22.1.1. Recovering disk space

In normal PostgreSQL operation, an UPDATE or DELETE of a row does not immediately remove the old version of the row. This approach is necessary to gain the benefits of multiversion concurrency control (see Chapter 12): the row version must not be deleted while it is still potentially visible to other transactions. But eventually, an outdated or deleted row version is no longer of interest to any transaction. The space it occupies must be reclaimed for reuse by new rows, to avoid infinite growth of disk space requirements. This is done by running VACUUM.

Clearly, a table that receives frequent updates or deletes will need to be vacuumed more often than tables that are seldom updated. It may be useful to set up periodic cron tasks that VACUUM only selected tables, skipping tables that are known not to change often. This is only likely to be helpful if you have both large heavily-updated tables and large seldom-updated tables — the extra cost of vacuuming a small table isn't enough to be worth worrying about.

There are two variants of the VACUUM command. The first form, known as "lazy vacuum" or just VACUUM, marks expired data in tables and indexes for future reuse; it does not attempt to reclaim the space used by this expired data unless the space is at the end of the table and an exclusive table lock can be easily obtained. Unused space at the start or middle of the file does not result in the file being shortened and space returned to the operating system. This variant of VACUUM can be run concurrently with normal database operations.

The second form is the VACUUM FULL command. This uses a more aggressive algorithm for reclaiming the space consumed by expired row versions. Any space that is freed by VACUUM FULL is immediately returned to the operating system. Unfortunately, this variant of the VACUUM command acquires an exclusive lock on each table while VACUUM FULL is processing it. Therefore, frequently using VACUUM FULL can have an extremely negative effect on the performance of concurrent database queries.

The standard form of VACUUM is best used with the goal of maintaining a fairly level steady-state usage of disk space. If you need to return disk space to the operating system you can use VACUUM FULL — but what's the point of releasing disk space that will only have to be allocated again soon? Moderately frequent standard VACUUM runs are a better approach than infrequent VACUUM FULL runs for maintaining heavily-updated tables.

Recommended practice for most sites is to schedule a database-wide VACUUM once a day at a low-usage time of day, supplemented by more frequent vacuuming of heavily-updated tables if necessary. (Some installations with an extremely high rate of data modification VACUUM busy tables as often as once every few minutes.) If you have multiple databases in a cluster, don't forget to VACUUM each one; the program vacuumdb may be helpful.

VACUUM FULL is recommended for cases where you know you have deleted the majority of rows in a table, so that the steady-state size of the table can be shrunk substantially with VACUUM FULL's more aggressive approach. Use plain VACUUM, not VACUUM FULL, for routine vacuuming for space recovery.

If you have a table whose contents are deleted on a periodic basis, consider doing it with TRUNCATE rather than using DELETE followed by VACUUM. TRUNCATE removes the entire content of the table immediately, without requiring a subsequent VACUUM or VACUUM FULL to reclaim the now-unused disk space.

22.1.2. Updating planner statistics

The PostgreSQL query planner relies on statistical information about the contents of tables in order to generate good plans for queries. These statistics are gathered by the ANALYZE command, which can be invoked by itself or as an optional step in VACUUM. It is important to have reasonably accurate statistics, otherwise poor choices of plans may degrade database performance.

As with vacuuming for space recovery, frequent updates of statistics are more useful for heavily-updated tables than for seldom-updated ones. But even for a heavily-updated table, there may be no need for statistics updates if the statistical distribution of the data is not changing much. A simple rule of thumb is to think about how much the minimum and maximum values of the columns in the table change. For example, a timestamp column that contains the time of row update will have a constantly-increasing maximum value as rows are added and updated; such a column will probably need more frequent statistics updates than, say, a column containing URLs for pages accessed on a website. The URL column may receive changes just as often, but the statistical distribution of its values probably changes relatively slowly.

It is possible to run ANALYZE on specific tables and even just specific columns of a table, so the flexibility exists to update some statistics more frequently than others if your application requires it. In practice, however, the usefulness of this feature is doubtful. Beginning in PostgreSQL 7.2, ANALYZE is a fairly fast operation even on large tables, because it uses a statistical random sampling of the rows of a table rather than reading every single row. So it's probably much simpler to just run it over the whole database every so often.

Recommended practice for most sites is to schedule a database-wide ANALYZE once a day at a low-usage time of day; this can usefully be combined with a nightly VACUUM. However, sites with relatively slowly changing table statistics may find that this is overkill, and that less-frequent ANALYZE runs are sufficient.

22.1.3. Preventing transaction ID wraparound failures

PostgreSQL's MVCC transaction semantics depend on being able to compare transaction ID (XID) numbers: a row version with an insertion XID greater than the current transaction's XID is "in the future" and should not be visible to the current transaction. But since transaction IDs have limited size (32 bits at this writing) a cluster that runs for a long time (more than 4 billion transactions) would suffer transaction ID wraparound: the XID counter wraps around to zero, and all of a sudden transactions that were in the past appear to be in the future — which means their outputs become invisible. In short, catastrophic data loss. (Actually the data is still there, but that's cold comfort if you can't get at it.)

Prior to PostgreSQL 7.2, the only defense against XID wraparound was to re-initdb at least every 4 billion transactions. This of course was not very satisfactory for high-traffic sites, so a better solution has been devised. The new approach allows a server to remain up indefinitely, without initdb or any sort of restart. The price is this maintenance requirement: every table in the database must be vacuumed at least once every billion transactions .

In practice this isn't an onerous requirement, but since the consequences of failing to meet it can be complete data loss (not just wasted disk space or slow performance), some special provisions have been made to help database administrators avoid disaster. For each database in the cluster, PostgreSQL keeps track of the time of the last database-wide VACUUM. When any database approaches the billion-transaction danger level, the system begins to emit warning messages. If nothing is done, it will eventually shut down normal operations until appropriate manual maintenance is done. The remainder of this section gives the details.

The new approach to XID comparison distinguishes two special XIDs, numbers 1 and 2 (BootstrapXID and FrozenXID). These two XIDs are always considered older than every normal XID. Normal XIDs (those greater than 2) are compared using modulo-2³¹ arithmetic. This means that for every normal XID, there are two billion XIDs that are "older" and two billion that are "newer"; another way to say it is that the normal XID space is circular with no endpoint. Therefore, once a row version has been created with a particular normal XID, the row version will appear to be "in the past" for the next two billion transactions, no matter which normal XID we are talking about. If the row version still exists after more than two billion transactions, it will suddenly appear to be in the future. To prevent data loss, old row versions must be reassigned the XID FrozenXID sometime before they reach the two-billion-transactions-old mark. Once they are assigned this special XID, they will appear to be "in the past" to all normal transactions regardless of wraparound issues, and so such row versions will be good until deleted, no matter how long that is. This reassignment of XID is handled by VACUUM.

VACUUM's normal policy is to reassign FrozenXID to any row version with a normal XID more than one billion transactions in the past. This policy preserves the original insertion XID until it is not likely to be of interest anymore. (In fact, most row versions will probably live and die without ever being "frozen".) With this policy, the maximum safe interval between VACUUM runs on any table is exactly one billion transactions: if you wait longer, it's possible that a row version that was not quite old enough to be reassigned last time is now more than two billion transactions old and has wrapped around into the future — i.e., is lost to you. (Of course, it'll reappear after another two billion transactions, but that's no help.)

Since periodic VACUUM runs are needed anyway for the reasons described earlier, it's unlikely that any table would not be vacuumed for as long as a billion transactions. But to help administrators ensure this constraint is met, VACUUM stores transaction ID statistics in the system table pg_database. In particular, the datfrozenxid column of a database's pg_database row is updated at the completion of any database-wide VACUUM operation (i.e., VACUUM that does not name a specific table). The value stored in this field is the freeze cutoff XID that was used by that VACUUM command. All normal XIDs older than this cutoff XID are guaranteed to have been replaced by FrozenXID within that database. A convenient way to examine this information is to execute the query

SELECT datname, age(datfrozenxid) FROM pg_database;

The age column measures the number of transactions from the cutoff XID to the current transaction's XID.

With the standard freezing policy, the age column will start at one billion for a freshly-vacuumed database. When the age approaches two billion, the database must be vacuumed again to avoid risk of wraparound failures. Recommended practice is to VACUUM each database at least once every half-a-billion (500 million) transactions, so as to provide plenty of safety margin. To help meet this rule, each database-wide VACUUM automatically delivers a warning if there are any pg_database entries showing an age of more than 1.5 billion transactions, for example:

play=# VACUUM;
WARNING:  database "mydb" must be vacuumed within 177009986 transactions
HINT:  To avoid a database shutdown, execute a full-database VACUUM in "mydb".
VACUUM

If the warnings emitted by VACUUM go ignored, then PostgreSQL will begin to emit a warning like the above on every transaction start once there are fewer than 10 million transactions left until wraparound. If those warnings also are ignored, the system will shut down and refuse to execute any new transactions once there are fewer than 1 million transactions left until wraparound:

play=# select 2+2;
ERROR:  database is shut down to avoid wraparound data loss in database "mydb"
HINT:  Stop the postmaster and use a standalone backend to VACUUM in "mydb".

The 1-million-transaction safety margin exists to let the administrator recover without data loss, by manually executing the required VACUUM commands. However, since the system will not execute commands once it has gone into the safety shutdown mode, the only way to do this is to stop the postmaster and use a standalone backend to execute VACUUM. The shutdown mode is not enforced by a standalone backend. See the postgres reference page for details about using a standalone backend.

VACUUM with the FREEZE option uses a more aggressive freezing policy: row versions are frozen if they are old enough to be considered good by all open transactions. In particular, if a VACUUM FREEZE is performed in an otherwise-idle database, it is guaranteed that all row versions in that database will be frozen. Hence, as long as the database is not modified in any way, it will not need subsequent vacuuming to avoid transaction ID wraparound problems. This technique is used by initdb to prepare the template0 database. It should also be used to prepare any user-created databases that are to be marked datallowconn = false in pg_database, since there isn't any convenient way to VACUUM a database that you can't connect to.

22.1.4. The auto-vacuum daemon

Beginning in PostgreSQL 8.1, there is a separate optional server process called the autovacuum daemon, whose purpose is to automate the execution of VACUUM and ANALYZE commands. When enabled, the autovacuum daemon runs periodically and checks for tables that have had a large number of inserted, updated or deleted tuples. These checks use the row-level statistics collection facility; therefore, the autovacuum daemon cannot be used unless stats_start_collector and stats_row_level are set to true. Also, it's important to allow a slot for the autovacuum process when choosing the value of superuser_reserved_connections.

The autovacuum daemon, when enabled, runs every autovacuum_naptime seconds and determines which database to process. Any database which is close to transaction ID wraparound is immediately processed. In this case, autovacuum issues a database-wide VACUUM call, or VACUUM FREEZE if it's a template database, and then terminates. If no database fulfills this criterion, the one that was least recently processed by autovacuum is chosen. In this case each table in the selected database is checked, and individual VACUUM or ANALYZE commands are issued as needed.

For each table, two conditions are used to determine which operation(s) to apply. If the number of obsolete tuples since the last VACUUM exceeds the "vacuum threshold", the table is vacuumed. The vacuum threshold is defined as:

vacuum threshold = vacuum base threshold + vacuum scale factor * number of tuples

where the vacuum base threshold is autovacuum_vacuum_threshold, the vacuum scale factor is autovacuum_vacuum_scale_factor, and the number of tuples is pg_class.reltuples. The number of obsolete tuples is obtained from the statistics collector; it is a semi-accurate count updated by each UPDATE and DELETE operation. (It is only semi-accurate because some information may be lost under heavy load.) For analyze, a similar condition is used: the threshold, defined as

analyze threshold = analyze base threshold + analyze scale factor * number of tuples

is compared to the total number of tuples inserted, updated, or deleted since the last ANALYZE.

The default thresholds and scale factors are taken from postgresql.conf, but it is possible to override them on a table-by-table basis by making entries in the system catalog pg_autovacuum . If a pg_autovacuum row exists for a particular table, the settings it specifies are applied; otherwise the global settings are used. See Section 17.9 for more details on the global settings.

Besides the base threshold values and scale factors, there are three more parameters that can be set for each table in pg_autovacuum. The first, pg_autovacuum.enabled, can be set to false to instruct the autovacuum daemon to skip that particular table entirely. In this case autovacuum will only touch the table when it vacuums the entire database to prevent transaction ID wraparound. The other two parameters, the vacuum cost delay (pg_autovacuum.vac_cost_delay) and the vacuum cost limit (pg_autovacuum.vac_cost_limit), are used to set table-specific values for the Cost-Based Vacuum Delay feature.

If any of the values in pg_autovacuum are set to a negative number, or if a row is not present at all in pg_autovacuum for any particular table, the corresponding values from postgresql.conf are used.

There is not currently any support for making pg_autovacuum entries, except by doing manual INSERTs into the catalog. This feature will be improved in future releases, and it is likely that the catalog definition will change.