Neon Deploy: Learn how Neon helps developers ship faster with Postgres. Grab your ticket

The pg_prewarm extension

Load data into your Postgres buffer cache with the pg_prewarm extension

You can use the pg_prewarm extension to preload data into the Postgres buffer cache after a restart. Doing so improves query response times by ensuring that your data is readily available in memory. Otherwise, data must be loaded into the buffer cache from disk on-demand, which can result in slower query response times.

Try it on Neon!

Neon is Serverless Postgres built for the cloud. Explore Postgres features and functions in our user-friendly SQL editor. Sign up for a free account to get started.

Sign Up

In this guide, we'll explore the pg_prewarm extension, how to enable it, and how to use it to prewarm your Postgres buffer cache.

note

The pg_prewarm extension is open-source and can be installed on any Postgres setup. Detailed information about the extension is available in the PostgreSQL Documentation.

Version availability

Please refer to the list of extensions available in Neon for information about the version of pg_prewarm that Neon supports.

Enable the pg_prewarm extension

Enable the pg_prewarm extension by running the CREATE EXTENSION statement in your Postgres client:

CREATE EXTENSION IF NOT EXISTS pg_prewarm;

For information about using the Neon SQL Editor, see Query with Neon's SQL Editor. For information about using the psql client with Neon, see Connect with psql.

Basic usage

To prewarm a specific table, simply use the pg_prewarm function with the name of the table you want to cache.

SELECT pg_prewarm('table_name');

Replace table_name with the actual name of your table.

The output of SELECT pg_prewarm() is the number of blocks from the specified table that was loaded into the Postgres buffer cache. The default block size in Postgres is 8192 bytes (8KB).

The pg_prewarm function does not support specifying multiple table names in a single command. It's designed to work with a single table at a time. If you want to prewarm multiple tables, you would need to call pg_prewarm separately for each.

Running pg_prewarm on indexes

Running pg_prewarm on frequently-used indexes can help improve query performance after a Postgres restart. You might also run pg_prewarm on indexes that are not frequently used but will be involved in upcoming heavy read operations.

Running pg_prewarm on an index is similar to running it on a table, but you specify the index's fully qualified name (schema name plus index name) or OID (Object Identifier) instead.

Here's an example that demonstrates how to use pg_prewarm to preload an index into memory:

SELECT pg_prewarm('schema_name.index_name');

Replace schema_name.index_name with the actual schema and index name you want to prewarm. If you're not sure about the index name or want to list all indexes for a specific table, you can use the pg_indexes view to find out. Here's how you might query for index names:

SELECT indexname FROM pg_indexes WHERE tablename = 'your_table_name';

Replace your_table_name with the name of the table whose indexes you're interested in. Once you have the index name, you can then use pg_prewarm as shown above.

Additionally, if you prefer to use the index's OID, you can find it using the pg_class system catalog. Here's how to find an index's OID:

SELECT oid FROM pg_class WHERE relname = 'index_name';

Then, you can use the OID with pg_prewarm like so:

SELECT pg_prewarm(your_index_oid);

Check the proportion of a table loaded into memory

In this example, you create a table, check its data size, run pg_prewarm, and then check to see how much of the table's data was loaded into memory.

  1. First, create a table and populate it with some data:

    CREATE TABLE t_test AS
    SELECT * FROM generate_series(1, 1000000) AS id;
  2. Check the size of the table:

    SELECT pg_size_pretty(pg_relation_size('t_test')) AS table_size_pretty,
        pg_relation_size('t_test') AS table_size_bytes;

    This command returns the size of the table in both MB and bytes.

    table_size_pretty | table_size_bytes
    -------------------+------------------
    35 MB              |         36700160
  3. Load the table data into the Postgres buffer cache using pg_prewarm:

    SELECT pg_prewarm('public.t_test') AS blocks_loaded;

    This will output the number of blocks that were loaded:

    blocks_loaded
    ---------------
            4480
  4. To understand the calculation that follows, check the block size of your Postgres instance:

    SHOW block_size;

    The default block size in Postgres is 8192 bytes (8KB). We'll use this value in the next step.

    block_size
    ------------
    8192
  5. Calculate the total size of the data loaded into the cache using the block size and the number of blocks loaded:

    -- Assuming 4480 blocks were loaded (replace with your actual number from pg_prewarm output)
    SELECT 4480 * 8192 AS loaded_data_bytes;

    You can now compare this value with the size of your table.

    loaded_data_bytes
    -------------------
            36700160

    note

    The values for the size of the table and the size of the data loaded into the buffer cache as shown in the example above match exactly, which is an ideal scenario. However, there are cases where these values might not match, indicating that not all the data was loaded into the buffer cache; for example, this can happen if pg_prewarm only partially loads the table into the buffer cache due to lack of memory availability. Concurrent data modifications could also cause sizes to differ.

    To understand how much memory is available to your Postgres instance on Neon, see How to size your compute.

Demonstrating the effect of pg_prewarm

This example shows how preloading data can improve query performance. We'll create two tables with the same data, preload one table, and then run EXPLAIN ANALYZE to compare execution time results.

  1. Create two sample tables with the same data for comparison:

    CREATE TABLE tbl_transactions_1
    (
        tran_id_ SERIAL,
        transaction_date TIMESTAMPTZ,
        transaction_name TEXT
    );
    
    INSERT INTO tbl_transactions_1
    (transaction_date, transaction_name)
    SELECT x, 'dbrnd'
    FROM generate_series('2010-01-01 00:00:00'::timestamptz, '2018-02-01 00:00:00'::timestamptz, '1 minutes'::interval) a(x);
    CREATE TABLE tbl_transactions_2
    (
        tran_id_ SERIAL,
        transaction_date TIMESTAMPTZ,
        transaction_name TEXT
    );
    
    INSERT INTO tbl_transactions_2
    (transaction_date, transaction_name)
    SELECT x, 'dbrnd'
    FROM generate_series('2010-01-01 00:00:00'::timestamptz, '2018-02-01 00:00:00'::timestamptz, '1 minutes'::interval) a(x);
  2. Restart your Postgres instance to clear the cache. On Neon, you can do this by restarting your compute.

  3. Prewarm the first sample table:

    SELECT pg_prewarm('tbl_transactions_1') AS blocks_loaded;

    This will output the number of blocks that were loaded into the cache:

    blocks_loaded
    ---------------
            27805
  4. Now, compare the execution plan of the prewarmed table vs. a non-prewarmed table to see the performance improvement.

    EXPLAIN ANALYZE SELECT * FROM tbl_transactions_1;
    EXPLAIN ANALYZE SELECT * FROM tbl_transactions_2;

    The execution time for the prewarmed table should be significantly lower than for the table that has not been prewarmed, as shown here:

    EXPLAIN ANALYZE SELECT * FROM tbl_transactions_1;
                                                          QUERY PLAN
    -------------------------------------------------------------------------------------------------------------------------------
    Seq Scan on tbl_transactions_1  (cost=0.00..69608.21 rows=4252321 width=18) (actual time=0.017..228.995 rows=4252321 loops=1)
    Planning Time: 1.134 ms
    Execution Time: 344.028 ms
    (3 rows)
    
    EXPLAIN ANALYZE SELECT * FROM tbl_transactions_2;
                                                            QUERY PLAN
    ---------------------------------------------------------------------------------------------------------------------------------
    Seq Scan on tbl_transactions_2  (cost=0.00..69608.21 rows=4252321 width=18) (actual time=2.251..11859.232 rows=4252321 loops=1)
    Planning Time: 0.216 ms
    Execution Time: 11994.066 ms
    (3 rows)

Conclusion

Prewarming your table data and indexes can help improve read performance, especially after a database restart or for indexes that are not frequently used but will be involved in upcoming heavy read operations. However, it's important to use this feature cautiously, especially on systems with limited memory, to avoid potential negative impacts on overall performance.

Resources

Need help?

Join our Discord Server to ask questions or see what others are doing with Neon. Users on paid plans can open a support ticket from the console. For more details, see Getting Support.

Last updated on

Was this page helpful?