be viewed in the Activity log.

Use PowerShell

View the current database status using the following PowerShell example:

Get-AzSqlDatabase -ResourceGroupName $resourcegroupname -ServerName $servername -DatabaseName $databasename `
  | Select -ExpandProperty "Status"

Use Azure CLI

View the current database status using the following Azure CLI example:

az sql db show --name $databasename --resource-group $resourcegroupname --server $servername --query 'status' -o json

Resource limits

For resource limits, see serverless compute tier.

Billing

The amount of compute billed for a serverless database is the maximum of CPU used and memory used each second. If the amount of CPU and memory used is less than the minimum amount provisioned for each resource, then the provisioned amount is billed. In order to compare CPU with memory for billing purposes, memory is normalized into units of vCores by rescaling the number of GB by 3 GB per vCore.

• Resource billed: CPU and memory
• Amount billed: vCore unit price * maximum (minimum vCores, vCores used, minimum memory GB * 1/3, memory GB used * 1/3)
• Billing frequency: Per second

The vCore unit price is the cost per vCore per second.

Refer to the Azure SQL Database pricing page for specific unit prices in a given region.

The amount of compute billed in serverless for a General Purpose database, or a Hyperscale primary or named replica is exposed by the following metric:

• Metric: app_cpu_billed (vCore seconds)
• Definition: maximum (minimum vCores, vCores used, minimum memory GB * 1/3, memory GB used * 1/3)
• Reporting frequency: Per minute based on per second measurements aggregated over 1 minute.

Minimum compute bill

If a serverless database is paused, then the compute bill is zero. If a serverless database is not paused, then the minimum compute bill is no less than the amount of vCores based on maximum (minimum vCores, minimum memory GB * 1/3).

Examples:

• Suppose a serverless database in the General Purpose tier is not paused and configured with 8 maximum vCores and 1 minimum vCore corresponding to 3.0 GB minimum memory. Then the minimum compute bill is based on maximum (1 vCore, 3.0 GB * 1 vCore / 3 GB) = 1 vCore.
• Suppose a serverless database in the General Purpose tier is not paused and configured with 4 maximum vCores and 0.5 minimum vCores corresponding to 2.1 GB minimum memory. Then the minimum compute bill is based on maximum (0.5 vCores, 2.1 GB * 1 vCore / 3 GB) = 0.7 vCores.
• Suppose a serverless database in the Hyperscale tier has a primary replica with one HA replica and one named replica with no HA replicas. Suppose each replica is configured with 8 maximum vCores and 1 minimum vCore corresponding to 3 GB minimum memory. Then the minimum compute bill for the primary replica, HA replica, and named replica are each based on maximum (1 vCore, 3 GB * 1 vCore / 3 GB) = 1 vCore.

The Azure SQL Database pricing calculator for serverless can be used to determine the minimum memory configurable based on the number of maximum and minimum vCores configured. As a rule, if the minimum vCores configured is greater than 0.5 vCores, then the minimum compute bill is independent of the minimum memory configured and based only on the number of minimum vCores configured.

Scenario examples

Azure Hybrid Benefit and reserved capacity

Azure Hybrid Benefit (AHB) and reserved capacity discounts do not apply to the serverless compute tier.

Available regions

Serverless for General Purpose and Hyperscale tiers with support up to 40 maximum vCores is available worldwide except the following regions:

• China East
• China North

Related content

• To get started, see Quickstart: Create a single database - Azure SQL Database.
• For serverless service tier choices, see General Purpose and Hyperscale.