Managing Bloat in the Database

Managing Bloat in the Database

Greenplum Database heap tables use the PostgreSQL Multiversion Concurrency Control (MVCC) storage implementation. A deleted or updated row is logically deleted from the database, but a non-visible image of the row remains in the table. These deleted rows, also called expired rows, are tracked in a free space map. Running VACUUM marks the expired rows as free space that is available for reuse by subsequent inserts.

If the free space map is not large enough to accommodate all of the expired rows, the VACUUM command is unable to reclaim space for expired rows that overflowed the free space map. The disk space may only be recovered by running VACUUM FULL, which locks the table, copies rows one-by-one to the beginning of the file, and truncates the file. This is an expensive operation that can take an exceptional amount of time to complete with a large table. It should be used only on smaller tables. If you attempt to kill a VACUUM FULL operation, the system can be disrupted.

Important:

It is very important to run VACUUM after large UPDATE and DELETE operations to avoid the necessity of ever running VACUUM FULL.

If the free space map overflows and it is necessary to recover the space it is recommended to use the CREATE TABLE...AS SELECT command to copy the table to a new table, which will create a new compact table. Then drop the original table and rename the copied table.

It is normal for tables that have frequent updates to have a small or moderate amount of expired rows and free space that will be reused as new data is added. But when the table is allowed to grow so large that active data occupies just a small fraction of the space, the table has become significantly "bloated." Bloated tables require more disk storage and additional I/O that can slow down query execution.

Bloat affects heap tables, system catalogs, and indexes.

Running the VACUUM statement on tables regularly prevents them from growing too large. If the table does become significantly bloated, the VACUUM FULL statement (or an alternative procedure) must be used to compact the file. If a large table becomes significantly bloated, it is better to use one of the alternative methods described in Removing Bloat from Database Tables to remove the bloat.

CAUTION:
Never run VACUUM FULL <database_name> and do not run VACUUM FULL on large tables in a Greenplum Database.

Sizing the Free Space Map

Expired rows in heap tables are added to a shared free space map when you run VACUUM. The free space map must be adequately sized to accommodate these rows. If the free space map is not large enough, any space occupied by rows that overflow the free space map cannot be reclaimed by a regular VACUUM statement. You will have to use VACUUM FULL or an alternative method to recover the space.

You can avoid overflowing the free space map by running the VACUUM statement regularly. The more bloated a table becomes, the more rows that must be tracked in the free space map. For very large databases with many objects, you may need to increase the size of the free space map to prevent overflow.

The max_fsm_pages configuration parameter sets the maximum number of disk pages for which free space will be tracked in the shared free-space map. Each page slot consumes six bytes of shared memory. The default value for max_fsm_pages is 200,000.

The max_fsm_relations configuration parameter sets the maximum number of relations for which free space will be tracked in the shared memory free-space map. It should be set to a value larger than the total number of tables, indexes, and system tables in the database. It costs about 60 bytes of memory for each relation per segment instance. The default value is 1000.

See the Greenplum Database Reference Guide for detailed information about these configuration parameters.

Detecting Bloat

The statistics collected by the ANALYZE statement can be used to calculate the expected number of disk pages required to store a table. The difference between the expected number of pages and the actual number of pages is a measure of bloat. The gp_toolkit schema provides a gp_bloat_diag view that identifies table bloat by comparing the ratio of expected to actual pages. To use it, make sure statistics are up to date for all of the tables in the database, then run the following SQL:
gpadmin=# SELECT * FROM gp_toolkit.gp_bloat_diag;
 bdirelid | bdinspname | bdirelname | bdirelpages | bdiexppages |                bdidiag                
----------+------------+------------+-------------+-------------+---------------------------------------
    21488 | public     | t1         |          97 |           1 | significant amount of bloat suspected
(1 row)

The results include only tables with moderate or significant bloat. Moderate bloat is reported when the ratio of actual to expected pages is greater than four and less than ten. Significant bloat is reported when the ratio is greater than ten.

The gp_toolkit.gp_bloat_expected_pages view lists the actual number of used pages and expected number of used pages for each database object.
gpadmin=# SELECT * FROM gp_toolkit.gp_bloat_expected_pages LIMIT 5;
 btdrelid | btdrelpages | btdexppages 
----------+-------------+-------------
    10789 |           1 |           1
    10794 |           1 |           1
    10799 |           1 |           1
     5004 |           1 |           1
     7175 |           1 |           1
(5 rows)

The btdrelid is the object ID of the table. The btdrelpages column reports the number of pages the table uses; the btdexppages column is the number of pages expected. Again, the numbers reported are based on the table statistics, so be sure to run ANALYZE on tables that have changed.

Removing Bloat from Database Tables

The VACUUM command adds expired rows to the shared free space map so that the space can be reused. When VACUUM is run regularly on a table that is frequently updated, the space occupied by the expired rows can be promptly reused, preventing the table file from growing larger. It is also important to run VACUUM before the free space map is filled. For heavily updated tables, you may need to run VACUUM at least once a day to prevent the table from becoming bloated.

Warning: When a table is significantly bloated, it is better to run ANALYZE before running VACUUM. Because ANALYZE uses block-level sampling, a table with a high ratio of blocks containing no valid rows can cause ANALYZE to set the reltuples column of the pg_class system catalog to an inaccurate value or 0, which can lead to poorly optimized queries. The VACUUM command produces a more accurate count and when run after ANALYZE will correct an inaccurate row count estimate.

When a table accumulates significant bloat, running the VACUUM command is insufficient. For small tables, running VACUUM FULL <table_name> can reclaim space used by rows that overflowed the free space map and reduce the size of the table file. However, a VACUUM FULL statement is an expensive operation that requires an ACCESS EXCLUSIVE lock and may take an exceptionally long and unpredictable amount of time to finish. Rather than run VACUUM FULL on a large table, an alternative method is required to remove bloat from a large file. Note that every method for removing bloat from large tables is resource intensive and should be done only under extreme circumstances.

The first method to remove bloat from a large table is to create a copy of the table excluding the expired rows, drop the original table, and rename the copy. This method uses the CREATE TABLE <table_name> AS SELECT statement to create the new table, for example:
gpadmin=# CREATE TABLE mytable_tmp AS SELECT * FROM mytable;
gpadmin=# DROP TABLE mytable;
gpadmin=# ALTER TABLE mytabe_tmp RENAME TO mytable;
A second way to remove bloat from a table is to redistribute the table, which rebuilds the table without the expired rows. Follow these steps:
  1. Make a note of the table's distribution columns.
  2. Change the table's distribution policy to random:
    ALTER TABLE mytable SET WITH (REORGANIZE=false) 
    DISTRIBUTED randomly;

    This changes the distribution policy for the table, but does not move any data. The command should complete instantly.

  3. Change the distribution policy back to its initial setting:
    ALTER TABLE mytable SET WITH (REORGANIZE=true) 
    DISTRIBUTED BY (<original distribution columns>);

    This step redistributes the data. Since the table was previously distributed with the same distribution key, the rows are simply rewritten on the same segment, excluding expired rows.

Removing Bloat from Indexes

The VACUUM command only recovers space from tables. To recover the space from indexes, recreate them using the REINDEX command.

To rebuild all indexes on a table run REINDEX table_name;. To rebuild a particular index, run REINDEX index_name;. REINDEX does not update the reltuples and relpages statistics for the index, so it is important to ANALYZE the table to update these statistics after reindexing.

Removing Bloat from System Catalogs

Greenplum Database system catalogs are also heap tables and can become bloated over time. As database objects are created, altered, or dropped, expired rows are left in the system catalogs. Using gpload to load data contributes to the bloat since gpload creates and drops external tables. (Rather than use gpload, it is recommended to use gpfdist to load data.)

Bloat in the system catalogs increases the time require to scan the tables, for example, when creating explain plans. System catalogs are scanned frequently and if they become bloated, overall system performance is degraded.

It is recommended to run VACUUM on the system catalog nightly and at least weekly. At the same time, running REINDEX SYSTEM removes bloat from the indexes. Alternatively, you can reindex system tables using the reindexdb utility with the -s (--system) option. After reindexing, it is also important to run ANALYZE, because the REINDEX command rebuilds indexes with empty statistics.

The following script runs VACUUM, REINDEX, and ANALYZE on the system catalogs.
#!/bin/bash
DBNAME="<database_name>"
SYSTABLES="' pg_catalog.' || relname || ';' from pg_class a, pg_namespace b \
where a.relnamespace=b.oid and b.nspname='pg_catalog' and a.relkind='r'"
psql -tc "SELECT 'VACUUM' || $SYSTABLES" $DBNAME | psql -a $DBNAME
reindexdb -s -d $DBNAME
analyzedb -s pg_catalog -d $DBNAME

If the system catalogs become significantly bloated, you must perform an intensive system catalog maintenance procedure. The CREATE TABLE AS SELECT and redistribution key methods for removing bloat cannot be used with system catalogs. You must instead run VACUUM FULL during a scheduled downtime period. During this period, stop all catalog activity on the system; VACUUM FULL takes exclusive locks against the system catalog. Running VACUUM regularly can prevent the need for this more costly procedure.

Removing Bloat from Append-Optimized Tables

Append-optimized tables are handled much differently than heap tables. Although append-optimized tables allow updates, inserts, and deletes, they are not optimized for these operations and it is recommended to not use them with append-optimized tables. If you heed this advice and use append-optimized for load-once/read-many workloads, VACUUM on an append-optimized table runs almost instantaneously.

If you do run UPDATE or DELETE commands on an append-optimized table, expired rows are tracked in an auxiliary bitmap instead of the free space map. VACUUM is the only way to recover the space. Running VACUUM on an append-optimized table with expired rows compacts a table by rewriting the entire table without the expired rows. However, no action is performed if the percentage of expired rows in the table exceeds the value of the gp_appendonly_compaction_threshold configuration parameter, which is 10 (10%) by default. The threshold is checked on each segment, so it is possible that a VACUUM statement will compact an append-only table on some segments and not others. Compacting append-only tables can be disabled by setting the gp_appendonly_compaction parameter to no.