| Oracle® Database Data Warehousing Guide 11g Release 2 (11.2) Part Number E16579-02 |
|
|
PDF · Mobi · ePub |
| Oracle® Database Data Warehousing Guide 11g Release 2 (11.2) Part Number E16579-02 |
|
|
PDF · Mobi · ePub |
This chapter discusses how to load and refresh a data warehouse, and discusses:
ETL (Extraction, Transformation and Loading) is done on a scheduled basis to reflect changes made to the original source system. During this step, you physically insert the new, clean data into the production data warehouse schema, and take all of the other steps necessary (such as building indexes, validating constraints, taking backups) to make this new data available to the end users. Once all of this data has been loaded into the data warehouse, the materialized views have to be updated to reflect the latest data.
The partitioning scheme of the data warehouse is often crucial in determining the efficiency of refresh operations in the data warehouse load process. In fact, the load process is often the primary consideration in choosing the partitioning scheme of data warehouse tables and indexes.
The partitioning scheme of the largest data warehouse tables (for example, the fact table in a star schema) should be based upon the loading paradigm of the data warehouse.
Most data warehouses are loaded with new data on a regular schedule. For example, every night, week, or month, new data is brought into the data warehouse. The data being loaded at the end of the week or month typically corresponds to the transactions for the week or month. In this very common scenario, the data warehouse is being loaded by time. This suggests that the data warehouse tables should be partitioned on a date column. In our data warehouse example, suppose the new data is loaded into the sales table every month. Furthermore, the sales table has been partitioned by month. These steps show how the load process will proceed to add the data for a new month (January 2001) to the table sales.
Place the new data into a separate table, sales_01_2001. This data can be directly loaded into sales_01_2001 from outside the data warehouse, or this data can be the result of previous data transformation operations that have already occurred in the data warehouse. sales_01_2001 has the exact same columns, datatypes, and so forth, as the sales table. Gather statistics on the sales_01_2001 table.
Create indexes and add constraints on sales_01_2001. Again, the indexes and constraints on sales_01_2001 should be identical to the indexes and constraints on sales. Indexes can be built in parallel and should use the NOLOGGING and the COMPUTE STATISTICS options. For example:
CREATE BITMAP INDEX sales_01_2001_customer_id_bix
ON sales_01_2001(customer_id)
TABLESPACE sales_idx NOLOGGING PARALLEL 8 COMPUTE STATISTICS;
Apply all constraints to the sales_01_2001 table that are present on the sales table. This includes referential integrity constraints. A typical constraint would be:
ALTER TABLE sales_01_2001 ADD CONSTRAINT sales_customer_id
REFERENCES customer(customer_id) ENABLE NOVALIDATE;
If the partitioned table sales has a primary or unique key that is enforced with a global index structure, ensure that the constraint on sales_pk_jan01 is validated without the creation of an index structure, as in the following:
ALTER TABLE sales_01_2001 ADD CONSTRAINT sales_pk_jan01 PRIMARY KEY (sales_transaction_id) DISABLE VALIDATE;
The creation of the constraint with ENABLE clause would cause the creation of a unique index, which does not match a local index structure of the partitioned table. You must not have any index structure built on the nonpartitioned table to be exchanged for existing global indexes of the partitioned table. The exchange command would fail.
Add the sales_01_2001 table to the sales table.
In order to add this new data to the sales table, we must do two things. First, we must add a new partition to the sales table. We will use the ALTER TABLE ... ADD PARTITION statement. This will add an empty partition to the sales table:
ALTER TABLE sales ADD PARTITION sales_01_2001
VALUES LESS THAN (TO_DATE('01-FEB-2001', 'DD-MON-YYYY'));
Then, we can add our newly created table to this partition using the EXCHANGE PARTITION operation. This will exchange the new, empty partition with the newly loaded table.
ALTER TABLE sales EXCHANGE PARTITION sales_01_2001 WITH TABLE sales_01_2001 INCLUDING INDEXES WITHOUT VALIDATION UPDATE GLOBAL INDEXES;
The EXCHANGE operation will preserve the indexes and constraints that were already present on the sales_01_2001 table. For unique constraints (such as the unique constraint on sales_transaction_id), you can use the UPDATE GLOBAL INDEXES clause, as shown previously. This will automatically maintain your global index structures as part of the partition maintenance operation and keep them accessible throughout the whole process. If there were only foreign-key constraints, the exchange operation would be instantaneous.
The benefits of this partitioning technique are significant. First, the new data is loaded with minimal resource utilization. The new data is loaded into an entirely separate table, and the index processing and constraint processing are applied only to the new partition. If the sales table was 50 GB and had 12 partitions, then a new month's worth of data contains approximately 4 GB. Only the new month's worth of data must be indexed. None of the indexes on the remaining 46 GB of data must be modified at all. This partitioning scheme additionally ensures that the load processing time is directly proportional to the amount of new data being loaded, not to the total size of the sales table.
Second, the new data is loaded with minimal impact on concurrent queries. All of the operations associated with data loading are occurring on a separate sales_01_2001 table. Therefore, none of the existing data or indexes of the sales table is affected during this data refresh process. The sales table and its indexes remain entirely untouched throughout this refresh process.
Third, in case of the existence of any global indexes, those are incrementally maintained as part of the exchange command. This maintenance does not affect the availability of the existing global index structures.
The exchange operation can be viewed as a publishing mechanism. Until the data warehouse administrator exchanges the sales_01_2001 table into the sales table, end users cannot see the new data. Once the exchange has occurred, then any end user query accessing the sales table will immediately be able to see the sales_01_2001 data.
Partitioning is useful not only for adding new data but also for removing and archiving data. Many data warehouses maintain a rolling window of data. For example, the data warehouse stores the most recent 36 months of sales data. Just as a new partition can be added to the sales table (as described earlier), an old partition can be quickly (and independently) removed from the sales table. These two benefits (reduced resources utilization and minimal end-user impact) are just as pertinent to removing a partition as they are to adding a partition.
Removing data from a partitioned table does not necessarily mean that the old data is physically deleted from the database. There are two alternatives for removing old data from a partitioned table. First, you can physically delete all data from the database by dropping the partition containing the old data, thus freeing the allocated space:
ALTER TABLE sales DROP PARTITION sales_01_1998;
Also, you can exchange the old partition with an empty table of the same structure; this empty table is created equivalent to steps 1 and 2 described in the load process. Assuming the new empty table stub is named sales_archive_01_1998, the following SQL statement will empty partition sales_01_1998:
ALTER TABLE sales EXCHANGE PARTITION sales_01_1998 WITH TABLE sales_archive_01_1998 INCLUDING INDEXES WITHOUT VALIDATION UPDATE GLOBAL INDEXES;
Note that the old data is still existent as the exchanged, nonpartitioned table sales_archive_01_1998.
If the partitioned table was setup in a way that every partition is stored in a separate tablespace, you can archive (or transport) this table using Oracle Database's transportable tablespace framework before dropping the actual data (the tablespace). See "Transportation Using Transportable Tablespaces" for further details regarding transportable tablespaces.
In some situations, you might not want to drop the old data immediately, but keep it as part of the partitioned table; although the data is no longer of main interest, there are still potential queries accessing this old, read-only data. You can use Oracle's data compression to minimize the space usage of the old data. We also assume that at least one compressed partition is already part of the partitioned table.
See Also:
Chapter 3, "Physical Design in Data Warehouses" for a generic discussion of table compression
Oracle Database VLDB and Partitioning Guide for more information regarding partitioning and table compression
A typical scenario might not only need to compress old data, but also to merge several old partitions to reflect the granularity for a later backup of several merged partitions. Let us assume that a backup (partition) granularity is on a quarterly base for any quarter, where the oldest month is more than 36 months behind the most recent month. In this case, we are therefore compressing and merging sales_01_1998, sales_02_1998, and sales_03_1998 into a new, compressed partition sales_q1_1998.
Create the new merged partition in parallel in another tablespace. The partition will be compressed as part of the MERGE operation:
ALTER TABLE sales MERGE PARTITIONS sales_01_1998, sales_02_1998, sales_03_1998 INTO PARTITION sales_q1_1998 TABLESPACE archive_q1_1998 COMPRESS UPDATE GLOBAL INDEXES PARALLEL 4;
The partition MERGE operation invalidates the local indexes for the new merged partition. We therefore have to rebuild them:
ALTER TABLE sales MODIFY PARTITION sales_q1_1998 REBUILD UNUSABLE LOCAL INDEXES;
Alternatively, you can choose to create the new compressed table outside the partitioned table and exchange it back. The performance and the temporary space consumption is identical for both methods:
Create an intermediate table to hold the new merged information. The following statement inherits all NOT NULL constraints from the original table by default:
CREATE TABLE sales_q1_1998_out TABLESPACE archive_q1_1998
NOLOGGING COMPRESS PARALLEL 4 AS SELECT * FROM sales
WHERE time_id >= TO_DATE('01-JAN-1998','dd-mon-yyyy')
AND time_id < TO_DATE('01-APR-1998','dd-mon-yyyy');
Create the equivalent index structure for table sales_q1_1998_out than for the existing table sales.
Prepare the existing table sales for the exchange with the new compressed table sales_q1_1998_out. Because the table to be exchanged contains data actually covered in three partitions, we have to create one matching partition, having the range boundaries we are looking for. You simply have to drop two of the existing partitions. Note that you have to drop the lower two partitions sales_01_1998 and sales_02_1998; the lower boundary of a range partition is always defined by the upper (exclusive) boundary of the previous partition:
ALTER TABLE sales DROP PARTITION sales_01_1998; ALTER TABLE sales DROP PARTITION sales_02_1998;
You can now exchange table sales_q1_1998_out with partition sales_03_1998. Unlike what the name of the partition suggests, its boundaries cover Q1-1998.
ALTER TABLE sales EXCHANGE PARTITION sales_03_1998 WITH TABLE sales_q1_1998_out INCLUDING INDEXES WITHOUT VALIDATION UPDATE GLOBAL INDEXES;
Both methods apply to slightly different business scenarios: Using the MERGE PARTITION approach invalidates the local index structures for the affected partition, but it keeps all data accessible all the time. Any attempt to access the affected partition through one of the unusable index structures raises an error. The limited availability time is approximately the time for re-creating the local bitmap index structures. In most cases, this can be neglected, because this part of the partitioned table should not be accessed too often.
The CTAS approach, however, minimizes unavailability of any index structures close to zero, but there is a specific time window, where the partitioned table does not have all the data, because we dropped two partitions. The limited availability time is approximately the time for exchanging the table. Depending on the existence and number of global indexes, this time window varies. Without any existing global indexes, this time window is a matter of a fraction to few seconds.
These examples are a simplification of the data warehouse rolling window load scenario. Real-world data warehouse refresh characteristics are always more complex. However, the advantages of this rolling window approach are not diminished in more complex scenarios.
Note that before you add single or multiple compressed partitions to a partitioned table for the first time, all local bitmap indexes must be either dropped or marked unusable. After the first compressed partition is added, no additional actions are necessary for all subsequent operations involving compressed partitions. It is irrelevant how the compressed partitions are added to the partitioned table.
See Also:
Oracle Database VLDB and Partitioning Guide for more information regarding partitioning and table compression
Oracle Database Administrator's Guide for further details about partitioning and table compression.
This section contains two typical scenarios where partitioning is used with refresh.
Data is loaded daily. However, the data warehouse contains two years of data, so that partitioning by day might not be desired.
The solution is to partition by week or month (as appropriate). Use INSERT to add the new data to an existing partition. The INSERT operation only affects a single partition, so the benefits described previously remain intact. The INSERT operation could occur while the partition remains a part of the table. Inserts into a single partition can be parallelized:
INSERT /*+ APPEND*/ INTO sales PARTITION (sales_01_2001) SELECT * FROM new_sales;
The indexes of this sales partition will be maintained in parallel as well. An alternative is to use the EXCHANGE operation. You can do this by exchanging the sales_01_2001 partition of the sales table and then using an INSERT operation. You might prefer this technique when dropping and rebuilding indexes is more efficient than maintaining them.
New data feeds, although consisting primarily of data for the most recent day, week, and month, also contain some data from previous time periods.
Use parallel SQL operations (such as CREATE TABLE ... AS SELECT) to separate the new data from the data in previous time periods. Process the old data separately using other techniques.
New data feeds are not solely time based. You can also feed new data into a data warehouse with data from multiple operational systems on a business need basis. For example, the sales data from direct channels may come into the data warehouse separately from the data from indirect channels. For business reasons, it may furthermore make sense to keep the direct and indirect data in separate partitions.
Oracle supports composite range-list partitioning. The primary partitioning strategy of the sales table could be range partitioning based on time_id as shown in the example. However, the subpartitioning is a list based on the channel attribute. Each subpartition can now be loaded independently of each other (for each distinct channel) and added in a rolling window operation as discussed before. The partitioning strategy addresses the business needs in the most optimal manner.
You can optimize DML performance through the following techniques:
Commonly, the data that is extracted from a source system is not simply a list of new records that needs to be inserted into the data warehouse. Instead, this new data set is a combination of new records as well as modified records. For example, suppose that most of data extracted from the OLTP systems will be new sales transactions. These records are inserted into the warehouse's sales table, but some records may reflect modifications of previous transactions, such as returned merchandise or transactions that were incomplete or incorrect when initially loaded into the data warehouse. These records require updates to the sales table.
As a typical scenario, suppose that there is a table called new_sales that contains both inserts and updates that are applied to the sales table. When designing the entire data warehouse load process, it was determined that the new_sales table would contain records with the following semantics:
If a given sales_transaction_id of a record in new_sales already exists in sales, then update the sales table by adding the sales_dollar_amount and sales_quantity_sold values from the new_sales table to the existing row in the sales table.
Otherwise, insert the entire new record from the new_sales table into the sales table.
This UPDATE-ELSE-INSERT operation is often called a merge. A merge can be executed using one SQL statement.
MERGE INTO sales s USING new_sales n ON (s.sales_transaction_id = n.sales_transaction_id) WHEN MATCHED THEN UPDATE SET s.sales_quantity_sold = s.sales_quantity_sold + n.sales_quantity_sold, s.sales_dollar_amount = s.sales_dollar_amount + n.sales_dollar_amount WHEN NOT MATCHED THEN INSERT (sales_transaction_id, sales_quantity_sold, sales_dollar_amount) VALUES (n.sales_transcation_id, n.sales_quantity_sold, n.sales_dollar_amount);
In addition to using the MERGE statement for unconditional UPDATE ELSE INSERT functionality into a target table, you can also use it to:
Perform an UPDATE only or INSERT only statement.
Apply additional WHERE conditions for the UPDATE or INSERT portion of the MERGE statement.
The UPDATE operation can even delete rows if a specific condition yields true.
Example 16-2 Omitting the INSERT Clause
In some data warehouse applications, it is not allowed to add new rows to historical information, but only to update them. It may also happen that you do not want to update but only insert new information. The following example demonstrates INSERT-only with UPDATE-only functionality:
MERGE USING Product_Changes S -- Source/Delta table INTO Products D1 -- Destination table 1 ON (D1.PROD_ID = S.PROD_ID) -- Search/Join condition WHEN MATCHED THEN UPDATE -- update if join SET D1.PROD_STATUS = S.PROD_NEW_STATUS
Example 16-3 Omitting the UPDATE Clause
The following statement illustrates an example of omitting an UPDATE:
MERGE USING New_Product S -- Source/Delta table INTO Products D2 -- Destination table 2 ON (D2.PROD_ID = S.PROD_ID) -- Search/Join condition WHEN NOT MATCHED THEN -- insert if no join INSERT (PROD_ID, PROD_STATUS) VALUES (S.PROD_ID, S.PROD_NEW_STATUS)
When the INSERT clause is omitted, Oracle performs a regular join of the source and the target tables. When the UPDATE clause is omitted, Oracle performs an antijoin of the source and the target tables. This makes the join between the source and target table more efficient.
Example 16-4 Skipping the UPDATE Clause
In some situations, you may want to skip the UPDATE operation when merging a given row into the table. In this case, you can use an optional WHERE clause in the UPDATE clause of the MERGE. As a result, the UPDATE operation only executes when a given condition is true. The following statement illustrates an example of skipping the UPDATE operation:
MERGE USING Product_Changes S -- Source/Delta table INTO Products P -- Destination table 1 ON (P.PROD_ID = S.PROD_ID) -- Search/Join condition WHEN MATCHED THEN UPDATE -- update if join SET P.PROD_LIST_PRICE = S.PROD_NEW_PRICE WHERE P.PROD_STATUS <> "OBSOLETE" -- Conditional UPDATE
This shows how the UPDATE operation would be skipped if the condition P.PROD_STATUS <> "OBSOLETE" is not true. The condition predicate can refer to both the target and the source table.
Example 16-5 Conditional Inserts with MERGE Statements
You may want to skip the INSERT operation when merging a given row into the table. So an optional WHERE clause is added to the INSERT clause of the MERGE. As a result, the INSERT operation only executes when a given condition is true. The following statement offers an example:
MERGE USING Product_Changes S -- Source/Delta table INTO Products P -- Destination table 1 ON (P.PROD_ID = S.PROD_ID) -- Search/Join condition WHEN MATCHED THEN UPDATE -- update if join SET P.PROD_LIST_PRICE = S.PROD_NEW_PRICE WHERE P.PROD_STATUS <> "OBSOLETE" -- Conditional WHEN NOT MATCHED THEN INSERT (PROD_ID, PROD_STATUS, PROD_LIST_PRICE) -- insert if not join VALUES (S.PROD_ID, S.PROD_NEW_STATUS, S.PROD_NEW_PRICE) WHERE S.PROD_STATUS <> "OBSOLETE"; -- Conditional INSERT
This example shows that the INSERT operation would be skipped if the condition S.PROD_STATUS <> "OBSOLETE" is not true, and INSERT will only occur if the condition is true. The condition predicate can refer to the source table only. The condition predicate can only refer to the source table.
Example 16-6 Using the DELETE Clause with MERGE Statements
You may want to cleanse tables while populating or updating them. To do this, you may want to consider using the DELETE clause in a MERGE statement, as in the following example:
MERGE USING Product_Changes S INTO Products D ON (D.PROD_ID = S.PROD_ID) WHEN MATCHED THEN UPDATE SET D.PROD_LIST_PRICE =S.PROD_NEW_PRICE, D.PROD_STATUS = S.PROD_NEWSTATUS DELETE WHERE (D.PROD_STATUS = "OBSOLETE") WHEN NOT MATCHED THEN INSERT (PROD_ID, PROD_LIST_PRICE, PROD_STATUS) VALUES (S.PROD_ID, S.PROD_NEW_PRICE, S.PROD_NEW_STATUS);
Thus when a row is updated in products, Oracle checks the delete condition D.PROD_STATUS = "OBSOLETE", and deletes the row if the condition yields true.
The DELETE operation is not as same as that of a complete DELETE statement. Only the rows from the destination of the MERGE can be deleted. The only rows that are affected by the DELETE are the ones that are updated by this MERGE statement. Thus, although a given row of the destination table meets the delete condition, if it does not join under the ON clause condition, it is not deleted.
Example 16-7 Unconditional Inserts with MERGE Statements
You may want to insert all of the source rows into a table. In this case, the join between the source and target table can be avoided. By identifying special constant join conditions that always result to FALSE, for example, 1=0, such MERGE statements will be optimized and the join condition will be suppressed.
MERGE USING New_Product S -- Source/Delta table INTO Products P -- Destination table 1 ON (1 = 0) -- Search/Join condition WHEN NOT MATCHED THEN -- insert if no join INSERT (PROD_ID, PROD_STATUS) VALUES (S.PROD_ID, S.PROD_NEW_STATUS)
In some data warehousing environments, you might want to insert new data into tables in order to guarantee referential integrity. For example, a data warehouse may derive sales from an operational system that retrieves data directly from cash registers. sales is refreshed nightly. However, the data for the product dimension table may be derived from a separate operational system. The product dimension table may only be refreshed once for each week, because the product table changes relatively slowly. If a new product was introduced on Monday, then it is possible for that product's product_id to appear in the sales data of the data warehouse before that product_id has been inserted into the data warehouses product table.
Although the sales transactions of the new product may be valid, this sales data will not satisfy the referential integrity constraint between the product dimension table and the sales fact table. Rather than disallow the new sales transactions, you might choose to insert the sales transactions into the sales table. However, you might also wish to maintain the referential integrity relationship between the sales and product tables. This can be accomplished by inserting new rows into the product table as placeholders for the unknown products.
As in previous examples, we assume that the new data for the sales table is staged in a separate table, new_sales. Using a single INSERT statement (which can be parallelized), the product table can be altered to reflect the new products:
INSERT INTO product (SELECT sales_product_id, 'Unknown Product Name', NULL, NULL ... FROM new_sales WHERE sales_product_id NOT IN (SELECT product_id FROM product));
Occasionally, it is necessary to remove large amounts of data from a data warehouse. A very common scenario is the rolling window discussed previously, in which older data is rolled out of the data warehouse to make room for new data.
However, sometimes other data might need to be removed from a data warehouse. Suppose that a retail company has previously sold products from XYZ Software, and that XYZ Software has subsequently gone out of business. The business users of the warehouse may decide that they are no longer interested in seeing any data related to XYZ Software, so this data should be deleted.
One approach to removing a large volume of data is to use parallel delete as shown in the following statement:
DELETE FROM sales WHERE sales_product_id IN (SELECT product_id FROM product WHERE product_category = 'XYZ Software');
This SQL statement spawns one parallel process for each partition. This approach is much more efficient than a serial DELETE statement, and none of the data in the sales table will need to be moved. However, this approach also has some disadvantages. When removing a large percentage of rows, the DELETE statement will leave many empty row-slots in the existing partitions. If new data is being loaded using a rolling window technique (or is being loaded using direct-path INSERT or load), then this storage space will not be reclaimed. Moreover, even though the DELETE statement is parallelized, there might be more efficient methods. An alternative method is to re-create the entire sales table, keeping the data for all product categories except XYZ Software.
CREATE TABLE sales2 AS SELECT * FROM sales, product WHERE sales.sales_product_id = product.product_id AND product_category <> 'XYZ Software' NOLOGGING PARALLEL (DEGREE 8) #PARTITION ... ; #create indexes, constraints, and so on DROP TABLE SALES; RENAME SALES2 TO SALES;
This approach may be more efficient than a parallel delete. However, it is also costly in terms of the amount of disk space, because the sales table must effectively be instantiated twice.
An alternative method to utilize less space is to re-create the sales table one partition at a time:
CREATE TABLE sales_temp AS SELECT * FROM sales WHERE 1=0; INSERT INTO sales_temp SELECT * FROM sales PARTITION (sales_99jan), product WHERE sales.sales_product_id = product.product_id AND product_category <> 'XYZ Software'; <create appropriate indexes and constraints on sales_temp> ALTER TABLE sales EXCHANGE PARTITION sales_99jan WITH TABLE sales_temp;
Continue this process for each partition in the sales table.
When creating a materialized view, you have the option of specifying whether the refresh occurs ON DEMAND or ON COMMIT. In the case of ON COMMIT, the materialized view is changed every time a transaction commits, thus ensuring that the materialized view always contains the latest data. Alternatively, you can control the time when refresh of the materialized views occurs by specifying ON DEMAND. In this case, the materialized view can only be refreshed by calling one of the procedures in the DBMS_MVIEW package.
DBMS_MVIEW provides three different types of refresh operations.
Refresh one or more materialized views.
Refresh all materialized views.
Refresh all materialized views that depend on a specified master table or materialized view or list of master tables or materialized views.
See Also:
"Manual Refresh Using the DBMS_MVIEW Package" for more informationPerforming a refresh operation requires temporary space to rebuild the indexes and can require additional space for performing the refresh operation itself. Some sites might prefer not to refresh all of their materialized views at the same time: as soon as some underlying detail data has been updated, all materialized views using this data will become stale. Therefore, if you defer refreshing your materialized views, you can either rely on your chosen rewrite integrity level to determine whether or not a stale materialized view can be used for query rewrite, or you can temporarily disable query rewrite with an ALTER SYSTEM SET QUERY_REWRITE_ENABLED = FALSE statement. After refreshing the materialized views, you can re-enable query rewrite as the default for all sessions in the current database instance by specifying ALTER SYSTEM SET QUERY_REWRITE_ENABLED as TRUE. Refreshing a materialized view automatically updates all of its indexes. In the case of full refresh, this requires temporary sort space to rebuild all indexes during refresh. This is because the full refresh truncates or deletes the table before inserting the new full data volume. If insufficient temporary space is available to rebuild the indexes, then you must explicitly drop each index or mark it UNUSABLE prior to performing the refresh operation.
If you anticipate performing insert, update or delete operations on tables referenced by a materialized view concurrently with the refresh of that materialized view, and that materialized view includes joins and aggregation, Oracle recommends you use ON COMMIT fast refresh rather than ON DEMAND fast refresh.
See Also:
Oracle OLAP User's Guide for information regarding the refresh of cube organized materialized viewsA complete refresh occurs when the materialized view is initially defined as BUILD IMMEDIATE, unless the materialized view references a prebuilt table. For materialized views using BUILD DEFERRED, a complete refresh must be requested before it can be used for the first time. A complete refresh may be requested at any time during the life of any materialized view. The refresh involves reading the detail tables to compute the results for the materialized view. This can be a very time-consuming process, especially if there are huge amounts of data to be read and processed. Therefore, you should always consider the time required to process a complete refresh before requesting it.
There are, however, cases when the only refresh method available for an already built materialized view is complete refresh because the materialized view does not satisfy the conditions specified in the following section for a fast refresh.
Most data warehouses have periodic incremental updates to their detail data. As described in "Materialized View Schema Design", you can use the SQL*Loader or any bulk load utility to perform incremental loads of detail data. Fast refresh of your materialized views is usually efficient, because instead of having to recompute the entire materialized view, the changes are applied to the existing data. Thus, processing only the changes can result in a very fast refresh time.
When there have been some partition maintenance operations on the detail tables, this is the only method of fast refresh that can be used. PCT-based refresh on a materialized view is enabled only if all the conditions described in "Partition Change Tracking" are satisfied.
In the absence of partition maintenance operations on detail tables, when you request a FAST method (method => 'F') of refresh through procedures in DBMS_MVIEW package, Oracle uses a heuristic rule to try log-based rule fast refresh before choosing PCT refresh. Similarly, when you request a FORCE method (method => '?'), Oracle chooses the refresh method based on the following attempt order: log-based fast refresh, PCT refresh, and complete refresh. Alternatively, you can request the PCT method (method => 'P'), and Oracle uses the PCT method provided all PCT requirements are satisfied.
Oracle can use TRUNCATE PARTITION on a materialized view if it satisfies the conditions in "Benefits of Partitioning a Materialized View" and hence, make the PCT refresh process more efficient.
A materialized view can be refreshed automatically using the ON COMMIT method. Therefore, whenever a transaction commits which has updated the tables on which a materialized view is defined, those changes are automatically reflected in the materialized view. The advantage of using this approach is you never have to remember to refresh the materialized view. The only disadvantage is the time required to complete the commit will be slightly longer because of the extra processing involved. However, in a data warehouse, this should not be an issue because there is unlikely to be concurrent processes trying to update the same table.
When a materialized view is refreshed ON DEMAND, one of four refresh methods can be specified as shown in the following table. You can define a default option during the creation of the materialized view. Table 16-1 details the refresh options.
Table 16-1 ON DEMAND Refresh Methods
| Refresh Option | Parameter | Description |
|---|---|---|
|
|
|
Refreshes by recalculating the defining query of the materialized view. |
|
|
|
Refreshes by incrementally applying changes to the materialized view. For local materialized views, it chooses the refresh method which is estimated by optimizer to be most efficient. The refresh methods considered are log-based |
|
|
|
Refreshes by recomputing the rows in the materialized view affected by changed partitions in the detail tables. |
|
|
|
Attempts a fast refresh. If that is not possible, it does a complete refresh. For local materialized views, it chooses the refresh method which is estimated by optimizer to be most efficient. The refresh methods considered are log based |
Three refresh procedures are available in the DBMS_MVIEW package for performing ON DEMAND refresh. Each has its own unique set of parameters.
See Also:
Oracle Database PL/SQL Packages and Types Reference for detailed information about theDBMS_MVIEW package and Oracle Database Advanced Replication for information showing how to use it in a replication environmentUse the DBMS_MVIEW.REFRESH procedure to refresh one or more materialized views. Some parameters are used only for replication, so they are not mentioned here. The required parameters to use this procedure are:
The comma-delimited list of materialized views to refresh
The refresh method: F-Fast, P-Fast_PCT, ?-Force, C-Complete
The rollback segment to use
Refresh after errors (TRUE or FALSE)
A Boolean parameter. If set to TRUE, the number_of_failures output parameter is set to the number of refreshes that failed, and a generic error message indicates that failures occurred. The alert log for the instance gives details of refresh errors. If set to FALSE, the default, then refresh will stop after it encounters the first error, and any remaining materialized views in the list are not be refreshed.
The following four parameters are used by the replication process. For warehouse refresh, set them to FALSE, 0,0,0.
Atomic refresh (TRUE or FALSE)
If set to TRUE, then all refreshes are done in one transaction. If set to FALSE, then the refresh of each specified materialized view is done in a separate transaction. If set to FALSE, Oracle can optimize refresh by using parallel DML and truncate DDL on a materialized views. When a materialized view is refreshed in atomic mode, it is eligible for query rewrite if the rewrite integrity mode is set to stale_tolerated. Atomic refresh cannot be guaranteed when refresh is performed on nested views.
For example, to perform a fast refresh on the materialized view cal_month_sales_mv, the DBMS_MVIEW package would be called as follows:
DBMS_MVIEW.REFRESH('CAL_MONTH_SALES_MV', 'F', '', TRUE, FALSE, 0,0,0, FALSE);
Multiple materialized views can be refreshed at the same time, and they do not all have to use the same refresh method. To give them different refresh methods, specify multiple method codes in the same order as the list of materialized views (without commas). For example, the following specifies that cal_month_sales_mv be completely refreshed and fweek_pscat_sales_mv receive a fast refresh:
DBMS_MVIEW.REFRESH('CAL_MONTH_SALES_MV, FWEEK_PSCAT_SALES_MV', 'CF', '',
TRUE, FALSE, 0,0,0, FALSE);
If the refresh method is not specified, the default refresh method as specified in the materialized view definition is used.
An alternative to specifying the materialized views to refresh is to use the procedure DBMS_MVIEW.REFRESH_ALL_MVIEWS. This procedure refreshes all materialized views. If any of the materialized views fails to refresh, then the number of failures is reported.
The parameters for this procedure are:
The number of failures (this is an OUT variable)
The refresh method: F-Fast, P-Fast_PCT, ?-Force, C-Complete
Refresh after errors (TRUE or FALSE)
A Boolean parameter. If set to TRUE, the number_of_failures output parameter is set to the number of refreshes that failed, and a generic error message indicates that failures occurred. The alert log for the instance gives details of refresh errors. If set to FALSE, the default, then refresh stops after it encounters the first error, and any remaining materialized views in the list is not refreshed.
Atomic refresh (TRUE or FALSE)
If set to TRUE, then all refreshes are done in one transaction. If set to FALSE, then the refresh of each specified materialized view is done in a separate transaction. If set to FALSE, Oracle can optimize refresh by using parallel DML and truncate DDL on a materialized views. When a materialized view is refreshed in atomic mode, it is eligible for query rewrite if the rewrite integrity mode is set to stale_tolerated. Atomic refresh cannot be guaranteed when refresh is performed on nested views.
An example of refreshing all materialized views is the following:
DBMS_MVIEW.REFRESH_ALL_MVIEWS(failures,'C','', TRUE, FALSE);
The third procedure, DBMS_MVIEW.REFRESH_DEPENDENT, refreshes only those materialized views that depend on a specific table or list of tables. For example, suppose the changes have been received for the orders table but not for customer payments. The refresh dependent procedure can be called to refresh only those materialized views that reference the orders table.
The parameters for this procedure are:
The number of failures (this is an OUT variable)
The dependent table
The refresh method: F-Fast, P-Fast_PCT, ?-Force, C-Complete
The rollback segment to use
Refresh after errors (TRUE or FALSE)
A Boolean parameter. If set to TRUE, the number_of_failures output parameter is set to the number of refreshes that failed, and a generic error message indicates that failures occurred. The alert log for the instance gives details of refresh errors. If set to FALSE, the default, then refresh stops after it encounters the first error, and any remaining materialized views in the list are not refreshed.
Atomic refresh (TRUE or FALSE)
If set to TRUE, then all refreshes are done in one transaction. If set to FALSE, then the refresh of each specified materialized view is done in a separate transaction. If set to FALSE, Oracle can optimize refresh by using parallel DML and truncate DDL on a materialized views. When a materialized view is refreshed in atomic mode, it is eligible for query rewrite if the rewrite integrity mode is set to stale_tolerated. Atomic refresh cannot be guaranteed when refresh is performed on nested views.
Whether it is nested or not
If set to TRUE, refresh all the dependent materialized views of the specified set of tables based on a dependency order to ensure the materialized views are truly fresh with respect to the underlying base tables.
To perform a full refresh on all materialized views that reference the customers table, specify:
DBMS_MVIEW.REFRESH_DEPENDENT(failures, 'CUSTOMERS', 'C', '', FALSE, FALSE );
Job queues can be used to refresh multiple materialized views in parallel. If queues are not available, fast refresh sequentially refreshes each view in the foreground process. To make queues available, you must set the JOB_QUEUE_PROCESSES parameter. This parameter defines the number of background job queue processes and determines how many materialized views can be refreshed concurrently. Oracle tries to balance the number of concurrent refreshes with the degree of parallelism of each refresh. The order in which the materialized views are refreshed is determined by dependencies imposed by nested materialized views and potential for efficient refresh by using query rewrite against other materialized views (See "Scheduling Refresh" for details). This parameter is only effective when atomic_refresh is set to FALSE.
If the process that is executing DBMS_MVIEW.REFRESH is interrupted or the instance is shut down, any refresh jobs that were executing in job queue processes will be requeued and will continue running. To remove these jobs, use the DBMS_JOB.REMOVE procedure.
Not all materialized views may be fast refreshable. Therefore, use the package DBMS_MVIEW.EXPLAIN_MVIEW to determine what refresh methods are available for a materialized view. See Chapter 9, "Basic Materialized Views" for further information about the DBMS_MVIEW package.
If you are not sure how to make a materialized view fast refreshable, you can use the DBMS_ADVISOR.TUNE_MVIEW procedure, which will provide a script containing the statements required to create a fast refreshable materialized view. See the Oracle Database Performance Tuning Guide for more information.
The following initialization parameters need to be set properly for parallelism to be effective:
PARALLEL_MAX_SERVERS should be set high enough to take care of parallelism. You must consider the number of slaves needed for the refresh statement. For example, with a degree of parallelism of eight, you need 16 slave processes.
PGA_AGGREGATE_TARGET should be set for the instance to manage the memory usage for sorts and joins automatically. If the memory parameters are set manually, SORT_AREA_SIZE should be less than HASH_AREA_SIZE.
Remember to analyze all tables and indexes for better optimization.
See Also:
Oracle Database VLDB and Partitioning Guide for further informationWhile a job is running, you can query the V$SESSION_LONGOPS view to tell you the progress of each materialized view being refreshed.
SELECT * FROM V$SESSION_LONGOPS;
To look at the progress of which jobs are on which queue, use:
SELECT * FROM DBA_JOBS_RUNNING;
Three views are provided for checking the status of a materialized view: DBA_MVEIWS, ALL_MVIEWS, and USER_MVIEWS. To check if a materialized view is fresh or stale, issue the following statement:
SELECT MVIEW_NAME, STALENESS, LAST_REFRESH_TYPE, COMPILE_STATE FROM USER_MVIEWS ORDER BY MVIEW_NAME; MVIEW_NAME STALENESS LAST_REF COMPILE_STATE ---------- --------- -------- ------------- CUST_MTH_SALES_MV NEEDS_COMPILE FAST NEEDS_COMPILE PROD_YR_SALES_MV FRESH FAST VALID
If the compile_state column shows NEEDS COMPILE, the other displayed column values cannot be trusted as reflecting the true status. To revalidate the materialized view, issue the following statement:
ALTER MATERIALIZED VIEW [materialized_view_name] COMPILE;
Then reissue the SELECT statement.
Several views are available that enable you to verify the status of base table partitions and determine which ranges of materialized view data are fresh and which are stale. The views are as follows:
*_USER_MVIEWS
To determine Partition Change Tracking (PCT) information for the materialized view.
*_USER_MVIEW_DETAIL_RELATIONS
To display partition information for the detail table a materialized view is based on.
*_USER_MVIEW_DETAIL_PARTITION
To determine which partitions are fresh.
*_USER_MVIEW_DETAIL_SUBPARTITION
To determine which subpartitions are fresh.
The use of these views is illustrated in the following examples. Figure 16-1 illustrates a range-list partitioned table and a materialized view based on it. The partitions are P1, P2, P3, and P4, while the subpartitions are SP1, SP2, and SP3.
This section illustrates examples of determining the PCT and freshness information for materialized views and their detail tables.
Example 16-8 Verifying the PCT Status of a Materialized View
Query USER_MVIEWS to access PCT information about the materialized view, as shown in the following:
SELECT MVIEW_NAME, NUM_PCT_TABLES, NUM_FRESH_PCT_REGIONS,
NUM_STALE_PCT_REGIONS
FROM USER_MVIEWS
WHERE MVIEW_NAME = MV1;
MVIEW_NAME NUM_PCT_TABLES NUM_FRESH_PCT_REGIONS NUM_STALE_PCT_REGIONS
---------- -------------- --------------------- ---------------------
MV1 1 9 3
Example 16-9 Verifying the PCT Status in a Materialized View's Detail Table
Query USER_MVIEW_DETAIL_RELATIONS to access PCT detail table information, as shown in the following:
SELECT MVIEW_NAME, DETAILOBJ_NAME, DETAILOBJ_PCT, NUM_FRESH_PCT_PARTITIONS, NUM_STALE_PCT_PARTITIONS FROM USER_MVIEW_DETAIL_RELATIONS WHERE MVIEW_NAME = MV1;
MVIEW_NAME DETAILOBJ_NAME DETAIL_OBJ_PCT NUM_FRESH_PCT_PARTITIONS NUM_STALE_PCT_PARTITIONS
---------- -------------- -------------- ------------------------ ------------------------
MV1 T1 Y 3 1
Example 16-10 Verifying Which Partitions are Fresh
Query USER_MVIEW_DETAIL_PARTITION to access PCT freshness information for partitions, as shown in the following:
SELECT MVIEW_NAME,DETAILOBJ_NAME,DETAIL_PARTITION_NAME, DETAIL_PARTITION_POSITION,FRESHNESS FROM USER_MVIEW_DETAIL_PARTITION WHERE MVIEW_NAME = MV1;
MVIEW_NAME DETAILOBJ_NAME DETAIL_PARTITION_NAME DETAIL_PARTITION_POSITION FRESHNESS
---------- -------------- --------------------- ------------------------- ---------
MV1 T1 P1 1 FRESH
MV1 T1 P2 2 FRESH
MV1 T1 P3 3 STALE
MV1 T1 P4 4 FRESH
Example 16-11 Verifying Which Subpartitions are Fresh
Query USER_MVIEW_DETAIL_SUBPARTITION to access PCT freshness information for subpartitions, as shown in the following:
SELECT MVIEW_NAME,DETAILOBJ_NAME,DETAIL_PARTITION_NAME, DETAIL_SUBPARTITION_NAME,
DETAIL_SUBPARTITION_POSITION,FRESHNESS
FROM USER_MVIEW_DETAIL_SUBPARTITION
WHERE MVIEW_NAME = MV1;
MVIEW_NAME DETAILOBJ DETAIL_PARTITION DETAIL_SUBPARTITION_NAME DETAIL_SUBPARTITION_POS FRESHNESS
---------- --------- ---------------- ------------------------ ----------------------- ---------
MV1 T1 P1 SP1 1 FRESH
MV1 T1 P1 SP2 1 FRESH
MV1 T1 P1 SP3 1 FRESH
MV1 T1 P2 SP1 1 FRESH
MV1 T1 P2 SP2 1 FRESH
MV1 T1 P2 SP3 1 FRESH
MV1 T1 P3 SP1 1 STALE
MV1 T1 P3 SP2 1 STALE
MV1 T1 P3 SP3 1 STALE
MV1 T1 P4 SP1 1 FRESH
MV1 T1 P4 SP2 1 FRESH
MV1 T1 P4 SP3 1 FRESH
Very often you have multiple materialized views in the database. Some of these can be computed by rewriting against others. This is very common in data warehousing environment where you may have nested materialized views or materialized views at different levels of some hierarchy.
In such cases, you should create the materialized views as BUILD DEFERRED, and then issue one of the refresh procedures in DBMS_MVIEW package to refresh all the materialized views. Oracle Database computes the dependencies and refreshes the materialized views in the right order. Consider the example of a complete hierarchical cube described in "Examples of Hierarchical Cube Materialized Views". Suppose all the materialized views have been created as BUILD DEFERRED. Creating the materialized views as BUILD DEFERRED only creates the metadata for all the materialized views. And, then, you can just call one of the refresh procedures in DBMS_MVIEW package to refresh all the materialized views in the right order:
DECLARE numerrs PLS_INTEGER;
BEGIN DBMS_MVIEW.REFRESH_DEPENDENT (
number_of_failures => numerrs, list=>'SALES', method => 'C');
DBMS_OUTPUT.PUT_LINE('There were ' || numerrs || ' errors during refresh');
END;
/
The procedure refreshes the materialized views in the order of their dependencies (first sales_hierarchical_mon_cube_mv, followed by sales_hierarchical_qtr_cube_mv, then, sales_hierarchical_yr_cube_mv and finally, sales_hierarchical_all_cube_mv). Each of these materialized views gets rewritten against the one prior to it in the list).
The same kind of rewrite can also be used while doing PCT refresh. PCT refresh recomputes rows in a materialized view corresponding to changed rows in the detail tables. And, if there are other fresh materialized views available at the time of refresh, it can go directly against them as opposed to going against the detail tables.
Hence, it is always beneficial to pass a list of materialized views to any of the refresh procedures in DBMS_MVIEW package (irrespective of the method specified) and let the procedure figure out the order of doing refresh on materialized views.
Following are some guidelines for using the refresh mechanism for materialized views with aggregates.
For fast refresh, create materialized view logs on all detail tables involved in a materialized view with the ROWID, SEQUENCE and INCLUDING NEW VALUES clauses.
Include all columns from the table likely to be used in materialized views in the materialized view logs.
Fast refresh may be possible even if the SEQUENCE option is omitted from the materialized view log. If it can be determined that only inserts or deletes will occur on all the detail tables, then the materialized view log does not require the SEQUENCE clause. However, if updates to multiple tables are likely or required or if the specific update scenarios are unknown, make sure the SEQUENCE clause is included.
Use Oracle's bulk loader utility or direct-path INSERT (INSERT with the APPEND hint for loads).
This is a lot more efficient than conventional insert. During loading, disable all constraints and re-enable when finished loading. Note that materialized view logs are required regardless of whether you use direct load or conventional DML.
Try to optimize the sequence of conventional mixed DML operations, direct-path INSERT and the fast refresh of materialized views. You can use fast refresh with a mixture of conventional DML and direct loads. Fast refresh can perform significant optimizations if it finds that only direct loads have occurred, as illustrated in the following:
Direct-path INSERT (SQL*Loader or INSERT /*+ APPEND */) into the detail table
Refresh materialized view
Conventional mixed DML
Refresh materialized view
You can use fast refresh with conventional mixed DML (INSERT, UPDATE, and DELETE) to the detail tables. However, fast refresh will be able to perform significant optimizations in its processing if it detects that only inserts or deletes have been done to the tables, such as:
DML INSERT or DELETE to the detail table
Refresh materialized views
DML update to the detail table
Refresh materialized view
Even more optimal is the separation of INSERT and DELETE.
If possible, refresh should be performed after each type of data change (as shown earlier) rather than issuing only one refresh at the end. If that is not possible, restrict the conventional DML to the table to inserts only, to get much better refresh performance. Avoid mixing deletes and direct loads.
Furthermore, for refresh ON COMMIT, Oracle keeps track of the type of DML done in the committed transaction. Therefore, do not perform direct-path INSERT and DML to other tables in the same transaction, as Oracle may not be able to optimize the refresh phase.
For ON COMMIT materialized views, where refreshes automatically occur at the end of each transaction, it may not be possible to isolate the DML statements, in which case keeping the transactions short will help. However, if you plan to make numerous modifications to the detail table, it may be better to perform them in one transaction, so that refresh of the materialized view will be performed just once at commit time rather than after each update.
Oracle recommends partitioning the tables because it enables you to use:
Parallel DML
For large loads or refresh, enabling parallel DML helps shorten the length of time for the operation.
Partition Change Tracking (PCT) fast refresh
You can refresh your materialized views fast after partition maintenance operations on the detail tables. "Partition Change Tracking" for details on enabling PCT for materialized views.
Partitioning the materialized view also helps refresh performance as refresh can update the materialized view using parallel DML. For example, assume that the detail tables and materialized view are partitioned and have a parallel clause. The following sequence would enable Oracle to parallelize the refresh of the materialized view.
Bulk load into the detail table.
Enable parallel DML with an ALTER SESSION ENABLE PARALLEL DML statement.
Refresh the materialized view.
For refresh using DBMS_MVIEW.REFRESH, set the parameter atomic_refresh to FALSE.
For COMPLETE refresh, this will TRUNCATE to delete existing rows in the materialized view, which is faster than a delete.
For PCT refresh, if the materialized view is partitioned appropriately, this will use TRUNCATE PARTITION to delete rows in the affected partitions of the materialized view, which is faster than a delete.
For FAST or FORCE refresh, if COMPLETE or PCT refresh is chosen, this will be able to use the TRUNCATE optimizations described earlier.
When using DBMS_MVIEW.REFRESH with JOB_QUEUES, remember to set atomic to FALSE. Otherwise, JOB_QUEUES is not used. Set the number of job queue processes greater than the number of processors.
If job queues are enabled and there are many materialized views to refresh, it is faster to refresh all of them in a single command than to call them individually.
Use REFRESH FORCE to ensure refreshing a materialized view so that it can definitely be used for query rewrite. The best refresh method is chosen. If a fast refresh cannot be done, a complete refresh is performed.
Refresh all the materialized views in a single procedure call. This gives Oracle an opportunity to schedule refresh of all the materialized views in the right order taking into account dependencies imposed by nested materialized views and potential for efficient refresh by using query rewrite against other materialized views.
If a materialized view contains joins but no aggregates, then having an index on each of the join column rowids in the detail table will enhance refresh performance greatly, because this type of materialized view tends to be much larger than materialized views containing aggregates. For example, consider the following materialized view:
CREATE MATERIALIZED VIEW detail_fact_mv BUILD IMMEDIATE AS SELECT s.rowid "sales_rid", t.rowid "times_rid", c.rowid "cust_rid", c.cust_state_province, t.week_ending_day, s.amount_sold FROM sales s, times t, customers c WHERE s.time_id = t.time_id AND s.cust_id = c.cust_id;
Indexes should be created on columns sales_rid, times_rid and cust_rid. Partitioning is highly recommended, as is enabling parallel DML in the session before invoking refresh, because it will greatly enhance refresh performance.
This type of materialized view can also be fast refreshed if DML is performed on the detail table. It is recommended that the same procedure be applied to this type of materialized view as for a single table aggregate. That is, perform one type of change (direct-path INSERT or DML) and then refresh the materialized view. This is because Oracle Database can perform significant optimizations if it detects that only one type of change has been done.
Also, Oracle recommends that the refresh be invoked after each table is loaded, rather than load all the tables and then perform the refresh.
For refresh ON COMMIT, Oracle keeps track of the type of DML done in the committed transaction. Oracle therefore recommends that you do not perform direct-path and conventional DML to other tables in the same transaction because Oracle may not be able to optimize the refresh phase. For example, the following is not recommended:
Direct load new data into the fact table
DML into the store table
Commit
Also, try not to mix different types of conventional DML statements if possible. This would again prevent using various optimizations during fast refresh. For example, try to avoid the following:
Insert into the fact table
Delete from the fact table
Commit
If many updates are needed, try to group them all into one transaction because refresh will be performed just once at commit time, rather than after each update.
In a data warehousing environment, assuming that the materialized view has a parallel clause, the following sequence of steps is recommended:
Bulk load into the fact table
Enable parallel DML
An ALTER SESSION ENABLE PARALLEL DML statement
Refresh the materialized view
All underlying objects are treated as ordinary tables when refreshing materialized views. If the ON COMMIT refresh option is specified, then all the materialized views are refreshed in the appropriate order at commit time. In other words, Oracle builds a partially ordered set of materialized views and refreshes them such that, after the successful completion of the refresh, all the materialized views are fresh. The status of the materialized views can be checked by querying the appropriate USER_, DBA_, or ALL_MVIEWS view.
If any of the materialized views are defined as ON DEMAND refresh (irrespective of whether the refresh method is FAST, FORCE, or COMPLETE), you will need to refresh them in the correct order (taking into account the dependencies between the materialized views) because the nested materialized view will be refreshed with respect to the current contents of the other materialized views (whether fresh or not). This can be achieved by invoking the refresh procedure against the materialized view at the top of the nested hierarchy and specifying the nested parameter as TRUE.
If a refresh fails during commit time, the list of materialized views that has not been refreshed is written to the alert log, and you must manually refresh them along with all their dependent materialized views.
Use the same DBMS_MVIEW procedures on nested materialized views that you use on regular materialized views.
These procedures have the following behavior when used with nested materialized views:
If REFRESH is applied to a materialized view my_mv that is built on other materialized views, then my_mv will be refreshed with respect to the current contents of the other materialized views (that is, the other materialized views will not be made fresh first) unless you specify nested => TRUE.
If REFRESH_DEPENDENT is applied to materialized view my_mv, then only materialized views that directly depend on my_mv will be refreshed (that is, a materialized view that depends on a materialized view that depends on my_mv will not be refreshed) unless you specify nested => TRUE.
If REFRESH_ALL_MVIEWS is used, the order in which the materialized views will be refreshed is guaranteed to respect the dependencies between nested materialized views.
GET_MV_DEPENDENCIES provides a list of the immediate (or direct) materialized view dependencies for an object.
You can use fast refresh for materialized views that use the UNION ALL operator by providing a maintenance column in the definition of the materialized view. For example, a materialized view with a UNION ALL operator can be made fast refreshable as follows:
CREATE MATERIALIZED VIEW fast_rf_union_all_mv AS SELECT x.rowid AS r1, y.rowid AS r2, a, b, c, 1 AS marker FROM x, y WHERE x.a = y.b UNION ALL SELECT p.rowid, r.rowid, a, c, d, 2 AS marker FROM p, r WHERE p.a = r.y;
The form of a maintenance marker column, column MARKER in the example, must be numeric_or_string_literal AS column_alias, where each UNION ALL member has a distinct value for numeric_or_string_literal.
You can often improve fast refresh performance by ensuring that your materialized view logs on the base table contain a WITH COMMIT SCN clause, often significantly. By optimizing materialized view log processing WITH COMMIT SCN, the fast refresh process can save time. The following example illustrates how to use this clause:
CREATE MATERIALIZED VIEW LOG ON sales WITH ROWID (prod_id, cust_id, time_id, channel_id, promo_id, quantity_sold, amount_sold), COMMIT SCN INCLUDING NEW VALUES;
The materialized view refresh will then automatically use the commit SCN-based materialized view log to save refresh time.
Note that only new materialized view logs can take advantage of COMMIT SCN. Existing materialized view logs cannot be altered to add COMMIT SCN unless they are dropped and recreated.
When a materialized view is created on both base tables with timestamp-based materialized view logs and base tables with commit SCN-based materialized view logs, an error (ORA-32414) will be raised stating that materialized view logs are not compatible with each other for fast refresh.
After you have performed a load or incremental load and rebuilt the detail table indexes, you must re-enable integrity constraints (if any) and refresh the materialized views and materialized view indexes that are derived from that detail data. In a data warehouse environment, referential integrity constraints are normally enabled with the NOVALIDATE or RELY options. An important decision to make before performing a refresh operation is whether the refresh needs to be recoverable. Because materialized view data is redundant and can always be reconstructed from the detail tables, it might be preferable to disable logging on the materialized view. To disable logging and run incremental refresh non-recoverably, use the ALTER MATERIALIZED VIEW ... NOLOGGING statement prior to refreshing.
If the materialized view is being refreshed using the ON COMMIT method, then, following refresh operations, consult the alert log alert_SID.log and the trace file ora_SID_number.trc to check that no errors have occurred.
A major maintenance component of a data warehouse is synchronizing (refreshing) the materialized views when the detail data changes. Partitioning the underlying detail tables can reduce the amount of time taken to perform the refresh task. This is possible because partitioning enables refresh to use parallel DML to update the materialized view. Also, it enables the use of Partition Change Tracking.
In a data warehouse, changes to the detail tables can often entail partition maintenance operations, such as DROP, EXCHANGE, MERGE, and ADD PARTITION. To maintain the materialized view after such operations used to require manual maintenance (see also CONSIDER FRESH) or complete refresh. You now have the option of using an addition to fast refresh known as Partition Change Tracking (PCT) refresh.
For PCT to be available, the detail tables must be partitioned. The partitioning of the materialized view itself has no bearing on this feature. If PCT refresh is possible, it will occur automatically and no user intervention is required in order for it to occur. See "Partition Change Tracking" for PCT requirements.
The following examples illustrate the use of this feature. In "PCT Fast Refresh Scenario 1", assume sales is a partitioned table using the time_id column and products is partitioned by the prod_category column. The table times is not a partitioned table.
The following materialized view satisfies requirements for PCT.
CREATE MATERIALIZED VIEW cust_mth_sales_mv
BUILD IMMEDIATE
REFRESH FAST ON DEMAND
ENABLE QUERY REWRITE AS
SELECT s.time_id, s.prod_id, SUM(s.quantity_sold), SUM(s.amount_sold),
p.prod_name, t.calendar_month_name, COUNT(*),
COUNT(s.quantity_sold), COUNT(s.amount_sold)
FROM sales s, products p, times t
WHERE s.time_id = t.time_id AND s.prod_id = p.prod_id
GROUP BY t.calendar_month_name, s.prod_id, p.prod_name, s.time_id;
You can use the DBMS_MVIEW.EXPLAIN_MVIEW procedure to determine which tables will allow PCT refresh.
MVNAME CAPABILITY_NAME POSSIBLE RELATED_TEXT MSGTXT
----------------- --------------- -------- ------------ ----------------
CUST_MTH_SALES_MV PCT Y SALES
CUST_MTH_SALES_MV PCT_TABLE Y SALES
CUST_MTH_SALES_MV PCT_TABLE N PRODUCTS no partition key
or PMARKER
in SELECT list
CUST_MTH_SALES_MV PCT_TABLE N TIMES relation is not
partitionedtable
As can be seen from the partial sample output from EXPLAIN_MVIEW, any partition maintenance operation performed on the sales table will allow PCT fast refresh. However, PCT is not possible after partition maintenance operations or updates to the products table as there is insufficient information contained in cust_mth_sales_mv for PCT refresh to be possible. Note that the times table is not partitioned and hence can never allow for PCT refresh. Oracle Database will apply PCT refresh if it can determine that the materialized view has sufficient information to support PCT for all the updated tables. You can verify which partitions are fresh and stale with views such as DBA_MVIEWS and DBA_MVIEW_DETAIL_PARTITION.
See "Analyzing Materialized View Capabilities" for information on how to use this procedure and also some details regarding PCT-related views.
Suppose at some later point, a SPLIT operation of one partition in the sales table becomes necessary.
ALTER TABLE SALES
SPLIT PARTITION month3 AT (TO_DATE('05-02-1998', 'DD-MM-YYYY'))
INTO (PARTITION month3_1 TABLESPACE summ,
PARTITION month3 TABLESPACE summ);
Insert some data into the sales table.
Fast refresh cust_mth_sales_mv using the DBMS_MVIEW.REFRESH procedure.
EXECUTE DBMS_MVIEW.REFRESH('CUST_MTH_SALES_MV', 'F',
'',TRUE,FALSE,0,0,0,FALSE);
Fast refresh will automatically do a PCT refresh as it is the only fast refresh possible in this scenario. However, fast refresh will not occur if a partition maintenance operation occurs when any update has taken place to a table on which PCT is not enabled. This is shown in "PCT Fast Refresh Scenario 2".
"PCT Fast Refresh Scenario 1" would also be appropriate if the materialized view was created using the PMARKER clause as illustrated in the following:
CREATE MATERIALIZED VIEW cust_sales_marker_mv
BUILD IMMEDIATE
REFRESH FAST ON DEMAND
ENABLE QUERY REWRITE AS
SELECT DBMS_MVIEW.PMARKER(s.rowid) s_marker, SUM(s.quantity_sold),
SUM(s.amount_sold), p.prod_name, t.calendar_month_name, COUNT(*),
COUNT(s.quantity_sold), COUNT(s.amount_sold)
FROM sales s, products p, times t
WHERE s.time_id = t.time_id AND s.prod_id = p.prod_id
GROUP BY DBMS_MVIEW.PMARKER(s.rowid),
p.prod_name, t.calendar_month_name;
In "PCT Fast Refresh Scenario 2", the first three steps are the same as in "PCT Fast Refresh Scenario 1". Then, the SPLIT partition operation to the sales table is performed, but before the materialized view refresh occurs, records are inserted into the times table.
The same as in "PCT Fast Refresh Scenario 1".
The same as in "PCT Fast Refresh Scenario 1".
The same as in "PCT Fast Refresh Scenario 1".
After issuing the same SPLIT operation, as shown in "PCT Fast Refresh Scenario 1", some data will be inserted into the times table.
ALTER TABLE SALES
SPLIT PARTITION month3 AT (TO_DATE('05-02-1998', 'DD-MM-YYYY'))
INTO (PARTIITION month3_1 TABLESPACE summ,
PARTITION month3 TABLESPACE summ);
Refresh cust_mth_sales_mv.
EXECUTE DBMS_MVIEW.REFRESH('CUST_MTH_SALES_MV', 'F',
'',TRUE,FALSE,0,0,0,FALSE);
ORA-12052: cannot fast refresh materialized view SH.CUST_MTH_SALES_MV
The materialized view is not fast refreshable because DML has occurred to a table on which PCT fast refresh is not possible. To avoid this occurring, Oracle recommends performing a fast refresh immediately after any partition maintenance operation on detail tables for which partition tracking fast refresh is available.
If the situation in "PCT Fast Refresh Scenario 2" occurs, there are two possibilities; perform a complete refresh or switch to the CONSIDER FRESH option outlined in the following, if suitable. However, it should be noted that CONSIDER FRESH and partition change tracking fast refresh are not compatible. Once the ALTER MATERIALIZED VIEW cust_mth_sales_mv CONSIDER FRESH statement has been issued, PCT refresh will no longer be applied to this materialized view, until a complete refresh is done. Moreover, you should not use CONSIDER FRESH unless you have taken manual action to ensure that the materialized view is indeed fresh.
A common situation in a data warehouse is the use of rolling windows of data. In this case, the detail table and the materialized view may contain say the last 12 months of data. Every month, new data for a month is added to the table and the oldest month is deleted (or maybe archived). PCT refresh provides a very efficient mechanism to maintain the materialized view in this case.
The new data is usually added to the detail table by adding a new partition and exchanging it with a table containing the new data.
ALTER TABLE sales ADD PARTITION month_new ... ALTER TABLE sales EXCHANGE PARTITION month_new month_new_table
Next, the oldest partition is dropped or truncated.
ALTER TABLE sales DROP PARTITION month_oldest;
Now, if the materialized view satisfies all conditions for PCT refresh.
EXECUTE DBMS_MVIEW.REFRESH('CUST_MTH_SALES_MV', 'F', '', TRUE, FALSE,0,0,0,FALSE);
Fast refresh will automatically detect that PCT is available and perform a PCT refresh.
In a data warehouse, you may often wish to accumulate historical information in the materialized view even though this information is no longer in the detailed tables. In this case, you could maintain the materialized view using the ALTER MATERIALIZED VIEW materialized_view_name CONSIDER FRESH statement.
Note that CONSIDER FRESH declares that the contents of the materialized view are FRESH (in sync with the detail tables). Care must be taken when using this option in this scenario in conjunction with query rewrite because you may see unexpected results.
After using CONSIDER FRESH in an historical scenario, you will be able to apply traditional fast refresh after DML and direct loads to the materialized view, but not PCT fast refresh. This is because if the detail table partition at one time contained data that is currently kept in aggregated form in the materialized view, PCT refresh in attempting to resynchronize the materialized view with that partition could delete historical data which cannot be recomputed.
Assume the sales table stores the prior year's data and the cust_mth_sales_mv keeps the prior 10 years of data in aggregated form.
Remove old data from a partition in the sales table:
ALTER TABLE sales TRUNCATE PARTITION month1;
The materialized view is now considered stale and requires a refresh because of the partition operation. However, as the detail table no longer contains all the data associated with the partition fast refresh cannot be attempted.
Therefore, alter the materialized view to tell Oracle Database to consider it fresh.
ALTER MATERIALIZED VIEW cust_mth_sales_mv CONSIDER FRESH;
This statement informs Oracle Database that cust_mth_sales_mv is fresh for your purposes. However, the materialized view now has a status that is neither known fresh nor known stale. Instead, it is UNKNOWN. If the materialized view has query rewrite enabled in QUERY_REWRITE_INTEGRITY = stale_tolerated mode, it will be used for rewrite.
Insert data into sales.
Refresh the materialized view.
EXECUTE DBMS_MVIEW.REFRESH('CUST_MTH_SALES_MV', 'F', '', TRUE, FALSE,0,0,0,FALSE);
Because the fast refresh detects that only INSERT statements occurred against the sales table it will update the materialized view with the new data. However, the status of the materialized view will remain UNKNOWN. The only way to return the materialized view to FRESH status is with a complete refresh which, also will remove the historical data from the materialized view.
|
 Copyright © 2001, 2010, Oracle and/or its affiliates. All rights reserved. Legal Notices |
|