Find and fix inconsistent B-tree indexes

Inconsistencies in database indexes can occur for a variety of reasons including software defects, hardware issues, or underlying changes in behavior such as sort order changes.

The PostgreSQL community has built tools to identify and remediate such issues. This includes tools like amcheck, which is recommended by the PostgreSQL community to identify consistency issues, including issues that earlier versions of PostgreSQL 14 exhibited.

We have written this playbook as a reference for AlloyDB for PostgreSQL who experience such issues. We hope this page provides information that may also aid other PostgreSQL users in identifying and remediating inconsistent B-tree indexes. Our goal is to continually improve this document as a resource for the broader open source community. If you have any feedback, please use the Send feedback button at the top and bottom of this page.

Resolving an index's inconsistencies involves the following steps:

  1. Before you begin.

    Before you begin reindexing, you should back up your database, set the correct permissions, verify your psql client version, and enable the amcheck extension.

  2. Check for inconsistent B-tree indexes.

    To identify the indexes you need to fix inconsistencies for, you need to identify all B-tree indexes with inconsistencies and identify all unique and primary key violations.

  3. Fix the index's inconsistencies.

    Reindexing an index fixes all its inconsistencies. You may need to adjust your instance's memory settings to improve performance.

  4. Monitor reindexing operations.

    We recommend that you monitor the progress of the reindexing operation to ensure that the operation is progressing and is not blocked.

  5. Verify that the indexes are consistent.

    After you have successfully reindexed your index, we recommend that you verify that your index does not contain any inconsistencies.

Before you begin

Back up your AlloyDB cluster's data

To ensure that no data is lost during reindexing, we recommend that you back up your cluster's data. For more information, see Create an on-demand backup.

Set the alloydbsuperuser permission

To complete the steps on this page, you must have alloydbsuperuser permissions. For more information see AlloyDB's predefined PostgreSQL roles.

Ensure that the psql client version is 9.6 or higher

To complete the steps on this page, you must ensure that your psql client version is 9.6 or higher. Run the command psql --version to verify your current psql client version.

Enable the amcheck extension

To check for index inconsistencies, you must enable the amcheck extension.

Code Sample

  CREATE EXTENSION amcheck;

Check for inconsistent B-tree indexes

The following sections describe how to check for inconsistent B-tree indexes by checking for an index's inconsistencies as well as unique and primary key violations.

Check for inconsistencies

Run the following statement in each of your databases to check for inconsistencies in all B-tree indexes:

Code Sample

DO $$
DECLARE
  r RECORD;
  version varchar(100);
BEGIN
  RAISE NOTICE 'Started amcheck on database: %', current_database();
  SHOW server_version into version;
  SELECT split_part(version, '.', 1) into version;
  FOR r IN
    SELECT c.oid, c.oid::regclass relname, i.indisunique
      FROM pg_index i
      JOIN pg_opclass op ON i.indclass[0] = op.oid
      JOIN pg_am am ON op.opcmethod = am.oid
      JOIN pg_class c ON i.indexrelid = c.oid
      JOIN pg_namespace n ON c.relnamespace = n.oid
     WHERE am.amname = 'btree'
       AND c.relpersistence != 't'
       AND c.relkind = 'i'
       AND i.indisready AND i.indisvalid LOOP
    BEGIN
      RAISE NOTICE 'Checking index %:', r.relname;
      PERFORM bt_index_check(index => r.oid, heapallindexed => r.indisunique);
    EXCEPTION
      WHEN undefined_function THEN
        RAISE EXCEPTION 'Failed to find the amcheck extension';
      WHEN OTHERS THEN
        RAISE LOG 'Failed to check index %: %', r.relname, sqlerrm;
        RAISE WARNING 'Failed to check index %: %', r.relname, sqlerrm;
        END;
  END LOOP;
  RAISE NOTICE 'Finished amcheck on database: %', current_database();
END $$;

You should receive output similar to the following:

Code Sample

  NOTICE:  Checking index t_pkey:
  NOTICE:  Checking index t_i_key:
  WARNING:  Failed to check index t_i_key: item order invariant violated for index "t_i_key"
  NOTICE:  Checking index t_j_key:
  WARNING:  Failed to check index t_j_key: item order invariant violated for index "t_j_key"
  NOTICE:  Checking index ij:
  WARNING:  Failed to check index ij: item order invariant violated for index "ij"

For more information about viewing logs, see View logs by using the Logs Explorer.

Identify and fix unique and primary key violations

This section describes how to check your index for unique and primary key violations, and if some exist, how you fix them.

Identify unique key violations

Unique key violations must be fixed before you reindex an index. To check for all unique key violations, run the following command in each database:

Code Sample

WITH q AS (
    /* this gets info for all UNIQUE indexes */
    SELECT indexrelid::regclass as idxname,
           indrelid::regclass as tblname,
           indcollation,
           pg_get_indexdef(indexrelid),
           format('(%s)',(select string_agg(quote_ident(attname), ', ')
              from pg_attribute a
              join unnest(indkey) ia(nr) on ia.nr = a.attnum
             where attrelid = indrelid)) as idxfields,
           COALESCE(substring(pg_get_indexdef(indexrelid) FROM '[)] (WHERE .*)$'), '') as whereclause
      FROM pg_index
     WHERE indisunique
     /* next line excludes indexes not affected by collation changes */
       AND trim(replace(indcollation::text, '0', '')) != ''
)
SELECT
 /* the format constructs the query to execute for each index */
 format(
$sql$
DO $$ BEGIN RAISE NOTICE 'checking index %3$I%2$I on table %1$I %4$I'; END;$$;
SELECT this,
       prev,
       /* we detect both reversed ordering or just not unique */
       (CASE WHEN this = prev THEN 'DUPLICATE' ELSE 'BACKWARDS' END) as violation_type
  FROM (SELECT %2$s AS this,
               lag(%2$s) OVER (ORDER BY %2$s) AS prev
          FROM %1$s %4$s
        ) s
 WHERE this <= prev and this IS NOT NULL and prev IS NOT NULL;  /* change to just '<' if looking for reverse order in index */
$sql$, tblname, idxfields, idxname, whereclause
)
  FROM q
-- LIMIT 20 /* may use limit for testing */
-- the next line tells psql to executes this query and then execute each returned line separately
\gexec

The output of the script is similar to the following:

Output

  NOTICE:  checking index=users_email_key on table=users key_columns="(email)"
  NOTICE:  checking index=games_title_key on table=games  key_columns="(title)"
          this        |        prev        | violation_type
  --------------------+--------------------+----------------
  Game #16 $soccer 2  | Game #16 $soccer 2 | DUPLICATE
  Game #18 $soccer 2  | Game #18 $soccer 2 | DUPLICATE
  Game #2 $soccer 2   | Game #2 $soccer 2  | DUPLICATE
  Game #5 $soccer 2   | Game #5 $soccer 2  | DUPLICATE

In this output, the table header NOTICE shows the index, column, and table for the values displayed below it. If your output contains rows displaying DUPLICATE or BACKWARDS, then this shows corruption in the index and may need to be fixed. Rows with BACKWARDS indicate possible duplicate values that might be hidden. If you see either of these entries in the table, see Fix duplicate key violations.

Fix duplicate key violations

If you have identified a duplicate unique index or if a reindex operation fails due to a duplicate key violation error, complete the following steps to find and remove the duplicate key(s).

  1. Extract the key_columns from the NOTICE table header, as shown in the preceding sample output. In the following example, the key column is email.

    Output

      NOTICE:  checking index=users_email_key on table=users key_columns="(email)"

    Use these values in KEY_COLUMNS in the query in step 3.

  2. Find the schema for your table. Use psql to connect to your database and run the following command:

    Code Sample

    \dt TABLE_NAME
     The value in the schema column is the value you use for SCHEMA_NAME in the query in step 3.

    For example, for the following query:

     \dt games

    The output is similar to the following:

     List of relations
 Schema  | Name  | Type  | Owner
 --------+-------+-------+----------
 public  | games | table | postgres
 (1 row)

  3. Run the following statements to force a full table scan and get duplicate keys.

    Code Sample

    SET enable_indexscan = off;
SET enable_bitmapscan = off;
SET enable_indexonlyscan = off;

SELECT KEY_COLUMNS, count(*)
  FROM SCHEMA_NAME.TABLE_NAME
GROUP BY KEY_COLUMNS
HAVING count(*) > 1;

    In the above statement, KEY_COLUMNS are one or more columns covered by the unique index or primary key in the table you are checking. These were identified when you checked for unique key violations. The statement returns the duplicate keys and a count of the duplicates for each.

    For example, for the following query:

      SELECT name,count(*)
    FROM public.TEST_NAMES
  GROUP BY name
  HAVING count(*) > 1;

    The output is similar to the following:

      name                | count
  --------------------+-------
  Johnny              |     2
  Peter               |     2
(2 rows)

    In this case, continue to the next step to remove the duplicate keys.

    If any of the columns in KEY_COLUMNS are null, you can ignore them because unique constraints do not apply for NULL columns.

    If no duplicate keys are found, you can move to Fix inconsistent indexes.

  4. Optional but recommended: Create a backup for the records containing duplicate keys. Run the following statement to create backup records:

    Code Sample

    CREATE TABLE SCHEMA_NAME.TABLE_NAME_bak
AS SELECT * FROM SCHEMA_NAME.TABLE_NAME
WHERE (KEY_COLUMNS)
IN (KEY_VALUES);
    In this statement, KEY_VALUES is a list of values copied from the result of the previous step. For example:

    Code Sample

      CREATE TABLE public.TEST_NAMES_bak
  AS SELECT * FROM public.TEST_NAMES
  WHERE (name) IN (('Johnny'),('Peter'))

    For a large number of rows, it is easier to replace the ((KEY_VALUES)) parameter in the IN statement with the SELECT statement from step 2 without the count parameter. For example:

    Code Sample

      CREATE TABLE SCHEMA_NAME.TABLE_NAME_bak
  AS SELECT * FROM SCHEMA_NAME.TABLE_NAME
  WHERE (KEY_COLUMNS)
  IN ( SELECT (KEY_COLUMNS)
  FROM SCHEMA_NAME.TABLE_NAME
  GROUP BY (KEY_COLUMNS)
  HAVING count(*) > 1);

  5. Add a replication role to the user to disable triggers:

    Code Sample

    ALTER USER CURRENT_USER with REPLICATION;
SET session_replication_role = replica;

  6. Run the following statement to delete the duplicate keys:

    Code Sample

    BEGIN;

DELETE FROM  SCHEMA_NAME.TABLE_NAME a
      USING  (
              SELECT   min(ctid) AS ctid,
                      KEY_COLUMNS
              FROM     SCHEMA_NAME.TABLE_NAME
              GROUP BY KEY_COLUMNS
                      HAVING count(*) > 1 ) b
      WHERE a.KEY_COLUMNS = b.KEY_COLUMNS
      AND   a.ctid <> b.ctid;

    For example, for multi-column KEY_COLUMNS:

    Code Sample

      DELETE FROM public.test_random a
      USING (
             SELECT min(ctid) AS ctid,
             day, rnum
      FROM public.test_random
      GROUP BY day, rnum
             HAVING count(*) > 1 ) b
      WHERE a.day=b.day and a.rnum = b.rnum
      AND a.ctid <> b.ctid;
     Where day and rnum are KEY_COLUMNS.

    Running this statement keeps one row and deletes others for each set of duplicate rows. If you want to control which version of the row gets deleted, run the following filter in the delete statement:

    Code Sample

    DELETE FROM  SCHEMA_NAME.TABLE_NAME
WHERE ( KEY_COLUMNS, ctid) = (KEY_VALUES, CTID_VALUE);

  7. Complete the following steps to check that the DELETE command returned the expected number of rows without any errors:

    1. Run the following statement to identify the rows in which tables were changed:

      Code Sample

      SELECT schemaname, relname, n_tup_del, n_tup_upd
  FROM pg_stat_xact_all_tables
WHERE n_tup_del+n_tup_upd > 0;

    2. If all the rows are correct, commit the DELETE transaction:

      Code Sample

      END;

    3. If there are errors, roll back the changes to fix the errors:

      Code Sample

      ROLLBACK;

  8. After the duplicate keys are deleted, you can reindex your index.

Fix inconsistent indexes

The following sections describe how you can fix the index inconsistencies found in your instance.

Depending on how your database is configured, you may need to do the following for each index identified in the previous steps:

  1. Prepare to reindex your index.

  2. Reindex your index.

  3. If the reindex operation fails due to foreign key violations, you must find and fix these violations.

  4. Run the reindex operation again.

Prepare to reindex your index

Find the index size

Indexing larger databases requires more time than smaller databases. To improve the speed of index and reindex operations of larger databases, you can allocate more memory and CPU power to these operations. This is an important step in planning your reindex operation. After you know the index size, you can set the memory size used by the reindex operation and set the number of parallel workers.

Run the following statement to find the index size, in kilobytes, of the index that you want to fix:

Code Sample

SELECT i.relname                                      AS index_name,
       pg_size_pretty(pg_relation_size(x.indexrelid)) AS index_size
FROM   pg_index x
       JOIN pg_class i
         ON i.oid = x.indexrelid
WHERE  i.relname = 'INDEX_NAME';

The output of this statement is similar to the following:

Output

 index_name | index_size
------------+------------
 my_index   | 16 kB
(1 row)

Set memory size to use for reindexing

Based on the size of your index as determined in the previous section, it is important to set the appropriate value for the maintenance_work_mem configuration parameter. This parameter specifies the amount of memory to use for the reindexing operation. For example, if your index size is greater than 15 GB, we recommend that you adjust your maintenance memory.

The following example shows how to set maintenance_work_mem:

Code Sample

SET maintenance_work_mem TO "1GB";

Indexing larger databases requires more time than smaller databases. To improve the speed of index and reindex operations, we recommend setting maintenance_work_mem to at least 2% of the instance memory for instances with 4GB or more memory during this reindexing operation.

Set the number of parallel workers

You can increase the number of parallel workers for reindexing by setting the max_parallel_maintenance_workers configuration parameter in the database. The default value of this parameter is 2 but can be set to a higher value to increase the number of workers for reindexing. For instances with 8 or more vCPU cores, we recommend setting the max_parallel_maintenance_workers flag value to 4.

The following shows how to identify the values set for these parameters:

Code Sample

SHOW max_parallel_maintenance_workers;
SHOW max_worker_processes;
SHOW max_parallel_workers;

The max_parallel_maintenance_workers parameter is a subset of max_worker_processes and limited by max_parallel_workers. If you require more parallel workers, increase the value of max_worker_processes and max_parallel_workers.

The following example shows how to set max_parallel_maintenance_workers:

Code Sample

SET max_parallel_maintenance_workers TO 4;

The max_parallel_maintenance_workers parameter does not guarantee the allocation of workers. To confirm that you have multiple parallel workers started by the reindexing, run the following query from another session after you have started the reindex:

Code Sample

SELECT
leader.leader_pid,leader.pid "worker_pid",leader.query
FROM
pg_stat_activity leader, pg_stat_activity worker
WHERE leader.leader_pid = worker.pid;

Reindex your index

You can reindex an index without blocking your production workload using the pg_repack utility. This utility automates and simplifies the concurrent reindex process, enabling you to reindex without downtime. For this procedure, use pg_repack version 1.4.7.

Complete the following steps to reindex your index using pg_repack:

  1. Download, compile, and install the pg_repack utility from the pg_repack page.

  2. Create the pg_repack extension:

    Code Sample

    CREATE EXTENSION pg_repack;

  3. Run the following command to reindex your index concurrently:

    Code Sample

      pg_repack -h HOSTIP -p 5432 -U USERNAME -d "DATABASE_NAME" -i "INDEX_NAME" --no-superuser-check --no-kill-backend --wait-timeout=3600

    This command has output similar to the following:

    Output

    INFO: repacking index "public.t_i_key"

    If any errors occurred when running pg_repack, you can correct the error and try again. After you have fixed all of your unique key indexes and primary key indexes, you should check for foreign key violations and fix any that are found.

Find and fix foreign key violations

For information about how to find and fix foreign key violations, see Find and fix foreign key violations.

Monitor reindexing operations

Occasionally, the reindex operation may be blocked by other sessions. We recommended that you check this every 4 hours. If the reindex operation is blocked, you can cancel the blocking session so the reindex operation can complete.

Complete the following steps to identify blocking and waiting sessions and then cancel them in the INDEX operation:

  1. To identify blocking sessions, run the following query:

    Code Sample

    SELECT pid,
      usename,
      pg_blocking_pids(pid) AS blocked_by,
      query                 AS blocked_query
FROM   pg_stat_activity
WHERE  cardinality(pg_blocking_pids(pid)) > 0;

  2. To cancel a session, run the following query using the PID of the blocking session from the previous query:

    Code Sample

    SELECT pg_cancel_backend(PID);

Verify that your indexes are consistent

You must continue to check for index inconsistencies for each inconsistent index. After you have fixed all your instance's inconsistent indexes and key violations, you can check that no issues exist by following the steps in the previous sections: