Edit

Limits in Azure Database for PostgreSQL - Flexible Server

  • Article

APPLIES TO: Azure Database for PostgreSQL - Flexible Server

The following sections describe capacity and functional limits in the database service. If you'd like to learn about resource (compute, memory, storage) tiers, see the compute and storage article.

Maximum connections

The default maximum number of connections per pricing tier and vCores are shown below. The Azure system requires three connections to monitor the Azure Database for PostgreSQL - Flexible Server.

SKU Name vCores Memory Size Max Connections Max User Connections
Burstable
B1ms 1 2 GiB 50 47
B2s 2 4 GiB 100 97
B2ms 2 4 GiB 100 97
B4ms 4 8 GiB 859 856
B8ms 8 16 GiB 1719 1716
B12ms 12 24 GiB 2578 2575
B16ms 16 32 GiB 3438 3435
B20ms 20 40 GiB 4297 4294
General Purpose
D2s_v3 / D2ds_v4 / D2ds_v5 2 8 GiB 859 856
D4s_v3 / D4ds_v4 / D4ds_v5 4 16 GiB 1719 1716
D8s_v3 / D8ds_V4 / D8ds_v5 8 32 GiB 3438 3435
D16s_v3 / D16ds_v4 / D16ds_v5 16 64 GiB 5000 4997
D32s_v3 / D32ds_v4 / D32ds_v5 32 128 GiB 5000 4997
D48s_v3 / D48ds_v4 / D48ds_v5 48 192 GiB 5000 4997
D64s_v3 / D64ds_v4 / D64ds_v5 64 256 GiB 5000 4997
D96ds_v5 96 384 GiB 5000 4997
Memory Optimized
E2s_v3 / E2ds_v4 / E2ds_v5 2 16 GiB 1719 1716
E4s_v3 / E4ds_v4 / E4ds_v5 4 32 GiB 3438 3433
E8s_v3 / E8ds_v4 / E8ds_v5 8 64 GiB 5000 4997
E16s_v3 / E16ds_v4 / E16ds_v5 16 128 GiB 5000 4997
E20ds_v4 / E20ds_v5 20 160 GiB 5000 4997
E32s_v3 / E32ds_v4 / E32ds_v5 32 256 GiB 5000 4997
E48s_v3 / E48ds_v4 / E48ds_v5 48 384 GiB 5000 4997
E64s_v3 / E64ds_v4 / E64ds_v5 64 432 GiB 5000 4997
E96ds_v5 96 672 GiB 5000 4997

Changing the max_connections value

Customers can change the value maximum number of connections using either of the following methods:

  • Change the default value for the max_connections parameter using server parameter. This parameter is static and requires an instance restart.

Caution

While it is possible to increase the value of "max_connections" beyond the default setting, it is not advisable. The rationale behind this recommendation is that instances may encounter difficulties when the workload expands and demands more memory. As the number of connections increases, memory usage also rises. Instances with limited memory may face issues such as crashes or high latency. Although a higher value for "max_connections" might be acceptable when most connections are idle, it can lead to significant performance problems once they become active. Instead, if you require additional connections, we suggest utilizing pgBouncer, Azure's built-in connection pool management solution, in transaction mode. To start, it is recommended to use conservative values by multiplying the vCores within the range of 2 to 5. Afterward, carefully monitor resource utilization and application performance to ensure smooth operation. For detailed information on pgBouncer, please refer to the PgBouncer in Azure Database for PostgreSQL - Flexible Server documentation.

  • Scale your Azure Postgres instance up to a SKU with more memory size.

Note

Scaling up Azure Postgres instances impacts the account’s billing. To learn more, refer Azure Database for PostgreSQL pricing.

When connections exceed the limit, you may receive the following error:

FATAL: sorry, too many clients already.

When using PostgreSQL for a busy database with a large number of concurrent connections, there may be a significant strain on resources. This strain can result in high CPU utilization, particularly when many connections are established simultaneously and when connections have short durations (less than 60 seconds). These factors can negatively impact overall database performance by increasing the time spent on processing connections and disconnections. It's important to note that each connection in Postgres, regardless of whether it is idle or active, consumes a significant amount of resources from your database. This consumption can lead to performance issues beyond high CPU utilization, such as disk and lock contention. The topic is discussed in more detail in the PostgreSQL Wiki article on the Number of Database Connections. To learn more, visit Identify and solve connection performance in Azure Postgres.

Functional limitations

Scale operations

  • At this time, scaling up the server storage requires a server restart.
  • Server storage can only be scaled in 2x increments, see Compute and Storage for details.
  • Decreasing server storage size is currently not supported. Only way to do is dump and restore it to a new Flexible Server.

Server version upgrades

  • Automated migration between major database engine versions is currently not supported. If you would like to upgrade to the next major version, take a dump and restore it to a server that was created with the new engine version.

Storage

  • Once configured, storage size can't be reduced. You have to create a new server with desired storage size, perform manual dump and restore and migrate your database(s) to the new server.
  • Currently, storage auto-grow feature isn't available. You can monitor the usage and increase the storage to a higher size.
  • When the storage usage reaches 95% or if the available capacity is less than 5 GiB whichever is more, the server is automatically switched to read-only mode to avoid errors associated with disk-full situations. In rare cases, if the rate of data growth outpaces the time it takes switch to read-only mode, your Server may still run out of storage.
  • We recommend setting alert rules for storage used or storage percent when they exceed certain thresholds so that you can proactively take action such as increasing the storage size. For example, you can set an alert if the storage percent exceeds 80% usage.
  • If you're using logical replication, then you must drop the logical replication slot in the primary server if the corresponding subscriber no longer exists. Otherwise the WAL files start to get accumulated in the primary filling up the storage. If the storage threshold exceeds certain threshold and if the logical replication slot isn't in use (due to non-available subscriber), Flexible server automatically drops that unused logical replication slot. That action releases accumulated WAL files and avoids your server becoming unavailable due to storage getting filled situation.
  • We don't support the creation of tablespaces, so if you're creating a database, don’t provide a tablespace name. PostgreSQL will use the default one that is inherited from the template database. It's unsafe to provide a tablespace like the temporary one because we can't ensure that such objects will remain persistent after server restarts, HA failovers, etc.

Networking

  • Moving in and out of VNET is currently not supported.
  • Combining public access with deployment within a VNET is currently not supported.
  • Firewall rules aren't supported on VNET, Network security groups can be used instead.
  • Public access database servers can connect to public internet, for example through postgres_fdw, and this access can't be restricted. VNET-based servers can have restricted outbound access using Network Security Groups.

High availability (HA)

Availability zones

  • Manually moving servers to a different availability zone is currently not supported. However, you can enable HA using the preferred AZ as the standby zone. Once established, you can fail over to the standby and then disable HA.

Postgres engine, extensions, and PgBouncer

  • Postgres 10 and older aren't supported as those are already retired by the open-source community. If you must use one of these versions, you need to use the Single Server option, which supports the older major versions 95, 96 and 10.
  • Flexible Server supports all contrib extensions and more. Please refer to PostgreSQL extensions.
  • Built-in PgBouncer connection pooler is currently not available for Burstable servers.

Stop/start operation

  • Once you stop the Flexible Server, it automatically starts after 7- days.

Scheduled maintenance

  • You can change custom maintenance window to any day/time of the week. However, any changes made after receiving the maintenance notification will have no impact on the next maintenance. Changes only take effect with the following monthly scheduled maintenance.

Backing up a server

  • Backups are managed by the system, there's currently no way to run these backups manually. We recommend using pg_dump instead.
  • The first snapshot is a full backup and consecutive snapshots are differential backups. The differential backups only back up the changed data since the last snapshot backup. For example, if the size of your database is 40 GB and your provisioned storage is 64 GB, the first snapshot backup will be 40 GB. Now, if you change 4 GB of data, then the next differential snapshot backup size will only be 4 GB. The transaction logs (write ahead logs - WAL) are separate from the full/differential backups, and are archived continuously.

Restoring a server

  • When using the Point-in-time-Restore feature, the new server is created with the same compute and storage configurations as the server it is based on.
  • VNET based database servers are restored into the same VNET when you restore from a backup.
  • The new server created during a restore doesn't have the firewall rules that existed on the original server. Firewall rules need to be created separately for the new server.
  • Restore to a different subscription isn't supported but as a workaround, you can restore the server within the same subscription and then migrate the restored server to a different subscription.

Next steps