Manage computes

A compute is a virtualized service that runs applications. In Neon, a compute runs Postgres.

Each project has a primary read-write compute for its default branch. Neon supports both read-write and read replica computes. A branch can have one primary (read-write) compute and multiple read replica computes. A compute is required to connect to a Neon branch (where your database resides) from a client or application.

To connect to a database in a branch, you must use a compute associated with that branch. The following diagram illustrates how an application connects to a branch via its compute:

Project
    |---- default branch (main) ---- compute <--- application/client
             |    |
             |    |---- database
             |
             ---- child branch ---- compute <--- application/client
                            |
                            |---- database

Your Neon plan determines the resources available to a compute. The Neon Free plan supports computes with up to 2 CU (8 GB of RAM). Paid plans offer larger compute sizes. Larger computes consume more compute hours over the same period of active time than smaller computes.

View a compute

A compute is associated with a branch.

Create a compute

You can only create a single primary read-write compute for a branch that does not have a compute, but a branch can have multiple read replica computes.

Edit a compute

You can edit a compute to change the compute size or scale to zero configuration.

For information about selecting an appropriate compute size or autoscaling configuration, see How to size your compute.

What happens to the compute when making changes

Some key points to understand about how your endpoint responds when you make changes to your compute settings:

• Changing the size of your fixed compute restarts the endpoint and temporarily disconnects all existing connections.

  note

  When your compute resizes automatically as part of the autoscaling feature, there are no restarts or disconnects; it just scales.

• Editing minimum or maximum autoscaling sizes also requires a restart; existing connections are temporarily disconnected.

• If you disable scale to zero, you may need to restart your compute manually to get the latest compute-related release updates from Neon if updates are not applied automatically by a scheduled update. Scheduled updates are applied according to certain criteria, so not all computes receive these updates automatically. See Restart a compute.

To avoid prolonged interruptions resulting from compute restarts, we recommend configuring your clients and applications to reconnect automatically in case of a dropped connection. See Handling connection disruptions.

Compute size and autoscaling configuration

You can change compute size settings when editing a compute.

Compute size is the number of Compute Units (CUs) assigned to a Neon compute. The number of CUs determines the processing capacity of the compute. Each CU allocates approximately 4 GB of RAM to the database instance, along with associated CPU and local SSD resources. Scaling up increases these resources linearly, as shown in the table below.

Neon supports fixed-size and autoscaling compute configurations.

• Fixed size: Select a fixed compute size ranging from .25 CUs to 56 CUs. A fixed-size compute does not scale to meet workload demand.
• Autoscaling: Specify a minimum and maximum compute size. Neon scales the compute size up and down within the selected compute size boundaries in response to the current load. Currently, the Autoscaling feature supports a range of .25 CU to 16 CU. The maximum permitted autoscaling range is 8 CU, meaning the difference between your maximum and minimum cannot exceed 8 CU. The .25 CU and .5 CU settings are shared compute. For information about how Neon implements the Autoscaling feature, see Autoscaling.

How to size your compute

The size of your compute determines the amount of frequently accessed data you can cache in memory and the maximum number of simultaneous connections you can support. As a result, if your compute size is too small, this can lead to suboptimal query performance and connection limit issues.

In Postgres, the shared_buffers setting defines the amount of data that can be held in memory. In Neon, up to 75% of your compute's RAM is used for data caching.

The Postgres max_connections setting defines your compute's maximum simultaneous connection limit and is set according to your compute size configuration.

The following table outlines the RAM, compute cache size (75% of RAM), and the max_connections limit for each compute size that Neon supports. To understand how max_connections is determined for an autoscaling configuration, see Parameter settings that differ by compute size.

When selecting a compute size, ideally, you want to keep as much of your dataset in memory as possible. This improves performance by reducing the amount of reads from storage. If your dataset is not too large, select a compute size that will hold the entire dataset in memory. For larger datasets that cannot be fully held in memory, select a compute size that can hold your working set. Selecting a compute size for a working set involves advanced steps, which are outlined below. See Sizing your compute based on the working set.

Regarding connection limits, you'll want a compute size that can support your anticipated maximum number of concurrent connections. If you are using Autoscaling, it is important to remember that your max_connections setting is based on both your minimum and the maximum compute size. See Parameter settings that differ by compute size for details. To avoid any max_connections constraints, you can use a pooled connection with your application, which supports up to 10,000 concurrent user connections. See Connection pooling.

Sizing your compute based on the working set

If it's not possible to hold your entire dataset in memory, the next best option is to ensure that your working set is in memory. A working set is your frequently accessed or recently used data and indexes. To determine whether your working set is fully in memory, you can query the cache hit ratio for your Neon compute. The cache hit ratio tells you how many queries are served from memory. Queries not served from memory bypass the cache to retrieve data from database storage (the Pageserver), which can affect query performance.

You can monitor your compute cache hit rate and your working set size from Neon's Monitoring page, where you'll find the following charts:

• Compute cache hit rate
• Working set size

Neon also provides a neon extension with a neon_stat_file_cache view that you can use to query the cache hit ratio for your compute. For more information, see The neon extension.

Autoscaling considerations

Autoscaling is most effective when your data (either your full dataset or your working set) can be fully cached in memory on the minimum compute size in your autoscaling configuration.

Consider this scenario: If your data size is approximately 6 GB, starting with a compute size of .25 CU can lead to suboptimal performance because your data cannot be adequately cached. While your compute will scale up from .25 CU on demand, you may experience poor query performance until your compute scales up and fully caches your working set. You can avoid this issue if your minimum compute size can hold your working set in memory.

As mentioned above, your max_connections setting is based on both your minimum and maximum compute size settings. To avoid any max_connections constraints, you can use a pooled connection for your application. See Connection pooling.

Scale to zero configuration

Neon's Scale to Zero feature automatically transitions a compute into an idle state after 5 minutes of inactivity. You can disable scale to zero to maintain an "always-active" compute. An "always-active" configuration eliminates the few hundred milliseconds seconds of latency required to reactivate a compute but is likely to increase your compute time usage on systems where the database is not always active.

For more information, refer to Configuring scale to zero for Neon computes.

Restart a compute

It is sometimes necessary to restart a compute. Reasons for restarting a compute might include:

• Activating new limits after upgrading to a paid plan
• Getting the latest compute-related updates, which Neon typically releases weekly
• Accessing a recently released Postgres extension or extension version
• Resolving performance issues or unexpected behavior

Restarting ensures your compute is running with the latest configurations and improvements.

Delete a compute

A branch can have a single read-write compute and multiple read replica computes. You can delete any of these computes from a branch. However, be aware that a compute is required to connect to a branch and access its data. If you delete a compute and add it back later, the new compute will have different connection details.

Compute-related issues

This section outlines compute-related issues you may encounter and possible resolutions.

No space left on device

You may encounter an error similar to the following when your compute's local disk storage is full:

ERROR: could not write to file "base/pgsql_tmp/pgsql_tmp1234.56.fileset/o12of34.p1.0": No space left on device (SQLSTATE 53100)

Neon computes allocate 20 GiB of local disk space or 15 GiB x the maximum compute size (whichever is highest) for temporary files used by Postgres. Data-intensive operations can sometimes consume all of this space, resulting in No space left on device errors.

To resolve this issue, you can try the following strategies:

• Identify and terminate resource-intensive processes: These could be long-running queries, operations, or possibly sync or replication activities. You can start your investigation by listing running queries by duration.
• Optimize queries to reduce temporary file usage.
• Adjust pipeline settings for third-party sync or replication: If you're syncing or replicating data with an external service, modify the pipeline settings to control disk space usage.

If the issue persists, refer to our Neon Support channels.

Compute is not suspending

In some cases, you may observe that your compute remains constantly active for no apparent reason. Possible causes for a constantly active compute when not expected include:

• Connection requests: Frequent connection requests from clients, applications, or integrations can prevent a compute from suspending automatically. Each connection resets the scale to zero timer.
• Background processes: Some applications or background jobs may run periodic tasks that keep the connection active.

Possible steps you can take to identify the issues include:

1. Checking for active processes

   You can run the following query to identify active sessions and their states:

   SELECT
     pid,
     usename,
     query,
     state,
     query_start
   FROM
     pg_stat_activity
   WHERE
     query_start >= now() - interval '24 hours'
   ORDER BY
     query_start DESC;

   Look for processes initiated by your users, applications, or integrations that may be keeping your compute active.

2. Review connection patterns

   • Ensure that no applications are sending frequent, unnecessary connection requests.
   • Consider batching connections if possible, or use connection pooling to limit persistent connections.

3. Optimize any background jobs

   If background jobs are needed, reduce their frequency or adjust their timing to allow Neon's scale to zero feature to activate after the defined period of inactivity (the default is 5 minutes). For more information, refer to our Scale to zero guide.