Autovacuum tuning in Azure Database for PostgreSQL - Flexible Server

APPLIES TO: Azure Database for PostgreSQL - Flexible Server

This article provides an overview of the autovacuum feature for Azure Database for PostgreSQL flexible server and the feature troubleshooting guides that are available to monitor the database bloat, autovacuum blockers and also information around how far the database is from emergency or wraparound situation.

What is autovacuum

Internal data consistency in PostgreSQL is based on the Multi-Version Concurrency Control (MVCC) mechanism, which allows the database engine to maintain multiple versions of a row and provides greater concurrency with minimal blocking between the different processes.

PostgreSQL databases need appropriate maintenance. For example, when a row is deleted, it isn't removed physically. Instead, the row is marked as "dead". Similarly for updates, the row is marked as "dead" and a new version of the row is inserted. These operations leave behind dead records, called dead tuples, even after all the transactions that might see those versions finish. Unless cleaned up, dead tuples remain, consuming disk space and bloating tables and indexes which result in slow query performance.

PostgreSQL uses a process called autovacuum to automatically clean-up dead tuples.

Autovacuum internals

Autovacuum reads pages looking for dead tuples, and if none are found, autovacuum discards the page. When autovacuum finds dead tuples, it removes them. The cost is based on:

• vacuum_cost_page_hit: Cost of reading a page that is already in shared buffers and doesn't need a disk read. The default value is set to 1.
• vacuum_cost_page_miss: Cost of fetching a page that isn't in shared buffers. The default value is set to 10.
• vacuum_cost_page_dirty: Cost of writing to a page when dead tuples are found in it. The default value is set to 20.

The amount of work autovacuum does depends on two parameters:

• autovacuum_vacuum_cost_limit is the amount of work autovacuum does in one go.
• autovacuum_vacuum_cost_delay number of milliseconds that autovacuum is asleep after it has reached the cost limit specified by the autovacuum_vacuum_cost_limit parameter.

In all currently supported versions of Postgres the default for autovacuum_vacuum_cost_limit is 200 (actually, it is set to -1 which makes it equals to the value of the regular vacuum_cost_limit which, by default, is 200).

As for autovacuum_vacuum_cost_delay, in Postgres version 11 it defaults to 20 milliseconds, while in Postgres versions 12 and above it defaults to 2 milliseconds.

Autovacuum wakes up 50 times (50*20 ms=1000 ms) every second. Every time it wakes up, autovacuum reads 200 pages.

That means in one-second autovacuum can do:

• ~80 MB/Sec [ (200 pages/vacuum_cost_page_hit) * 50 * 8 KB per page] if all pages with dead tuples are found in shared buffers.
• ~8 MB/Sec [ (200 pages/vacuum_cost_page_miss) * 50 * 8 KB per page] if all pages with dead tuples are read from disk.
• ~4 MB/Sec [ (200 pages/vacuum_cost_page_dirty) * 50 * 8 KB per page] autovacuum can write up to 4 MB/sec.

Monitor autovacuum

Use the following queries to monitor autovacuum:

select schemaname,relname,n_dead_tup,n_live_tup,round(n_dead_tup::float/n_live_tup::float*100) dead_pct,autovacuum_count,last_vacuum,last_autovacuum,last_autoanalyze,last_analyze from pg_stat_all_tables where n_live_tup >0;

The following columns help determine if autovacuum is catching up to table activity:

• dead_pct: percentage of dead tuples when compared to live tuples.
• last_autovacuum: The date of the last time the table was autovacuumed.
• last_autoanalyze: The date of the last time the table was automatically analyzed.

When does PostgreSQL trigger autovacuum

An autovacuum action (either ANALYZE or VACUUM) triggers when the number of dead tuples exceeds a particular number that is dependent on two factors: the total count of rows in a table, plus a fixed threshold. ANALYZE, by default, triggers when 10% of the table plus 50 rows changes, while VACUUM triggers when 20% of the table plus 50 rows changes. Since the VACUUM threshold is twice as high as the ANALYZE threshold, ANALYZE gets triggered earlier than VACUUM.

The exact equations for each action are:

• Autoanalyze = autovacuum_analyze_scale_factor * tuples + autovacuum_analyze_threshold
• Autovacuum = autovacuum_vacuum_scale_factor * tuples + autovacuum_vacuum_threshold

For example, analyze triggers after 60 rows change on a table that contains 100 rows, and vacuum triggers when 70 rows change on the table, using the following equations:

Autoanalyze = 0.1 * 100 + 50 = 60
Autovacuum = 0.2 * 100 + 50 = 70

Use the following query to list the tables in a database and identify the tables that qualify for the autovacuum process:

 SELECT *
      ,n_dead_tup > av_threshold AS av_needed
      ,CASE
        WHEN reltuples > 0
          THEN round(100.0 * n_dead_tup / (reltuples))
        ELSE 0
        END AS pct_dead
    FROM (
      SELECT N.nspname
        ,C.relname
        ,pg_stat_get_tuples_inserted(C.oid) AS n_tup_ins
        ,pg_stat_get_tuples_updated(C.oid) AS n_tup_upd
        ,pg_stat_get_tuples_deleted(C.oid) AS n_tup_del
        ,pg_stat_get_live_tuples(C.oid) AS n_live_tup
        ,pg_stat_get_dead_tuples(C.oid) AS n_dead_tup
        ,C.reltuples AS reltuples
        ,round(current_setting('autovacuum_vacuum_threshold')::INTEGER + current_setting('autovacuum_vacuum_scale_factor')::NUMERIC * C.reltuples) AS av_threshold
        ,date_trunc('minute', greatest(pg_stat_get_last_vacuum_time(C.oid), pg_stat_get_last_autovacuum_time(C.oid))) AS last_vacuum
        ,date_trunc('minute', greatest(pg_stat_get_last_analyze_time(C.oid), pg_stat_get_last_analyze_time(C.oid))) AS last_analyze
      FROM pg_class C
      LEFT JOIN pg_namespace N ON (N.oid = C.relnamespace)
      WHERE C.relkind IN (
          'r'
          ,'t'
          )
        AND N.nspname NOT IN (
          'pg_catalog'
          ,'information_schema'
          )
        AND N.nspname !~ '^pg_toast'
      ) AS av
    ORDER BY av_needed DESC ,n_dead_tup DESC;

Note

The query doesn't take into consideration that autovacuum can be configured on a per-table basis using the "alter table" DDL command.

Common autovacuum problems

Review the following list of possible common problems with the autovacuum process.

Not keeping up with busy server

The autovacuum process estimates the cost of every I/O operation, accumulates a total for each operation it performs and pauses once the upper limit of the cost is reached. autovacuum_vacuum_cost_delay and autovacuum_vacuum_cost_limit are the two server parameters that are used in the process.

By default, autovacuum_vacuum_cost_limit is set to –1, meaning autovacuum cost limit is the same value as the parameter vacuum_cost_limit, which defaults to 200. vacuum_cost_limit is the cost of a manual vacuum.

If autovacuum_vacuum_cost_limit is set to -1 then autovacuum uses the vacuum_cost_limit parameter, but if autovacuum_vacuum_cost_limit itself is set to greater than -1 then autovacuum_vacuum_cost_limit parameter is considered.

In case the autovacuum isn't keeping up, the following parameters might be changed:

Parameter	Description
autovacuum_vacuum_scale_factor	Default: 0.2, range: 0.05 - 0.1. The scale factor is workload-specific and should be set depending on the amount of data in the tables. Before changing the value, investigate the workload and individual table volumes.
autovacuum_vacuum_cost_limit	Default: 200. Cost limit might be increased. CPU and I/O utilization on the database should be monitored before and after making changes.
autovacuum_vacuum_cost_delay	Postgres Version 11 - Default: 20 ms. The parameter might be decreased to 2-10 ms. Postgres Versions 12 and above - Default: 2 ms.

Note

The autovacuum_vacuum_cost_limit value is distributed proportionally among the running autovacuum workers, so that if there is more than one, the sum of the limits for each worker doesn't exceed the value of the autovacuum_vacuum_cost_limit parameter

Autovacuum constantly running

Continuously running autovacuum might affect CPU and IO utilization on the server. The following might be possible reasons:

maintenance_work_mem

Autovacuum daemon uses autovacuum_work_mem that is by default set to -1 meaning autovacuum_work_mem would have the same value as the parameter maintenance_work_mem. This document assumes autovacuum_work_mem is set to -1 and maintenance_work_mem is used by the autovacuum daemon.

If maintenance_work_mem is low, it might be increased to up to 2 GB on Azure Database for PostgreSQL flexible server. A general rule of thumb is to allocate 50 MB to maintenance_work_mem for every 1 GB of RAM.

Large number of databases

Autovacuum tries to start a worker on each database every autovacuum_naptime seconds.

For example, if a server has 60 databases and autovacuum_naptime is set to 60 seconds, then the autovacuum worker starts every second [autovacuum_naptime/Number of DBs].

It's a good idea to increase autovacuum_naptime if there are more databases in a cluster. At the same time, the autovacuum process can be made more aggressive by increasing the autovacuum_cost_limit and decreasing the autovacuum_cost_delay parameters and increasing the autovacuum_max_workers from the default of 3 to 4 or 5.

Out of memory errors

Overly aggressive maintenance_work_mem values could periodically cause out-of-memory errors in the system. It's important to understand available RAM on the server before any change to the maintenance_work_mem parameter is made.

Autovacuum is too disruptive

If autovacuum is consuming a lot of resources, the following can be done:

Autovacuum parameters

Evaluate the parameters autovacuum_vacuum_cost_delay, autovacuum_vacuum_cost_limit, autovacuum_max_workers. Improperly setting autovacuum parameters might lead to scenarios where autovacuum becomes too disruptive.

If autovacuum is too disruptive, consider the following:

• Increase autovacuum_vacuum_cost_delay and reduce autovacuum_vacuum_cost_limit if set higher than the default of 200.
• Reduce the number of autovacuum_max_workers if it's set higher than the default of 3.

Too many autovacuum workers

Increasing the number of autovacuum workers won't necessarily increase the speed of vacuum. Having a high number of autovacuum workers isn't recommended.

Increasing the number of autovacuum workers will result in more memory consumption, and depending on the value of maintenance_work_mem , could cause performance degradation.

Each autovacuum worker process only gets (1/autovacuum_max_workers) of the total autovacuum_cost_limit, so having a high number of workers causes each one to go slower.

If the number of workers is increased, autovacuum_vacuum_cost_limit should also be increased and/or autovacuum_vacuum_cost_delay should be decreased to make the vacuum process faster.

However, if we have changed table level autovacuum_vacuum_cost_delay or autovacuum_vacuum_cost_limit parameters then the workers running on those tables are exempted from being considered in the balancing algorithm [autovacuum_cost_limit/autovacuum_max_workers].

Autovacuum transaction ID (TXID) wraparound protection

When a database runs into transaction ID wraparound protection, an error message like the following can be observed:

Database isn't accepting commands to avoid wraparound data loss in database 'xx'
Stop the postmaster and vacuum that database in single-user mode.

Note

This error message is a long-standing oversight. Usually, you do not need to switch to single-user mode. Instead, you can run the required VACUUM commands and perform tuning for VACUUM to run fast. While you cannot run any data manipulation language (DML), you can still run VACUUM.

The wraparound problem occurs when the database is either not vacuumed or there are too many dead tuples that couldn't be removed by autovacuum. The reasons for this might be:

Heavy workload

The workload could cause too many dead tuples in a brief period that makes it difficult for autovacuum to catch up. The dead tuples in the system add up over a period leading to degradation of query performance and leading to wraparound situation. One reason for this situation to arise might be because autovacuum parameters aren't adequately set and it isn't keeping up with a busy server.

Long-running transactions

Any long-running transactions in the system won't allow dead tuples to be removed while autovacuum is running. They're a blocker to the vacuum process. Removing the long running transactions frees up dead tuples for deletion when autovacuum runs.

Long-running transactions can be detected using the following query:

    SELECT pid, age(backend_xid) AS age_in_xids,
    now () - xact_start AS xact_age,
    now () - query_start AS query_age,
    state,
    query
    FROM pg_stat_activity
    WHERE state != 'idle'
    ORDER BY 2 DESC
    LIMIT 10;

Prepared statements

If there are prepared statements that aren't committed, they would prevent dead tuples from being removed.
The following query helps find noncommitted prepared statements:

    SELECT gid, prepared, owner, database, transaction
    FROM pg_prepared_xacts
    ORDER BY age(transaction) DESC;

Use COMMIT PREPARED or ROLLBACK PREPARED to commit or roll back these statements.

Unused replication slots

Unused replication slots prevent autovacuum from claiming dead tuples. The following query helps identify unused replication slots:

    SELECT slot_name, slot_type, database, xmin
    FROM pg_replication_slots
    ORDER BY age(xmin) DESC;

Use pg_drop_replication_slot() to delete unused replication slots.

When the database runs into transaction ID wraparound protection, check for any blockers as mentioned previously, and remove those manually for autovacuum to continue and complete. You can also increase the speed of autovacuum by setting autovacuum_cost_delay to 0 and increasing the autovacuum_cost_limit to a value greater than 200. However, changes to these parameters won't be applied to existing autovacuum workers. Either restart the database or kill existing workers manually to apply parameter changes.

Table-specific requirements

Autovacuum parameters might be set for individual tables. It's especially important for small and big tables. For example, for a small table that contains only 100 rows, autovacuum triggers VACUUM operation when 70 rows change (as calculated previously). If this table is frequently updated, you might see hundreds of autovacuum operations a day. This prevents autovacuum from maintaining other tables on which the percentage of changes aren't as big. Alternatively, a table containing a billion rows needs to change 200 million rows to trigger autovacuum operations. Setting autovacuum parameters appropriately prevents such scenarios.

To set autovacuum setting per table, change the server parameters as the following examples:

    ALTER TABLE <table name> SET (autovacuum_analyze_scale_factor = xx);
    ALTER TABLE <table name> SET (autovacuum_analyze_threshold = xx);
    ALTER TABLE <table name> SET (autovacuum_vacuum_scale_factor = xx);
    ALTER TABLE <table name> SET (autovacuum_vacuum_threshold = xx);
    ALTER TABLE <table name> SET (autovacuum_vacuum_cost_delay = xx);
    ALTER TABLE <table name> SET (autovacuum_vacuum_cost_limit = xx);

Insert-only workloads

In versions of PostgreSQL prior to 13, autovacuum won't run on tables with an insert-only workload, because if there are no updates or deletes, there are no dead tuples and no free space that needs to be reclaimed. However, autoanalyze will run for insert-only workloads since there's new data. The disadvantages of this are:

• The visibility map of the tables isn't updated, and thus query performance, especially where there are Index Only Scans, starts to suffer over time.
• The database can run into transaction ID wraparound protection.
• Hint bits won't be set.

Solutions

Postgres versions prior to 13

Using the pg_cron extension, a cron job can be set up to schedule a periodic vacuum analyze on the table. The frequency of the cron job depends on the workload.

For step-by-step guidance using pg_cron, review Extensions.

Postgres 13 and higher versions

Autovacuum will run on tables with an insert-only workload. Two new server parameters autovacuum_vacuum_insert_threshold and  autovacuum_vacuum_insert_scale_factor help control when autovacuum can be triggered on insert-only tables.

Troubleshooting guides

Using the feature troubleshooting guides which is available on the Azure Database for PostgreSQL flexible server portal it is possible to monitor bloat at database or individual schema level along with identifying potential blockers to autovacuum process. Two troubleshooting guides are available first one is autovacuum monitoring that can be used to monitor bloat at database or individual schema level. The second troubleshooting guide is autovacuum blockers and wraparound which helps to identify potential autovacuum blockers along with information on how far the databases on the server are from wraparound or emergency situation. The troubleshooting guides also share recommendations to mitigate potential issues. How to set up the troubleshooting guides to use them please follow setup troubleshooting guides.

Related content

• High CPU Utilization
• High Memory Utilization
• Server Parameters