21 Extensible Optimizer Interface

EXPLAIN PLAN

EXPLAIN PLAN has been enhanced to show the user-defined CPU and I/O costs for domain indexes in the CPU_COST and IO_COST columns of PLAN_TABLE. For example, suppose we have a table Emp_tab and a user-defined operator Contains. Further, suppose that there is a domain index EmpResume_indx on the Resume_col column of Emp_tab, and that the indextype of EmpResume_indx supports the operator Contains. Then, the query

SELECT * FROM Emp_tab WHERE Contains(Resume_col, 'Oracle') = 1

might have the following plan:

SELECT STATEMENT

TABLE ACCESS

BY ROWID

EMP_TAB

DOMAIN INDEX

EMPRESUME_INDX

300

4

INDEX Hint

The index hint will apply to domain indexes. In other words, the index hint will force the optimizer to use the hinted index for a user-defined operator, if possible.

ORDERED_PREDICATES Hint

The hint ORDERED_PREDICATES forces the optimizer to preserve the order of predicate evaluation (except predicates used for index keys) as specified in the WHERE clause of a SQL DML statement.

Example

Consider an example of how the statistics functions might be used. Suppose, in the schema SCOTT, we define the following:

CREATE OPERATOR Contains binding (VARCHAR2(4000), VARCHAR2(30))
   RETURN NUMBER USING Contains_fn;

CREATE TYPE stat1 (
   ...,
   STATIC FUNCTION ODCIStatsSelectivity(pred ODCIPredInfo, sel OUT NUMBER,
      args ODCIArgDescList, start NUMBER, stop NUMBER, doc VARCHAR2(4000),
      key VARCHAR2(30)) return NUMBER,
   STACTIC FUNCTION ODCIStatsFunctionCost(func ODCIFuncInfo, cost OUT
      ODCICost, args ODCIArgDescList, doc VARCHAR2(4000), key VARCHAR2(30))
      return NUMBER,
   STATIC FUNCTION ODCIStatsIndexCost(ia ODCIIndexInfo, sel NUMBER,
      cost OUT ODCICost, qi ODCIQueryInfo, pred ODCIPredInfo,
      args ODCIArgDescList, start NUMBER, stop NUMBER,
      key VARCHAR2(30)) return NUMBER,
   ...
);

CREATE TABLE T (resume VARCHAR2(4000));

CREATE INDEX T_resume on T(resume) INDEXTYPE IS indtype;

ASSOCIATE STATISTICS WITH FUNCTIONS Contains_fn USING stat1;

ASSOCIATE STATISTICS WITH INDEXES T_resume USING stat1;

When the optimizer encounters the query

SELECT * FROM T WHERE Contains(resume, 'ORACLE') = 1,

it will compute the selectivity of the predicate by invoking the user-defined selectivity function for the functional implementation of the Contains operator. In this case, the selectivity function is stat1.ODCIStatsSelectivity. It will be called as follows:

stat1.ODCIStatsSelectivity (
ODCIPredInfo('SCOTT', 'Contains_fn', NULL, 29),
sel,
ODCIArgDescList(
   ODCIArgDesc(ODCIConst.ArgLit, NULL, NULL, NULL),
   ODCIArgDesc(ODCIConst.ArgLit, NULL, NULL, NULL),
   ODCIArgDesc(ODCIConst.ArgCol, 'T', 'SCOTT', '"resume"'),
   ODCIArgDesc(ODCIConst.ArgLit, NULL, NULL, NULL)),
   1,
   1,
   NULL,
   'ORACLE')

Suppose the selectivity function returns a selectivity of 3 (percent). When the domain index is being evaluated, then the optimizer will call the user-defined index cost function as follows:

stat1.ODCIStatsIndexCost (
   ODCIIndexInfo('SCOTT', 'T_resume',
      ODCIColInfoList(ODCIColInfo('SCOTT', 'T', '"resume"', NULL, NULL))),
      3,
      cost,
      NULL,
      ODCIPredInfo('SCOTT', 'Contains', NULL, 13),
      ODCIArgDescList( ODCIArgDesc(ODCIConst.ArgLit, NULL, NULL, NULL),
                       ODCIArgDesc(ODCIConst.ArgLit, NULL, NULL, NULL),
                       ODCIArgDesc(ODCIConst.ArgLit, NULL, NULL, NULL)),
      1,
      1,
      'ORACLE')

Suppose that the optimizer decides not to use the domain index because it is too expensive. Then it will call the user-defined cost function for the functional implementation of the operator as follows:

stat1.ODCIStatsFunctionCost (
   ODCIFuncInfo('SCOTT', 'Contains_fn', NULL, 1),
   cost,
   ODCIArgDescList(  ODCIArgDesc(ODCIConst.ArgCol, 'T', 'SCOTT', '"resume"'),
                     ODCIArgDesc(ODCIConst.ArgLit, NULL, NULL, NULL)),
   NULL,
   'ORACLE')

The following sections describe each statistics type function in greater detail.

ODCIGetInterfaces

ODCIGetInterfaces is invoked by the server to discover which version of the ODCIStats interface the user has implemented in the methods of the user-defined statistics type.

ODCIGetInterfaces(
   ifclist OUT ODCIObjectList)
RETURN NUMBER

ifclist (OUT)

ODCIConst.Success on success, ODCIConst.Error otherwise.

Different versions of ODCIStats functions are used by Oracle8i and subsequent versions of Oracle Database. More recent versions adds parameters to some functions to support working with statistics on partitions of a table or domain index. ODCIGetInterfaces must return the string 'SYS.ODCISTATS2' in the ODCIObjectList parameter, which indicates that the statistics type uses the current form of the ODCIStats interface.

ODCIStatsCollect

Called by the DBMS_STATS package to collect user-defined statistics on a table or a partition of a table.

FUNCTION ODCIStatsCollect(
   col ODCIColInfo, 
   options ODCIStatsOptions, 
   statistics OUT RAW) 
return NUMBER

The function returns ODCIConst.Success, ODCIConst.Error, or ODCIConst.Warning.

• This function should be implemented as a static type method.

• If statistics are being collected for only one partition, the TablePartition field in the ODCIColInfo type is filled in with the name of the partition. Otherwise (if statistics need to be collected for all the partitions or for the entire table), the TablePartition field is null.

• If the DBMS_STATS package methods are executed to collect user-defined statistics on a partitioned table, then n+1 ODCIStatsCollect calls are made, where n is the number of partitions in the table. The first n calls are made with the TablePartition attribute in ODCIColInfo filled in with the partition name and the ODCIStatsOptions.CallProperty set to IntermediateCall. The last call is made with ODCIEnv.CallPropertyflag set to FinalCall to allow you to collect aggregate statistics for the entire table. The OUT statistics in the first call are ignored by the server. The OUT statistics in the subsequent n calls are inserted into the USTATS$ table corresponding to the partitions. The OUT statistics in the last call are the aggregate statistics for the table. The ODCIColInfo.Partition field is NULL in the first and last calls.

• If user-defined statistics are being collected for only one partition of the table, two ODCIStatsCollect calls are made. In the first, you should collect statistics for the partition. For this call, the TablePartition attribute of the ODCIColInfo structure is filled in and the ODCIEnv.CallProperty is set to FirstCall. The statistics in the OUT arguments in the ODCIStatsCollect call are inserted into the USTATS$ table corresponding to the partition.

• In the second call you can update the aggregate statistics of the table based upon the new statistics collected for the partition. In this call, the ODCIEnv.CallPropertyflag is set to FinalCall to indicate that it is the second call. If you do not want to modify the aggregate statistics, read the aggregate statistics of the table from the catalog and pass that back in the statistics field as the OUT argument. Whatever value is present in the statistics argument is written in the USTATS$ by the server. The ODCIColInfo.TablePartition is filled in with the partition name in both the calls.

• Return 'SYS.ODCISTATS2' in the ODCIGetInterfaces call to indicate that you are using a post Oracle8i version of the ODCISTATS interface that supports partitioning.

ODCIStatsCollect

ODCIStatsCollect is called by the to collect user-defined statistics on an index or a partition of an index.

FUNCTION ODCIStatsCollect(
   ia ODCIIndexInfo, 
   options ODCIStatsOptions, 
   statistics OUT RAW)
return NUMBER

The function returns ODCIConst.Success, ODCIConst.Error, or ODCIConst.Warning.

This function should be implemented as a static type method.

If statistics are being collected for the entire partitioned index, the IndexPartition field is null, and n+2 calls are made to the ODCIStatsCollect function. This scenario is similar to that described for the column version of ODCIStatsCollect.

If the statistics are being collected for a single partition of the index, the IndexPartition field contains the name of the partition, and two calls are made to the ODCIStatsCollect function. The first call is made to obtain the statistics for the index partition, and the second call is made to obtain the aggregate statistics for the domain index.

To collect statistics on a non-partitioned domain index only a single call is made to the ODCIStatsCollect function.

Return 'SYS.ODCISTATS2' in the ODCIGetInterfaces call to indicate that you are using a post-Oracle8i version of the ODCISTATS interface that supports partitioning.

ODCIStatsDelete

FUNCTION ODCIStatsDelete(
   col ODCIColInfo, 
   statistics OUT RAW,
   env ODCIEnv) 
return NUMBER

ODCIStatsDelete is called to delete user-defined statistics on a table or a partition of a table.

ODCIConst.Success, ODCIConst.Error, or ODCIConst.Warning.

This function should be implemented as a static method.

When the function is called for a non-partitioned table, the statistics argument in the ODCIStatsDelete interface is ignored.

If the statistics are being deleted for a partitioned table, the ODCIStatsDelete is called n+1 times. The first n calls are with the partition name filled in the ODCIColInfo structure and the ODCIEnv.CallProperty set to IntermediateCall. The last call is made with the ODCIEnv.CallProperty set to FinalCall.

The order of operations that you must perform for a delete are the inverse of what you do to collect statistics: In the first call, delete the table-level statistics from your statistics tables; in the intermediate n calls, delete the statistics for the specific partitions; and in the last call drop or clean up any structures created for holding statistics for the deleted table. The ODCIColInfo.TablePartition is set to null in the first and last calls. In the intermediate n calls, the TablePartition field is filled in.

If statistics are being deleted for only one partition and the _minimal_stats_aggregation parameter is set to FALSE, two ODCIStatsDelete calls are made. In each call, ODCIColInfo.TablePartition is filled in with the partition name. On the first call, delete any user-defined statistics collected for that partition. On the second call, update the aggregate statistics for the table and return these aggregate statistics as an OUT argument.

If statistics are being deleted for one partition and _minimal_stats_aggregation is set to TRUE, ODCIStatsDelete is only called one to delete any user-defined statistics collected for that partition.

The initial value of _minimal_stats_aggregation is TRUE.

Return 'SYS.ODCISTATS2' in the ODCIGetInterfaces call to indicate that you are using a post-Oracle8i version of the ODCISTATS interface that supports partitioning.

ODCIStatsDelete (Index)

FUNCTION ODCIStatsDelete(
   ia ODCIIndexInfo, 
   statistics OUT RAW, 
   env ODCIEnv) 
return NUMBER

ODCIStatsDelete is called to delete user-defined statistics on an index or a partition of an index.

ODCIConst.Success, ODCIConst.Error, or ODCIConst.Warning.

• This function should be implemented as a static method.

• When the function is called for a non-partitioned index, the statistics argument in the ODCIStatsDelete interface is ignored.

• If statistics are being deleted for a partitioned index, ODCIStatsDelete is called n+2 times. The first and the last call are made with the ODCIEnv.CallProperty set to FirstCall and FinalCall respectively and do not have the partition name set in the ODCIIndexInfo type. The intermediate n calls are made with the partition name filled in the ODCIIndexInfo structure and the ODCIEnv.CallProperty set to IntermediateCall.

• The order of operations that you must perform to delete statistics are the inverse of what you do to collect statistics: In the first call, delete the index-level statistics from your statistics tables; in the intermediate n calls, delete the statistics for the specific partitions; and in the last call drop or clean up any structures created for holding the deleted statistics. The ODCIIndexInfo.IndexPartition is set to null in the first and last calls. In the intermediate n calls, the IndexPartition field is filled in.

• If statistics are being deleted for only one partition and the parameter _minimal_stats_aggregation is set to FALSE, two ODCIStatsDelete calls are made. In each call, ODCIIndexInfo.IndexPartition is filled in with the partition name. On the first call, delete any user-defined statistics collected for that partition. On the second call, update the aggregate statistics for the index and return these aggregate statistics as an OUT argument.

• If statistics are being deleted for one partition and _minimal_stats_aggregation is set to TRUE, ODCIStatsDelete is only called one to delete any user-defined statistics collected for that index.

• The initial value of _minimal_stats_aggregation is TRUE.

• Return 'SYS.ODCISTATS2' in the ODCIGetInterfaces call to indicate that you are using a version of the ODCISTATS interface that supports partitioning.

ODCIStatsFunctionCost

Computes the cost of a function.

FUNCTION ODCIStatsFunctionCost(
   func ODCIFuncInfo, 
   cost OUT ODCICost, 
   args ODCIArgDescList, 
   <list of function arguments>) 
return NUMBER

func

cost

args

<list of function arguments>

ODCIConst.Success, ODCIConst.Error, or ODCIConst.Warning.

This function should be implemented as a static type method.

ODCIStatsIndexCost

Calculates the cost of a domain index scan—either a scan of the entire index or a scan of one or more index partitions if a local domain index has been built.

FUNCTION ODCIStatsIndexCost(
   ia ODCIIndexInfo, 
   sel NUMBER, 
   cost OUT ODCICost, 
   qi ODCIQueryInfo, 
   pred ODCIPredInfo, 
   args ODCIArgDescList,  
   start <operator_return_type>, 
   stop <operator_return_type>,  
   <list of operator arguments>, 
   env ODCIEnv) 
return NUMBER

ia

sel

cost

qi

pred

args

start

stop

<list of function arguments>

env

ODCIConst.Success, ODCIConst.Error, or ODCIConst.Warning

• For each table in the query, the optimizer uses partition pruning to determine the range of partitions that may be accessed. These partitions are called interesting partitions. The set of interesting partitions for a table is also the set of interesting partitions for all domain indexes on that table. The cost of a domain index can depend on the set of interesting partitions, so the optimizer passes a list of interesting index partitions to ODCIStatsIndexCost in the args argument (the type of this argument, ODCIArgDescList, is a list of ODCIArgDesc argument descriptor types) for those arguments that are columns. For non-partitioned domain indexes or for cases where no partition pruning is possible, no partition list is passed to ODCIStatsIndexCost, and you should assume that the entire index will be accessed.

• The domain index key can contain multiple column arguments (for example, the indexed column and column arguments from other tables appearing earlier in a join order). For each column appearing in the index key, the args argument contains the list of interesting partitions for the table. For example, for an index key

  op(T1.c1, T2.c2) = 1
  
  

  the optimizer passes a list of interesting partitions for tables T1 and T2 if they are partitioned and there is partition pruning for them.

• This function should be implemented as a static type method.

• Only a single call is made to the ODCIIndexCost function for queries on partitioned or non-partitioned tables. For queries on partitioned tables, additional information is passed in the ODCIIndexCost function. Note that some partitions in the list passed to ODCIStatsIndexCost may not actually be accessed by the query. The list of interesting partitions chiefly serves to exclude partitions that definitely will not be accessed.

• When the ODCIIndexCost function is invoked, users can fill in a string in the IndexCostInfo field of the cost attribute to supply any additional information that might be helpful. The string (255 characters maximum) is displayed in the OPTIONS column in the EXPLAIN PLAN output when an execution plan chooses a domain index scan.

• Users implementing this function must return 'SYS.ODCISTATS2' in the ODCIGetInterfaces call.

ODCIStatsSelectivity

Specifies the selectivity of a predicate. The selectivity of a predicate involving columns from a single table is the fraction of rows of that table that satisfy the predicate. For predicates involving columns from multiple tables (for example, join predicates), the selectivity should be computed as a fraction of rows in the Cartesian product of those tables.

FUNCTION ODCIStatsSelectivity(
   pred ODCIPredInfo, 
   sel OUT NUMBER, 
   args ODCIArgDescList, 
   start <function_return_type>, 
   stop <function_return_type>, 
   <list of function arguments>, 
   env ODCIEnv) 
return NUMBER

ODCIConst.Success, ODCIConst.Error, or ODCIConst.Warning

• As in ODCIStatsIndexCost, the args argument contains a list of interesting partitions for the tables whose columns are referenced in the predicate for which the selectivity has to be computed. These interesting partitions are partitions that cannot be eliminated by partition pruning as possible candidates to be accessed. The set of interesting partitions is passed to the function only if partition pruning has occurred (in other words, the interesting partitions are a strict subset of all the partitions).

• For example, when ODCIStatsSelectivity is called to compute the selectivity of the predicate:

  f(T1.c1, T2.c2) > 4
  
  

  the optimizer passes the list of interesting partitions for the table T1 (in the argument descriptor for column T1.c1) if partition pruning is possible; similarly for the table T2.

  If a predicate contains columns from more than one table, this information is indicated by the flag bit PredMultiTable, set in the Flags attribute of the pred argument.

• This function should be implemented as a static type method.

• Users implementing this interface must return 'SYS.ODCISTATS2' in the ODCIGetInterfaces call.

• The selectivity of a predicate involving columns from a single table is the fraction of rows of that table that satisfy the predicate. For predicates involving columns from multiple tables (for example, join predicates), the selectivity should be computed as a fraction of rows in the Cartesian product of those tables. For tables with partition pruning, the selectivity should be expressed relative to the cardinalities of the interesting partitions of the tables involved.

  The selectivity of predicates involving columns on partitioned tables is computed relative to the rows in the interesting partitions. Thus, the selectivity of the predicate

  g(T1.c1) < 5
  
  

  is the percentage of rows in the set of interesting partitions (or all partitions if no partition pruning is possible) that satisfies this predicate. For predicates with columns from multiple tables, the selectivity must be relative to the number of rows in the cartesian product of the tables.

• For example, consider the predicate:

  f(T1.c1, T2.c2) > 4
  
  

  Suppose that the number of rows in the interesting partitions is 1000 for T1 and 5000 for T2. The selectivity of this predicate must be expressed as the percentage of the 5,000,000 rows in the Cartesian product of T1 and T2 that satisfy the predicate.

• If a predicate contains columns from more than one table, this information is indicated by the flag bit PredMultiTable set in the Flags attribute of the pred argument.

• A selectivity expressed relative to the base cardinalities of the tables involved may be only an approximation of the true selectivity if cardinalities (and other statistics) of the tables have been reduced based on single-table predicates or other joins earlier in the join order. However, this approximation to the true selectivity should be acceptable to most applications.

• Only one call is made to the ODCIStatsSelectivity function for queries on partitioned or non-partitioned tables. In the case of queries on partitioned tables, additional information is passed while calling the ODCIStatsSelectivity function.

ODCIStatsTableFunction

This function provides cardinality statistics for table functions and input cursor expressions.

STATIC FUNCTION ODCIStatsTableFunction(
   func IN SYS.ODCIFuncInfo, 
   outStats OUT SYS.ODCITabFuncStats, 
   argDesc IN SYS.ODCIArgDescList, 
   <list of function arguments> 
RETURN NUMBER

ODCIConst.Success, ODCIConst.Error, or ODCIConst.Warning.