USING INDEX¶

USING INDEX clause should be specified after WHERE clause of SELECT, DELETE or UPDATE statement. USING INDEX clause forces a sequential/index scan to be used or an index that can improve the performance to be included.

If USING INDEX clause is specified with the list of index names, query optimizer creates optimized execution plan by calculating the query execution cost based on the specified indexes only and comparing the index scan cost and the sequential scan cost of the specified indexes(CUBRID performs cost-based query optimization to select an execution plan).

The USING INDEX clause is useful to get the results in the desired order without ORDER BY. When index scan is performed by CUBRID, the results are created in the order they were saved in the index. When there are more than one indexes in one table, you can use USING INDEX to get the query results in a given order of indexes.

SELECT ... WHERE ...
[USING INDEX { NONE | [ ALL EXCEPT ] <index_spec> [ {, <index_spec> } ...] } ] [ ; ]

DELETE ... WHERE ...
[USING INDEX { NONE | [ ALL EXCEPT ] <index_spec> [ {, <index_spec> } ...] } ] [ ; ]

UPDATE ... WHERE ...
[USING INDEX { NONE | [ ALL EXCEPT ] <index_spec> [ {, <index_spec> } ...] } ] [ ; ]

<index_spec> ::=
  [table_spec.]index_name [(+) | (-)] |
  table_spec.NONE

• NONE: If NONE is specified, a sequential scan is used on all tables.
• ALL EXCEPT: All indexes except the specified indexes can be used when the query is executed.
• index_name(+): If (+) is specified after the index_name, it is the first priority in index selection. IF this index is not proper to run the query, it is not selected.
• index_name(-): If (-) is specified after the index_name, it is excluded from index selection.
• table_spec.NONE: All indexes are excluded from the selection, so sequential scan is used.

USE, FORCE, IGNORE INDEX¶

Index hints can be specified through USE, FORCE, IGNORE INDEX syntax after table specification of FROM clause.

FROM table_spec [ <index_hint_clause> ] ...

<index_hint_clause> ::=
  { USE | FORCE | IGNORE } INDEX  ( <index_spec> [, <index_spec>  ...] )

<index_spec> ::=
  [table_spec.]index_name

• USE INDEX ( <index_spec> ): Only specified indexes are considered when choose them.
• FORCE INDEX ( <index_spec> ): Specified indexes are chosen as the first priority.
• IGNORE INDEX ( <index_spec> ): Specified indexes are excluded from the choice.

USE, FORCE, IGNORE INDEX syntax is automatically rewritten as the proper USING INDEX syntax by the system.

Examples of index hint¶

CREATE TABLE athlete2 (
   code             SMALLINT PRIMARY KEY,
   name             VARCHAR(40) NOT NULL,
   gender           CHAR(1),
   nation_code      CHAR(3),
   event            VARCHAR(30)
);
CREATE UNIQUE INDEX athlete2_idx1 ON athlete2 (code, nation_code);
CREATE INDEX athlete2_idx2 ON athlete2 (gender, nation_code);

Below two queries do the same behavior and they select index scan if the specified index, athlete2_idx2's scan cost is lower than sequential scan cost.

SELECT /*+ RECOMPILE */ *
FROM athlete2 USE INDEX (athlete2_idx2)
WHERE gender='M' AND nation_code='USA';

SELECT /*+ RECOMPILE */ *
FROM athlete2
WHERE gender='M' AND nation_code='USA'
USING INDEX athlete2_idx2;

Below two queries do the same behavior and they always use athlete2_idx2

SELECT /*+ RECOMPILE */ *
FROM athlete2 FORCE INDEX (athlete2_idx2)
WHERE gender='M' AND nation_code='USA';

SELECT /*+ RECOMPILE */ *
FROM athlete2
WHERE gender='M' AND nation_code='USA'
USING INDEX athlete2_idx2(+);

Below two queries do the same behavior and they always don't use athlete2_idx2

SELECT /*+ RECOMPILE */ *
FROM athlete2 IGNORE INDEX (athlete2_idx2)
WHERE gender='M' AND nation_code='USA';

SELECT /*+ RECOMPILE */ *
FROM athlete2
WHERE gender='M' AND nation_code='USA'
USING INDEX athlete2_idx2(-);

Below query always do the sequential scan.

SELECT *
FROM athlete2
WHERE gender='M' AND nation_code='USA'
USING INDEX NONE;

SELECT *
FROM athlete2
WHERE gender='M' AND nation_code='USA'
USING INDEX athlete2.NONE;

Below query forces to be possible to use all indexes except athlete2_idx2 index.

SELECT *
FROM athlete2
WHERE gender='M' AND nation_code='USA'
USING INDEX ALL EXCEPT athlete2_idx2;

When two or more indexes have been specified in the USING INDEX clause, the query optimizer selects the proper one of the specified indexes.

SELECT *
FROM athlete2 USE INDEX (athlete2_idx2, athlete2_idx1)
WHERE gender='M' AND nation_code='USA';

SELECT *
FROM athlete2
WHERE gender='M' AND nation_code='USA'
USING INDEX athlete2_idx2, athlete2_idx1;

When a query is run for several tables, you can specify a table to perform index scan by using a specific index and another table to perform sequential scan. The query has the following format.

SELECT *
FROM tab1, tab2
WHERE ...
USING INDEX tab1.idx1, tab2.NONE;

When executing a query with the index hint syntax, the query optimizer considers all available indexes on the table for which no index has been specified. For example, when the tab1 table includes idx1 and idx2 and the tab2 table includes idx3, idx4, and idx5, if indexes for only tab1 are specified but no indexes are specified for tab2, the query optimizer considers the indexes of tab2.

SELECT ...
FROM tab1, tab2 USE INDEX(tab1.idx1)
WHERE ... ;

SELECT ...
FROM tab1, tab2
WHERE ...
USING INDEX tab1.idx1;

The above query select the scan method of table tab1 after comparing the cost between the sequential scan of the table tab1 and the index scan of the index idx1, and select the scan method of table tab2 after comparing the cost between the sequential scan of the table tab2 and the index scan of the indexes idx3, idx4, idx5.

Filtered Index¶

The filtered index is used to sort, search, or operate a well-defined partials set for one table. It is called the partial index since only some data that satisfy the condition are kept in that index.

CREATE /*+ hints */ INDEX index_name
ON table_name (col1, col2, ...)
WHERE <filter_predicate>;

ALTER  /*+ hints */ INDEX index_name
[ ON table_name (col1, col2, ...)
[ WHERE <filter_predicate> ] ]
REBUILD;

<filter_predicate> ::= <filter_predicate> AND <expression> | <expression>

• <filter_predicate>: Condition to compare the column and the constant. When there are several conditions, filtering is available only when they are connected by using AND. The filter conditions can include most of the operators and functions supported by CUBRID. However, the date/time function that shows the current date/time (ex: SYS_DATETIME()) or random functions (ex: RAND()), which outputs different results for one input are not allowed.

If you want to apply the filtered index, that filtered index must be specified by USE INDEX syntax or FORCE INDEX syntax.

• When a filtered index is specified by USING INDEX clause or USE INDEX syntax:

  If columns of which the index consists are not included on the conditions of WHERE clause, the filtered index is not used.

  CREATE TABLE blogtopic
  (
      blogID BIGINT NOT NULL,
      title VARCHAR(128),
      author VARCHAR(128),
      content VARCHAR(8096),
      postDate TIMESTAMP NOT NULL,
      deleted SMALLINT DEFAULT 0
  );
  
  CREATE INDEX my_filter_index ON blogtopic(postDate) WHERE deleted=0;
  

  On the below query, postDate, a column of which my_filter_index consists, is included on the conditions of WHERE condition. Therefore, this index can be used by "USE INDEX" syntax.

  SELECT *
  FROM blogtopic USE INDEX (my_filter_index)
  WHERE postDate>'2010-01-01' AND deleted=0;
  

• When a filtered index is specified by USING INDEX <index_name>(+) clause or FORCE INDEX syntax:

  Even if a column of which the index consists is not included on the condition of WHERE clause, the filtered index is used.

  On the below query, my_filter_index cannot be used by "USE INDEX" syntax because a column of which my_filter_index consists is not included on the WHERE condition.

  SELECT *
  FROM blogtopic USE INDEX (my_filter_index)
  WHERE author = 'David' AND deleted=0;
  

  Therefore, to use my_filter_index, it should be forced by "FORCE INDEX".

  SELECT *
  FROM blogtopic FORCE INDEX (my_filter_index)
  WHERE author = 'David' AND deleted=0;
  

The following example shows a bug tracking system that maintains bugs/issues. After a specified period of development, the bugs table records bugs. Most of the bugs have already been closed. The bug tracking system makes queries on the table to find new open bugs. In this case, the indexes on the bug table do not need to know the records on closed bugs. Then the filtered indexes allow indexing of open bugs only.

CREATE TABLE bugs
(
    bugID BIGINT NOT NULL,
    CreationDate TIMESTAMP,
    Author VARCHAR(255),
    Subject VARCHAR(255),
    Description VARCHAR(255),
    CurrentStatus INTEGER,
    Closed SMALLINT
);

Indexes for open bugs can be created by using the following sentence:

CREATE INDEX idx_open_bugs ON bugs(bugID) WHERE Closed = 0;

To process queries that are interested in open bugs, specify the index as an index hint. It will allow creating query results by accessing less index pages through filtered indexes.

SELECT *
FROM bugs
WHERE Author = 'madden' AND Subject LIKE '%fopen%' AND Closed = 0
USING INDEX idx_open_bugs(+);

SELECT *
FROM bugs FORCE INDEX (idx_open_bugs)
WHERE CreationDate > CURRENT_DATE - 10 AND Closed = 0;

On the above example, if you use "USING INDEX idx_open_bugs" clause or "USE INDEX (idx_open_bugs)" syntax, a query is processed without using the idx_open_bugs index.

CREATE INDEX idx_open_bugs ON bugs (Author) WHERE Closed = 0;

ERROR: before ' ; '
Invalid filter expression (bugs.Closed=0) for index.

CREATE INDEX idx_open_bugs ON bugs (Author, CreationDate) WHERE Closed = 0;

Function-based Index¶

Function-based index is used to sort or find the data based on the combination of values of table rows by using a specific function. For example, to find the space-ignored string, it can be used to optimize the query by using the function that provides the feature. In addition, it is useful to search the non-case-sensitive names.

CREATE /*+ hints */ INDEX index_name
ON table_name (function_name (argument_list));

ALTER /*+ hints */ INDEX index_name
[ ON table_name (function_name (argument_list)) ]
REBUILD;

After the following indexes have been created, the SELECT query automatically uses the function-based index.

CREATE INDEX idx_trim_post ON posts_table(TRIM(keyword));

SELECT *
FROM posts_table
WHERE TRIM(keyword) = 'SQL';

If a function-based index is created by using the LOWER function, it can be used to search the non-case-sensitive names.

CREATE INDEX idx_last_name_lower ON clients_table(LOWER(LastName));

SELECT *
FROM clients_table
WHERE LOWER(LastName) = LOWER('Timothy');

To make an index selected while creating a query plan, the function used for the index should be used for the query condition in the same way. The SELECT query above uses the last_name_lower index created above. However, this index is not used for the following condition:

SELECT *
FROM clients_table
WHERE LOWER(CONCAT('Mr. ', LastName)) = LOWER('Mr. Timothy');

In addition, to make the function-based index used by force, use the USING INDEX syntax.

CREATE INDEX i_tbl_first_four ON tbl(LEFT(col, 4));
SELECT *
FROM clients_table
WHERE LEFT(col, 4) = 'CAT5'
USING INDEX i_tbl_first_four;

Functions with the function-based indexes are as follows:

ABS	ACOS	ADD_MONTHS	ADDDATE	ASIN
ATAN	ATAN2	BIT_COUNT	BIT_LENGTH	CEIL
CHAR_LENGTH	CHR	COS	COT	DATE
DATE_ADD	DATE_FORMAT	DATE_SUB	DATEDIFF	DAY
DAYOFMONTH	DAYOFWEEK	DAYOFYEAR	DEGREES	EXP
FLOOR	FORMAT	FROM_DAYS	FROM_UNIXTIME	GREATEST
HOUR	IFNULL	INET_ATON	INET_NTOA	INSTR
LAST_DAY	LEAST	LEFT	LN	LOCATE
LOG	LOG10	LOG2	LOWER	LPAD
LTRIM	MAKEDATE	MAKETIME	MD5	MID
MINUTE	MOD	MONTH	MONTHS_BETWEEN	NULLIF
NVL	NVL2	OCTET_LENGTH	POSITION	POWER
QUARTER	RADIANS	REPEAT	REPLACE	REVERSE
RIGHT	ROUND	RPAD	RTRIM	SECOND
SECTOTIME	SIN	SQRT	STR_TO_DATE	STRCMP
SUBDATE	SUBSTR	SUBSTRING	SUBSTRING_INDEX	TAN
TIME	TIME_FORMAT	TIMEDIFF	TIMESTAMP	TIMETOSEC
TO_CHAR	TO_DATE	TO_DATETIME	TO_DAYS	TO_NUMBER
TO_TIME	TO_TIMESTAMP	TRANSLATE	TRIM	TRUNC
UNIX_TIMESTAMP	UPPER	WEEK	WEEKDAY	YEAR

Arguments of functions which can be used in the function-based indexes, only column names and constants are allowed; nested expressions are not allowed. For example, a statement below will cause an error.

CREATE INDEX my_idx ON tbl (TRIM(LEFT(col, 3)));
CREATE INDEX my_idx ON tbl (LEFT(col1, col2 + 3));

However, implicit cast is allowed. In the example below, the first argument type of the LEFT () function should be VARCHAR and the second argument type should be INTEGER; it works normally.

CREATE INDEX my_idx ON tbl (LEFT(int_col, str_col));

Function-based indexes cannot be used with filtered indexes. The example will cause an error.

CREATE INDEX my_idx ON tbl (TRIM(col)) WHERE col > 'SQL';

Function-based indexes cannot become multiple-columns indexes. The example will cause an error.

CREATE INDEX my_idx ON tbl (TRIM(col1), col2, LEFT(col3, 5));

Covering Index¶

The covering index is the index including the data of all columns in the SELECT list and the WHERE, HAVING, GROUP BY, and ORDER BY clauses.

You only need to scan the index pages, as the covering index contains all the data necessary for executing a query, and it also reduces the I/O costs as it is not necessary to scan the data storage any further. To increase data search speed, you can consider creating a covering index but you should be aware that the INSERT and the DELETE processes may be slowed down due to the increase in index size.

The rules about the applicability of the covering index are as follows:

• If the covering index is applicable, you should use the CUBRID query optimizer first.
• For the join query, if the index includes columns of the table in the SELECT list, use this index.
• You cannot use the covering index if an index cannot be used.

CREATE TABLE t (col1 INT, col2 INT, col3 INT);
CREATE INDEX i_t_col1_col2_col3 ON t (col1,col2,col3);
INSERT INTO t VALUES (1,2,3),(4,5,6),(10,8,9);

The following example shows that the index is used as a covering index because columns of both SELECT and WHERE condition exist within the index.

-- csql> ;plan simple
SELECT * FROM t WHERE col1 < 6;

Query plan:
 Index scan(t t, i_t_col1_col2_col3, [(t.col1 range (min inf_lt t.col3))] (covers))

         col1         col2         col3
=======================================
            1            2            3
            4            5            6

CREATE TABLE tab(c VARCHAR(32));
INSERT INTO tab VALUES('abcd'),('abcd    '),('abcd ');
CREATE INDEX i_tab_c ON tab(c);

-- csql>;plan simple
SELECT * FROM tab WHERE c='abcd    ' USING INDEX i_tab_c(+);

Query plan:
 Index scan(tab tab, i_tab_c, (tab.c='abcd    ') (covers))

 c
======================
'abcd'
'abcd'
'abcd'

SELECT * FROM tab WHERE c='abcd    ' USING INDEX tab.NONE;

Query plan:
 Sequential scan(tab tab)

 c
======================
'abcd'
'abcd    '
'abcd '

Optimizing ORDER BY Clause¶

The index including all columns in the ORDER BY clause is referred to as the ordered index. Optimizing the query with ORDER BY clause is no need for the additional sorting process(skip order by), because the query results are searched by the ordered index. In general, for an ordered index, the columns in the ORDER BY clause should be located at the front of the index.

SELECT *
FROM tab
WHERE col1 > 0
ORDER BY col1, col2;

• The index consisting of tab (col1, col2) is an ordered index.
• The index consisting of tab (col1, col2, col3) is also an ordered index. This is because the col3, which is not referred to by the ORDER BY clause, comes after col1 and col2 .
• The index consisting of tab (col1) is not an ordered index.
• You can use the index consisting of tab (col3, col1, col2) or tab (col1, col3, col2) for optimization. This is because col3 is not located at the back of the columns in the ORDER BY clause.

Although the columns composing an index do not exist in the ORDER BY clause, you can use an ordered index if the column condition is a constant.

SELECT *
FROM tab
WHERE col2=val
ORDER BY col1,col3;

If the index consisting of tab (col1, col2, col3) exists and the index consisting of tab (col1, col2) do not exist when executing the above query, the query optimizer uses the index consisting of tab (col1, col2, col3) as an ordered index. You can get the result in the requested order when you execute an index scan, so you don't need to sort records.

If you can use the sorted index and the covering index, use the latter first. If you use the covering index, you don't need to retrieve additional data, because the data result requested is included in the index page, and you won't need to sort the result if you are satisfied with the index order.

If the query doesn't include any conditions and uses an ordered index, the ordered index will be used under the condition that the first column meets the NOT NULL condition.

CREATE TABLE tab (i INT, j INT, k INT);
CREATE INDEX i_tab_j_k on tab (j,k);
INSERT INTO tab VALUES (1,2,3),(6,4,2),(3,4,1),(5,2,1),(1,5,5),(2,6,6),(3,5,4);

The following example shows that indexes consisting of tab (j, k) become sorted indexes and no separate sorting process is required because GROUP BY is executed by j and k columns.

SELECT i,j,k
FROM tab
WHERE j > 0
ORDER BY j,k;

--  the  selection from the query plan dump shows that the ordering index i_tab_j_k was used and sorting was not necessary
--  (/* --> skip ORDER BY */)
Query plan:
iscan
    class: tab node[0]
    index: i_tab_j_k term[0]
    sort:  2 asc, 3 asc
    cost:  1 card 0
Query stmt:
select tab.i, tab.j, tab.k from tab tab where ((tab.j> ?:0 )) order by 2, 3
/* ---> skip ORDER BY */

            i            j            k
=======================================
            5            2            1
            1            2            3
            3            4            1
            6            4            2
            3            5            4
            1            5            5
            2            6            6

The following example shows that j and k columns execute ORDER BY and the index including all columns are selected so that indexes consisting of tab (j, k) are used as covering indexes; no separate process is required because the value is selected from the indexes themselves.

SELECT /*+ RECOMPILE */ j,k
FROM tab
WHERE j > 0
ORDER BY j,k;

--  in this case the index i_tab_j_k is a covering index and also respects the ordering index property.
--  Therefore, it is used as a covering index and sorting is not performed.

Query plan:
iscan
    class: tab node[0]
    index: i_tab_j_k term[0] (covers)
    sort:  1 asc, 2 asc
    cost:  1 card 0

Query stmt: select tab.j, tab.k from tab tab where ((tab.j> ?:0 )) order by 1, 2
/* ---> skip ORDER BY */

            j            k
==========================
            2            1
            2            3
            4            1
            4            2
            5            4
            5            5
            6            6

The following example shows that i column exists, ORDER BY is executed by j and k columns, and columns that perform SELECT are i, j, and k. Therefore, indexes consisting of tab (i, j, k) are used as covering indexes; separate sorting process is required for ORDER BY j, k even though the value is selected from the indexes themselves.

CREATE INDEX i_tab_j_k ON tab (i,j,k);
SELECT /*+ RECOMPILE */ i,j,k
FROM tab
WHERE i > 0
ORDER BY j,k;

-- since an index on (i,j,k) is now available, it will be used as covering index. However, sorting the results according to
-- the ORDER BY  clause is needed.
Query plan:
temp(order by)
    subplan: iscan
                 class: tab node[0]
                 index: i_tab_i_j_k term[0] (covers)
                 sort:  1 asc, 2 asc, 3 asc
                 cost:  1 card 1
    sort:  2 asc, 3 asc
    cost:  7 card 1

Query stmt: select tab.i, tab.j, tab.k from tab tab where ((tab.i> ?:0 )) order by 2, 3

            i            j            k
=======================================
            5            2            1
            1            2            3
            3            4            1
            6            4            2
            3            5            4
            1            5            5
            2            6            6

Index Scan in Descending Order¶

When a query is executed by sorting in descending order as follows, it usually creates a descending index. In this way, you do not have to go through addition procedure.

SELECT *
FROM tab
[WHERE ...]
ORDER BY a DESC;

However, if you create an ascending index and an descending index in the same column, the possibility of deadlock increases. In order to decrease the possibility of such case, CUBRID supports the descending scan only with ascending index. Users can use the USE_DESC_IDX hint to specify the use of the descending scan. If the hint is not specified, the following three query executions should be considered, provided that the columns listed in the ORDER BY clause can use the index.

• Sequential scan + Sort in descending order
• Scan in general ascending order + sort in descending
• Scan in descending order that does not require a separate scan

Although the USE_DESC_IDX hint is omitted for the scan in descending order, the query optimizer decides the last execution plan of the three listed for an optimal plan.

CREATE TABLE di (i INT);
CREATE INDEX i_di_i on di (i);
INSERT INTO di VALUES (5),(3),(1),(4),(3),(5),(2),(5);

The query will be executed as an ascending scan without USE_DESC_IDX hint.

-- The query will be executed with an ascending scan.

SELECT  *
FROM di
WHERE i > 0
LIMIT 3;

Query plan:

Index scan(di di, i_di_i, (di.i range (0 gt_inf max) and inst_num() range (min inf_le 3)) (covers))

            i
=============
            1
            2
            3

If you add USE_DESC_IDX hint to the above query, a different result will be shown by descending scan.

-- We now run the following query, using the ''use_desc_idx'' SQL hint:

SELECT /*+ USE_DESC_IDX */ *
FROM di
WHERE i > 0
LIMIT 3;

Query plan:
 Index scan(di di, i_di_i, (di.i range (0 gt_inf max) and inst_num() range (min inf_le 3)) (covers) (desc_index))

            i
=============
            5
            5
            5

The following example requires descending order by ORDER BY clause. In this case, there is no USE_DESC_IDX but do the descending scan.

-- We also run the same query, this time asking that the results are displayed in descending order.
-- However, no hint is given.
-- Since ORDER BY...DESC clause exists, CUBRID will use descending scan, even though the hint is not given,
-- thus avoiding to sort the records.

SELECT *
FROM di
WHERE i > 0
ORDER BY i DESC LIMIT 3;

Query plan:
 Index scan(di di, i_di_i, (di.i range (0 gt_inf max)) (covers) (desc_index))

            i
=============
            5
            5
            5

Optimizing GROUP BY Clause¶

GROUP BY clause optimization works on the premise that if all columns in the GROUP BY clause are included in an index, CUBRID can use the index upon executing a query, so CUBRID don't execute a separate sorting job. The columns in the GROUP BY clause must exist in front side of the column forming the index.

SELECT *
FROM tab
WHERE col1 > 0
GROUP BY col1,col2;

• You can use the index consisting of tab(col1, col2) for optimization.
• The index consisting of tab(col1, col2, col3) can be used because col3 which is not referred to by GROUP BY comes after col1 and col2.
• You cannot use the index consisting of tab(col1) for optimization.
• You also cannot use the index consisting of tab(col3, col1, col2) or tab(col1, col3, col2), because col3 is not located at the back of the column in the GROUP BY clause.

You can use the index if the column condition is a constant although the column consisting of the index doesn't exist in the GROUP BY clause.

SELECT *
FROM tab
WHERE col2=val
GROUP BY col1,col3;

If there is any index that consists of tab(col1, col2, col3) in the above example, use the index for optimizing GROUP BY.

Row sorting by GROUP BY is not required, because you can get the result as the requested order on the index scan.

If the index consisting of the GROUP BY column and the first column of the index is NOT NULL, even though there is no WHERE clause, the GROUP BY optimization will be applied.

If there is an index made of GROUP BY columns even when using aggregate functions, GROUP BY optimization is applied.

CREATE INDEX i_T_a_b_c ON T(a, b, c);
SELECT a, MIN(b), c, MAX(b) FROM T WHERE a > 18 GROUP BY a, b;

Example

CREATE TABLE tab (i INT, j INT, k INT);
CREATE INDEX i_tab_j_k ON tab (j, k);
INSERT INTO tab VALUES (1,2,3),(6,4,2),(3,4,1),(5,2,1),(1,5,5),(2,6,6),(3,5,4);

UPDATE STATISTICS on tab;

The following example shows that indexes consisting of tab(j,k) are used and no separate sorting process is required because GROUP BY is executed by j and k columns.

SELECT /*+ RECOMPILE */ j,k
FROM tab
WHERE j > 0
GROUP BY j,k;

--  the  selection from the query plan dump shows that the index i_tab_j_k was used and sorting was not necessary
--  (/* ---> skip GROUP BY */)

Query plan:
iscan
    class: tab node[0]
    index: i_tab_j_k term[0]
    sort:  2 asc, 3 asc
    cost:  1 card 0

Query stmt:
select tab.i, tab.j, tab.k from tab tab where ((tab.j> ?:0 )) group by tab.j, tab.k
/* ---> skip GROUP BY */
            i            j            k
            5            2            1
            1            2            3
            3            4            1
            6            4            2
            3            5            4
            1            5            5
            2            6            6

The following example shows that an index consisting of tab(j,k) is used and no separate sorting process is required while GROUP BY is executed by j and k columns, no condition exists for j, and j column has NOT NULL attribute.

ALTER TABLE tab CHANGE COLUMN j j INT NOT NULL;

SELECT *
FROM tab
GROUP BY j,k;

--  the  selection from the query plan dump shows that the index i_tab_j_k was used (since j has the NOT NULL constraint )
--  and sorting was not necessary (/* ---> skip GROUP BY */)
Query plan:
iscan
    class: tab node[0]
    index: i_tab_j_k
    sort:  2 asc, 3 asc
    cost:  1 card 0

Query stmt: select tab.i, tab.j, tab.k from tab tab group by tab.j, tab.k
/* ---> skip GROUP BY */
=== <Result of SELECT Command in Line 1> ===
            i            j            k
=======================================
            5            2            1
            1            2            3
            3            4            1
            6            4            2
            3            5            4
            1            5            5
            2            6            6

CREATE TABLE tab (k1 int, k2 int, k3 int, v double);
INSERT INTO tab
    SELECT
        RAND(CAST(UNIX_TIMESTAMP() AS INT)) MOD 5,
        RAND(CAST(UNIX_TIMESTAMP() AS INT)) MOD 10,
        RAND(CAST(UNIX_TIMESTAMP() AS INT)) MOD 100000,
        RAND(CAST(UNIX_TIMESTAMP() AS INT)) MOD 100000
    FROM db_class a, db_class b, db_class c, db_class d LIMIT 20000;
CREATE INDEX idx ON tab(k1, k2, k3);

If you create tables and indexes of the above, the following example runs the GROUP BY with k1, k2 columns and performs an aggregate function in k3; therefore, the index which consists of tab(k1, k2, k3) is used and no sort processing is required. In addition, because all columns of k1, k2, k3 of ** SELECT ** list are present in the index configured in the tab(k1, k2, k3), covering index is applied.

SELECT /*+ RECOMPILE INDEX_SS */ k1, k2, SUM(DISTINCT k3)
FROM tab
WHERE k2 > -1 GROUP BY k1, k2;

Query plan:

iscan
    class: tab node[0]
    index: idx term[0] (covers) (index skip scan)
    sort:  1 asc, 2 asc
    cost:  85 card 2000

Query stmt:

select tab.k1, tab.k2, sum(distinct tab.k3) from tab tab where (tab.k2> ?:0 ) group by tab.k1, tab.k2

/* ---> skip GROUP BY */

The following example performs GROUP BY clause with k1, k2 columns; therefore, the index composed with tab(k1, k2, k3) is used and no sort processing is required. However, v column in the SELECT list is not present in the index composed of tab(k1, k2, k3); therefore, it does not apply covering index.

SELECT /*+ RECOMPILE INDEX_SS */ k1, k2, stddev_samp(v)
FROM tab
WHERE k2 > -1 GROUP BY k1, k2;

Query plan:

iscan
    class: tab node[0]
    index: idx term[0] (index skip scan)
    sort:  1 asc, 2 asc
    cost:  85 card 2000

Query stmt:

select tab.k1, tab.k2, stddev_samp(tab.v) from tab tab where (tab.k2> ?:0 ) group by tab.k1, tab.k2

/* ---> skip GROUP BY */

Multiple Key Ranges Optimization¶

Optimizing the LIMIT clause is crucial for performance because the most queries have limit filter. A representative optimization of this case is Multiple Key Ranges Optimization.

Multiple Key Ranges Optimization generate the query result with Top N Sorting to scan only some key ranges in an index rather than doing a full index scan. Top N Sorting always keeps the best N tuples sorted rather than selecting all tuples and then sorting. Therefore, it shows the outstanding performance.

For example, when you search only the recent 10 posts which your friends wrote, CUBRID which applied Multiple KEY Ranges Optimization finds the result not by sorting after finding all your friends' posts, but by scanning the index which keeps the recent 10 sorted posts of each friends.

An example of Multiple Key Ranges Optimization is as follows.

CREATE TABLE t (a int, b int);
CREATE INDEX i_t_a_b ON t (a,b);

-- Multiple key range optimization
SELECT *
FROM t
WHERE a IN (1,2,3)
ORDER BY b
LIMIT 2;

Query plan:
iscan
class: t node[0]
index: i_t_a_b term[0] (covers) (multi_range_opt)
sort: 1 asc, 2 asc
cost: 1 card 0

On a single table, multiple key range optimization can be applied if below conditions are satisfied.

SELECT /*+ hints */ ...
FROM table
WHERE col_1 = ? AND col_2 = ? AND ... AND col(j-1) = ?
AND col_(j) IN (?, ?, ...)
AND col_(j+1) = ? AND ... AND col_(p-1) = ?
AND key_filter_terms
ORDER BY col_(p) [ASC|DESC], col_(p+1) [ASC|DESC], ... col_(p+k-1) [ASC|DESC]
LIMIT n;

Firstly, upper limit(n) for LIMIT should be less than or equal to the value of multi_range_optimization_limit system parameter.

And you need the proper index to the multiple key range optimization, this index should cover all k columns specified in the ORDER BY clause. In other words, this index should include all k columns specified in the ORDER BY clause and the sorting order should be the same as the columns' order. Also this index should include all columns used in WHERE clause.

Among columns that comprise the index,

• Columns in front of range condition(e.g. IN condition) are represented as equivalent condition(=).
• Only one column with range condition exists.
• Columns after range condition exist as key filters.
• There should be no data filtering condition. In other words, the index should include all columns used in WHERE clause.
• Columns after the key filter exist in ORDER BY clause.
• Columns of key filter condition always should not the column of ORDER BY clause.
• If key filter condition with correlated subquery exists, related columns to this should be included into WHERE clause with no range condition.

On the below example, Multiple Key Ranges Optimization can be applied.

CREATE TABLE t (a INT, b INT, c INT, d INT, e INT);
CREATE INDEX i_t_a_b ON t (a,b,c,d,e);

SELECT *
FROM t
WHERE a = 1 AND b = 3 AND c IN (1,2,3) AND d = 3
ORDER BY e
LIMIT 2;

Queries with multiple joined tables can also support Multiple Key Ranges Optimization:

SELECT /*+ hints */ ...
FROM table_1, table_2, ... table_(sort), ...
WHERE col_1 = ? AND col_2 = ? AND ...
AND col_(j) IN (?, ?, ... )
AND col_(j+1) = ? AND ... AND col_(p-1) = ?
AND key_filter_terms
AND join_terms
ORDER BY col_(p) [ASC|DESC], col_(p+1) [ASC|DESC], ... col_(p+k-1) [ASC|DESC]
LIMIT n;

If queries with multiple joined tables can support Multiple Key Ranges Optimization, below conditions should be satisfied:

• Columns in ORDER BY clause only exist on one table, and this table should satisfy all required conditions by Multiple Key Ranges Optimization on a single table query. Let the "sort table" be the table that contains all sorting columns.
• All columns of "sort table" specified in a JOIN condition between "sort table" and "outer tables" should be included on an index. In other words, there should be no data filtering condition.
• All columns of "sort table" specified in a JOIN condition between "sort table" and "outer tables" should be included on the WHERE clause with no range condition.

Index Skip Scan¶

Index Skip Scan (here after ISS) is an optimization method that allows ignoring the first column of an index when the first column of the index is not included in the condition but the following column is included in the condition (in most cases, =).

Applying ISS is considered when INDEX_SS for specific tables is specified through a query hint and the below cases are satisfied.

1. The query condition should be specified from the second column of the composite index.
2. The used index should not be a filtered index.
3. The first column of an index should not be a range filter or key filter.
4. A hierarchical query is not supported.
5. A query which an aggregate function is included is not supported.

In a INDEX_SS hint, a list of table to consider applying ISS, can be input; if a list of table is omitted, applying ISS for all tables can be considered.

/*+ INDEX_SS */
/*+ INDEX_SS(tbl1) */
/*+ INDEX_SS(tbl1, tbl2) */

CREATE TABLE t1 (id INT PRIMARY KEY, a INT, b INT, c INT);
CREATE TABLE t2 (id INT PRIMARY KEY, a INT, b INT, c INT);
CREATE INDEX i_t1_ac ON t1(a,c);
CREATE INDEX i_t2_ac ON t2(a,c);

INSERT INTO t1 SELECT rownum, rownum, rownum, rownum
FROM db_class x1, db_class x2, db_class LIMIT 10000;

INSERT INTO t2 SELECT id, a%5, b, c FROM t1;


SELECT /*+ INDEX_SS */ *
FROM t1, t2
WHERE t1.b<5 AND t1.c<5 AND t2.c<5
USING INDEX i_t1_ac, i_t2_ac limit 1;

SELECT /*+ INDEX_SS(t1) */ *
FROM t1, t2
WHERE t1.b<5 AND t1.c<5 AND t2.c<5
USING INDEX i_t1_ac, i_t2_ac LIMIT 1;

SELECT /*+ INDEX_SS(t1, t2) */ *
FROM t1, t2
WHERE t1.b<5 AND t1.c<5 AND t2.c<5
USING INDEX i_t1_ac, i_t2_ac LIMIT 1;

Generally, ISS should consider several columns (C1, C2, ..., Cn). Here, a query has the conditions for the consecutive columns and the conditions are started from the second column (C2) of the index.

INDEX (C1, C2, ..., Cn);

SELECT ... WHERE C2 = x AND C3 = y AND ... AND Cp = z; -- p <= n
SELECT ... WHERE C2 < x AND C3 >= y AND ... AND Cp BETWEEN (z AND w); -- other conditions than equal

The query optimizer eventually determines whether ISS is the most optimum access method based on the cost. ISS is applied under very specific situations, such as when the first column of an index has a very small number of DISTINCT values compared to the number of records. In this case, ISS provides higher performance compared to Index Full Scan. For example, when the first column of index columns has very low cardinality, such as the value of men/women or millions of records with the value of 1~100, it may be inefficient to perform index scan by using the first column value. So ISS is useful in this case.

ISS skips reading most of the index pages in the disk and uses range search which is dynamically readjusted. Generally, ISS can be applied to a specific scenario when the number of DISTINCT values in the first column is very small. If ISS is applied to this case, ISS provides significantly higher performance than the index full scan. However, it means improper index creation that ISS is applied to a lot queries. So DBA should consider whether readjusting the indexes or not.

CREATE TABLE tbl (name STRING, gender CHAR (1), birthday DATETIME);

INSERT INTO tbl
SELECT ROWNUM, CASE (ROWNUM MOD 2) WHEN 1 THEN 'M' ELSE 'F' END, SYSDATETIME
FROM db_class a, db_class b, db_class c, db_class d, db_class e LIMIT 360000;

CREATE INDEX idx_tbl_gen_name ON tbl (gender, name);
-- Note that gender can only have 2 values, 'M' and 'F' (low cardinality)

UPDATE STATISTICS ON ALL CLASSES;

-- csql>;plan simple
-- this will qualify to use Index Skip Scanning
SELECT /*+ RECOMPILE INDEX_SS */ *
FROM tbl
WHERE name = '1000';

Query plan:

 Index scan(tbl tbl, idx_tbl_gen_name, tbl.[name]= ?:0  (index skip scan))

-- csql>;plan simple
-- this will qualify to use Index Skip Scanning
SELECT /*+ RECOMPILE INDEX_SS */ *
FROM tbl
WHERE name between '1000' and '1050';

Query plan:

 Index scan(tbl tbl, idx_tbl_gen_name, (tbl.[name]>= ?:0  and tbl.[name]<= ?:1 ) (index skip scan))

Loose Index Scan¶

When GROUP BY clause or DISTINCT column includes a subkey of a index, loose index scan starts B-tree search by adjusting the range dynamically for unique value of each of the columns that make up the subkey. Therefore, it is possible to significantly reduce the scanning area of B-tree.

Applying loose index scan is advantageous when the cardinality of the grouped column is very small, compared to the total data amount.

Loose index scan optimization is considered to be applied when INDEX_LS is input as a hint and the below cases are satisfied:

1. when an index covers all SELECT list, that is, covered index is applied.
2. when the statement is SELECT DISTINCT, SELECT ... GROUP BY statement or a single tuple SELECT.
3. all aggregate functions (with the exception of MIN/MAX) must have DISTINCT input
4. COUNT(*) should not be used
5. when cardinality of the used subkey is 100 times smaller than the cardinality of the whole index

a subkey is a prefix part in a composite index; e.g. when there is INDEX(a, b, c, d), (a), (a, b) or (a, b, c) belongs to the subkey.

When you run the below query regarding the above table,

SELECT /*+ INDEX_LS */ a, b FROM tbl GROUP BY a;

CUBRID cannot use a subkey because there is no condition for the column a. However, if the condition of the subkey is specified as follows, loose index scan can be applied.

SELECT /*+ INDEX_LS */ a, b FROM tbl WHERE a > 10 GROUP BY a;

As follows, a subkey can be used when the grouped column is on the first and the WHERE-condition column is on the following position; therefore, also in this case, loose index scan can be applied.

SELECT /*+ INDEX_LS */ a, b FROM tbl WHERE b > 10 GROUP BY a;

The following shows the cases when loose index scan optimization is applied.

CREATE TABLE tbl1 (
    k1 INT,
    k2 INT,
    k3 INT,
    k4 INT
);

INSERT INTO tbl1
SELECT ROWNUM MOD 2, ROWNUM MOD 400, ROWNUM MOD 80000, ROWNUM
FROM db_class a, db_class b, db_class c, db_class d, db_class e LIMIT 360000;

CREATE INDEX idx ON tbl1 (k1, k2, k3);

CREATE TABLE tbl2 (
    k1 INT,
    k2 INT
);

INSERT INTO tbl2 VALUES (0, 0), (1, 1), (0, 2), (1, 3), (0, 4), (0, 100), (1000, 1000);

UPDATE STATISTICS ON ALL CLASSES;

-- csql>;plan simple
-- add a condition to the grouped column, k1 to enable loose index scan
SELECT /*+ RECOMPILE INDEX_LS */ DISTINCT k1
FROM tbl1
WHERE k1 > -1000000 LIMIT 20;

Query plan:

 Sort(distinct)
    Index scan(tbl1 tbl1, idx, (tbl1.k1> ?:0 ) (covers) (loose index scan on prefix 1))

-- csql>;plan simple
-- different key ranges/filters
SELECT /*+ RECOMPILE INDEX_LS */ DISTINCT k1
FROM tbl1
WHERE k1 >= 0 AND k1 <= 1;

Query plan:

 Sort(distinct)
    Index scan(tbl1 tbl1, idx, (tbl1.k1>= ?:0  and tbl1.k1<= ?:1 ) (covers) (loose index scan on prefix 1))

-- csql>;plan simple
SELECT /*+ RECOMPILE INDEX_LS */ DISTINCT k1, k2
FROM tbl1
WHERE k1 >= 0 AND k1 <= 1 AND k2 > 3 AND k2 < 11;

Query plan:

 Sort(distinct)
    Index scan(tbl1 tbl1, idx, (tbl1.k1>= ?:0  and tbl1.k1<= ?:1 ), [(tbl1.k2> ?:2  and tbl1.k2< ?:3 )] (covers) (loose index scan on prefix 2))

-- csql>;plan simple
SELECT /*+ RECOMPILE INDEX_LS */ DISTINCT k1, k2
FROM tbl1
WHERE k1 >= 0 AND k1 + k2 <= 10;

Query plan:

 Sort(distinct)
    Index scan(tbl1 tbl1, idx, (tbl1.k1>= ?:0 ), [tbl1.k1+tbl1.k2<=10] (covers) (loose index scan on prefix 2))

-- csql>;plan simple
SELECT /*+ RECOMPILE INDEX_LS */ tbl1.k1, tbl1.k2
FROM tbl2 INNER JOIN tbl1
ON tbl2.k1 = tbl1.k1 AND tbl2.k2 = tbl1.k2
GROUP BY tbl1.k1, tbl1.k2;

Sort(group by)
   Nested loops
       Sequential scan(tbl2 tbl2)
       Index scan(tbl1 tbl1, idx, tbl2.k1=tbl1.k1 and tbl2.k2=tbl1.k2 (covers) (loose index scan on prefix 2))

SELECT /*+ RECOMPILE INDEX_LS */ MIN(k2), MAX(k2)
FROM tbl1;

Query plan:

 Index scan(tbl1 tbl1, idx (covers) (loose index scan on prefix 2))

-- csql>;plan simple
SELECT /*+ RECOMPILE INDEX_LS */ SUM(DISTINCT k1), SUM(DISTINCT k2)
FROM tbl1;

Query plan:

 Index scan(tbl1 tbl1, idx (covers) (loose index scan on prefix 2))

-- csql>;plan simple
SELECT /*+ RECOMPILE INDEX_LS */ DISTINCT k1
FROM tbl1
WHERE k2 > 0;

Query plan:

 Sort(distinct)
    Index scan(tbl1 tbl1, idx, [(tbl1.k2> ?:0 )] (covers) (loose index scan on prefix 2))

The following shows the cases when loose index scan optimization is not applied.

-- csql>;plan simple
-- not enabled when full key is used
SELECT /*+ RECOMPILE INDEX_LS */ DISTINCT k1, k2, k3
FROM tbl1
ORDER BY 1, 2, 3 LIMIT 10;

Query plan:

 Sort(distinct)
    Sequential scan(tbl1 tbl1)

-- csql>;plan simple
SELECT /*+ RECOMPILE INDEX_LS */ k1, k2, k3
FROM tbl1
WHERE k1 > -10000 GROUP BY k1, k2, k3 LIMIT 10;

Query plan:

 Index scan(tbl1 tbl1, idx, (tbl1.k1> ?:0 ) (covers))
 skip GROUP BY

-- csql>;plan simple
-- not enabled when using count star
SELECT /*+ RECOMPILE INDEX_LS */ COUNT(*), k1
FROM tbl1
WHERE k1 > -10000 GROUP BY k1;

Query plan:

 Index scan(tbl1 tbl1, idx, (tbl1.k1> ?:0 ) (covers))
 skip GROUP BY

-- csql>;plan simple
-- not enabled when index is not covering
SELECT /*+ RECOMPILE INDEX_LS */ k1, k2, SUM(k4)
FROM tbl1
WHERE k1 > -1 AND k2 > -1 GROUP BY k1, k2 LIMIT 10;

Query plan:

 Index scan(tbl1 tbl1, idx, (tbl1.k1> ?:0 ), [(tbl1.k2> ?:1 )])
 skip GROUP BY

-- csql>;plan simple
-- not enabled for non-distinct aggregates
SELECT /*+ RECOMPILE INDEX_LS */ k1, SUM(k2)
FROM tbl1
WHERE k1 > -1 GROUP BY k1;

Query plan:

 Index scan(tbl1 tbl1, idx, (tbl1.k1> ?:0 ) (covers))
 skip GROUP BY

-- csql>;plan simple
SELECT /*+ RECOMPILE */ SUM(k1), SUM(k2)
FROM tbl1;

Query plan:

 Sequential scan(tbl1 tbl1)

In Memory Sort¶

The "in memory sort(IMS)" feature is an optimization applied to the LIMIT queries specifying ORDER BY. Normally, when executing a query which specifies ORDER BY and LIMIT clauses, CUBRID generates the full sorted result set and then applies the LIMIT operator to this result set. With the IMS optimization, instead of generating the whole result set, CUBRID uses an in-memory binary heap in which only tuples satisfying the ORDER BY and LIMIT clauses are allowed. This optimization improves performance by eliminating the need for a full unordered result set.

Whether this optimization is applied or not is not transparent to users. CUBRID decides to use in memory sort in the following situation:

• The query specifies ORDER BY and LIMIT clauses.
• The size of the final result (after applying the LIMIT clause) is less than the amount of memory used by external sort (see sort_buffer_size in Memory-Related Parameters).

Note that IMS considers the actual size of the result and not the count of tuples the result contains. For example, for the default sort buffer size (two megabytes), this optimization will be applied for a LIMIT value of 524,288 tuples consisting of one 4 byte INTEGER type but only for ~2,048 tuples of CHAR(1024) values. This optimization is not applied to queries requiring DISTINCT ordered result sets.

SORT-LIMIT optimization¶

The SORT-LIMIT optimization applies to queries specifying ORDER BY and LIMIT clauses. The idea behind it is to evaluate the LIMIT operator as soon as possible in the query plan in order to benefit from the reduced cardinality during joins.

A SORT-LIMIT plan can be generated when the following conditions are met:

• All referred tables in the ORDER BY clause belong to the SORT-LIMIT plan.

• A table belonging to a SORT-LIMIT plan is either:

  • The owner of a foreign key from a fk->pk join
  • The left side of a LEFT JOIN.
  • The right side of a RIGHT JOIN.

• LIMIT rows should be specified as less rows than the value of sort_limit_max_count system parameter(default: 1000).

• Query does not have cross joins.

• Query joins at least two relations.

• Query does not have a GROUP BY clause.

• Query does not specify DISTINCT.

• ORDER BY expressions can be evaluated during scan.

  For example, the below query cannot apply SORT-LIMIT plan because SUM cannot be evaluated during scan.

  SELECT SUM(u.i) FROM u, t where u.i = t.i ORDER BY 1 LIMIT 5;
  

The below is an example of planning SORT-LIMIT.

CREATE TABLE t(i int PRIMARY KEY, j int, k int);
CREATE TABLE u(i int, j int, k int);
ALTER TABLE u ADD constraint fk_t_u_i FOREIGN KEY(i) REFERENCES t(i);
CREATE INDEX i_u_j ON u(j);

INSERT INTO t SELECT ROWNUM, ROWNUM, ROWNUM FROM _DB_CLASS a, _DB_CLASS b LIMIT 1000;
INSERT INTO u SELECT 1+(ROWNUM % 1000), RANDOM(1000), RANDOM(1000) FROM _DB_CLASS a, _DB_CLASS b, _DB_CLASS c LIMIT 5000;

SELECT /*+ RECOMPILE */ * FROM u, t WHERE u.i = t.i AND u.j > 10 ORDER BY u.j LIMIT 5;

The above SELECT query's plan is printed out as below; we can see "(sort limit)".

Query plan:

idx-join (inner join)
    outer: temp(sort limit)
               subplan: iscan
                            class: u node[0]
                            index: i_u_j term[1]
                            cost:  1 card 0
               cost:  1 card 0
    inner: iscan
               class: t node[1]
               index: pk_t_i term[0]
               cost:  6 card 1000
    sort:  2 asc
    cost:  7 card 0