21 Extensible Optimizer Interface

EXPLAIN PLAN

EXPLAIN PLAN shows 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 applies to domain indexes. In other words, the index hint forces 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.

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.

Syntax

FUNCTION ODCIGetInterfaces(
   ifclist OUT ODCIObjectList)
RETURN NUMBER;

ifclist

OUT

Returns

ODCIConst.Success on success, ODCIConst.Error otherwise.

ODCIStatsCollect()

Called by the DBMS_STATS package to collect user-defined statistics.

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

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

Returns

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

Usage Notes

• 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 must 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.

• 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.

• 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. The ODCIColInfo.TablePartition is filled in with the partition name in both the calls.

• The ODCIStatsCollect() method is invoked only one time for a non-partitioned domain index, a partitioned domain index and a partition in a domain index. If the statistics are being collected only for one partition in a domain index, the IndexPartitionNum field in the ODCIIndexInfo type is filled in with the partition number. Otherwise, the IndexPartitionNum field is null.

• Because the statistics OUT RAW argument of statistics is not used in the new interface, the cartridge developer should store the user-defined statistics result in some user-defined tables.

• If a non-partitioned domain index is being ANALYZEd, the user should collect statistics for the domain index.

• If a partitioned domain index is being ANALYZEd,

  • ODCIEnv.CallProperty = StatsGlobalAndPartition means that the user should collect statistics for all partitions in the domain index and then aggregate statistics of the domain index based upon the statistics collected for all the partitions

  • ODCIEnv.CallProperty = StatsGlobal means that the user should aggregate domain index statistics from the statistics of all the domain index partitions.

  • ODCIEnv.CallProperty = StatsPartition means that the user should collect statistics for all index partitions in the domain index.

• If only one partition of the domain index is being ANALYZEd,

  • ODCIEnv.CallProperty = StatsGlobalAndPartition means that the user should collect statistics for the single index partition and then aggregate statistics of the domain index based upon the statistics of all the partitions.

  • ODCIEnv.CallProperty = StatsGlobal means that the user should aggregate domain index statistics from the statistics of all the index partitions.

  • ODCIEnv.CallProperty = StatsPartition means that the user should collect statistics for the single index partition.

• Note that when ODCIEnv.CallProperty = StatsGlobalAndPartition or StatsGlobal, the user should aggregate statistics for the domain index, depending on the availability of the statistics collected for the other index partitions. If the statistics for all the index partitions are available, aggregate these statistics. If any one statistics for an index partition is absent, do nothing.

ODCIStatsDelete()

ODCIStatsDelete is called to delete user-defined statistics.

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

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

OUT

Returns

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

Usage Notes

• 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.

• In the first call, 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 last call. In the first call, 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.

• 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.

• The ODCIStatsDelete() method is invoked only one time for non-partitioned domain index, partitioned domain index, or an index partition.

• If the statistics is being deleted for a non-partitioned domain index, the user should delete user-defined statistics for the domain index.

• If the statistics is being deleted for a partitioned domain index, the user should delete the aggregated statistics of the domain index and optionally delete user-defined statistics for all domain index partitions, depending on Options in ODCIEnv.CallProperty:

  • ODCIEnv.CallProperty = StatsGlobalAndPartition means that the user should delete statistics for all the domain index partitions and aggregated statistics of the domain index.

  • ODCIEnv.CallProperty = StatsGlobal means that the user should delete the aggregated statistics of the domain index.

  • ODCIEnv.CallProperty = StatsPartition is not valid option.

• If the statistics is being deleted for only one partition of the index, the user should delete user-defined statistics for the index partition.

ODCIStatsFunctionCost()

Computes the cost of a function.

Syntax

FUNCTION ODCIStatsFunctionCost(
   func ODCIFuncInfo, 
   cost OUT ODCICost, 
   args ODCIArgDescList, 
   list,
   env ODCIEnv) 
return NUMBER;

func

cost

OUT

args

list

env

Returns

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

Usage Notes

This function should be implemented as a static type method.

ODCIStatsExchangePartition()

Exchanges domain index statistics when an ALTER TABLE EXCHANGE PARTITION ... INCLULDING INDEXES command is issued.

Syntax

FUNCTION ODCIStatsExchangePartition(
   ia ODCIIndexInfo, 
   ia1 ODCIIndexInfo,
   env ODCIEnv) 
return NUMBER;

ia

sia1

env

Returns

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

Usage Notes

• This method should be implemented as a STATIC type.

• This method should be capable of converting the statistics associated with a domain index partition into statistics associated with a non-partitioned domain index, and the reverse. If the statistics are missing for one of the indexes or index partitions, the user should be able to delete these statistics.

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.

Syntax

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

ia

sel

cost

qi

pred

args

start

stop

list

env

Returns

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

Usage Notes

• 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 is 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 ODCIStatsIndexCost() function for queries on partitioned or non-partitioned tables. For queries on partitioned tables, additional information is passed in the ODCIStatsIndexCost() 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 are definitely not accessed.

• When the ODCIStatsIndexCost() 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.

Syntax

FUNCTION ODCIStatsSelectivity(
   pred ODCIPredInfo, 
   sel OUT NUMBER, 
   args ODCIArgDescList, 
   start function_return_type, 
   stop function_return_type, 
   list, 
   env ODCIEnv) 
return NUMBER;

Returns

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

Usage Notes

• 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 several tables, 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 several tables, 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.

Syntax

STATIC FUNCTION ODCIStatsTableFunction(
   func IN SYS.ODCIFuncInfo, 
   outStats OUT SYS.ODCITabFuncStats, 
   argDesc IN SYS.ODCIArgDescList, 
   list)
RETURN NUMBER;

Returns

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

ODCIStatsUpdPartStatistics()

Updates statistics during partition maintenance operations. This lets the statistics type patch up the domain index statistics to correctly reflect the partition maintenance operation.

Syntax

STATIC FUNCTION ODCIStatsCollect(
   ia ODCIIndexInfo,
   palist ODCIPartInfoList,
   env ODCIEnv) 
RETURN NUMBER

env

Returns

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

• When the statistics type is specified by the SYSTEM MANAGED approach, then the ODCIStatsUpdPartStatistics() method is invoked only one time during PMO. Only DML and query are allowed in the method implementation.

• If the user maintains the domain index statistics in a global non-partitioned table, then the user should delete the entry for the user-defined statistics for the dropped partition (and optionally add a NULL entry for added partition). They can then check if ODCIEnv.CallProperty is StatsGlobalAndPartition or StatsPartition. If ODCIEnv.CallProperty is StatsGlobalAndPartition then they should aggregate all the available index partition statistics. If ODCIEnv.CallProperty is StatsPartition they can simply delete the aggregate statistics, or leave the aggregate statistics as they are. ODCIEnv.CallProperty cannot be StatsGlobal for this call.

• The user should use the information passed in by the ODCIEnv.CallProperty to determine the type of statistics to delete and adjust.

• If the method returns ODCIConst.Error, the error is ignored and the partition management operation continues.