Logical and Physical Records
In Table 21-1, several of the keywords refer to "logical" rows. A logical row is a row that is inserted into the database. Depending on the structure of the input file, multiple physical rows may be combined to make a single logical row.
For example, the input file may look like this:
Good Record,Some Publisher, ADULTNF, 3
in which case there would be a one-to-one relationship between that physical record and the logical record it creates. But the datafile may look like this instead:
To combine the data, you need to use continuation rules. In this case, the column values are split one to a line, so there is a set number of physical records for each logical record. To combine them, use the concatenate
clause within the control file. In this case, you would specify concatenate 4
to create a single logical row from the four physical rows.
The logic for creating a single logical record from multiple physical records can be much more complex than a simple concatenation. You can use the continueif
clause to specify the conditions that cause logical records to be continued. You can further manipulate the input data to create multiple logical records from a single physical record (via the use of multiple into table
clauses). See the control file syntax in the "SQLLDR" entry of the Alphabetical Reference in this book and the notes in the following section.
You can use SQL*Loader to generate multiple inserts from a single physical row (similar to the multitable insert capability described in Chapter 15). For example, suppose the input data is denormalized, with fields City and Rainfall, while the input data is in the format City, Rainfall1, Rainfall2, Rainfall3. The control file would resemble the following (depending on the actual physical stop and start positions of the data in the file):
into table RAINFALL
when City != ' '
(City POSITION(1:5) CHAR,
Rainfall POSITION(6:10) INTEGER EXTERNAL) -- 1st row
into table RAINFALL
when City != ' '
(City POSITION(1:5) CHAR,
Rainfall POSITION(11:16) INTEGER EXTERNAL) -- 2nd row
into table RAINFALL
when City != ' '
(City POSITION(1:5) CHAR,
Rainfall POSITION(16:21) INTEGER EXTERNAL) -- 3rd row
Note that separate into table
clauses operate on each physical row. In this example, they generate separate rows in the RAINFALL table; they could also be used to insert rows into multiple tables.
Control File Syntax Notes
The full syntax for SQL*Loader control files is shown in the "SQLLDR" entry in the Alphabetical Reference, so it is not repeated here.
Within the load
clause, you can specify that the load is recoverable
. The unrecoverable
clause only applies to Direct Path loading, and is described in "Tuning Data Loads" later in this chapter.
In addition to using the concatenate
clause, you can use the continueif
clause to control the manner in which physical records are assembled into logical records. The this
clause refers to the current physical record, while the next
clause refers to the next physical record. For example, you could create a two-character continuation character at the start of each physical record. If that record should be concatenated to the preceding record, set that value equal to '**'
. You could then use the continueif next (1:2)='**'
clause to create a single logical record from the multiple physical records. The '**'
continuation character will not be part of the merged record.
The syntax for the into table
clause includes a when
clause. The when
clause, shown in the following listing, serves as a filter applied to rows prior to their insertion into the table. For example, you can specify
to load only books with ratings greater than 3 into the table. Any row that does not pass the when condition will be written to the discard file. Thus, the discard file contains rows that can be used for later loads, but that did not pass the current set of when conditions. You can use multiple when conditions, connected with and clauses.
Use the trailing nullcols
clause if you are loading variable-length records for which the last column does not always have a value. With this clause in effect, SQL*Loader will generate NULL
values for those columns.
As shown in an example earlier in this chapter, you can use the fields terminated by
clause to load variable-length data. Rather than being terminated by a character, the fields can be terminated by whitespace
or enclosed by
characters or optionally enclosed by
For example, the following entry loads AuthorName values and sets the values to uppercase during the insert. If the value is blank, a NULL
AuthorName POSITION(10:34) CHAR TERMINATED BY WHITESPACE
NULLIF AuthorName=BLANKS "UPPER(:AuthorName)"
When you load DATE datatype values, you can specify a date mask. For example, if you had a column named ReturnDate and the incoming data is in the format Mon-DD-YYYY in the first 11 places of the record, you could specify the ReturnDate portion of the load as follows:
ReturnDate POSITION (1:11) DATE "Mon-DD-YYYY"
Within the into table
clause, you can use the recnum
keyword to assign a record number to each logical record as it is read from the datafile, and that value will be inserted into the assigned column of the table. The constant
keyword allows you to assign a constant value to a column during the load. For character columns, enclose the constant value within single quotes. If you use the sysdate
keyword, the selected column will be populated with the current system date and time.
If you use the sequence
option, SQL*Loader will maintain a sequence of values during the load. As records are processed, the sequence value will be increased by the increment you specify. If the rows fail during insert
(and are sent to the bad file), those sequence values will not be reused. If you use the max
keyword within the sequence option, the sequence values will use the current maximum value of the column as the starting point for the sequence. The following listing shows the use of the sequence
You can also specify the starting value and increment for a sequence to use when inserting. The following example inserts values starting with a value of 100, incrementing by 2. If a row is rejected during the insert
, its sequence value is skipped.
If you store numbers in VARCHAR2 columns, avoid using the sequence
option for those columns. For example, if your table already contains the values 1 through 10 in a VARCHAR2 column, then the maximum value within that column is 9the greatest character string. Using that as the basis for a sequence
option will cause SQL*Loader to attempt to insert
a record using 10 as the newly created valueand that may conflict with the existing record.
SQL*Loader control files can support complex logic and business rules. For example, your input data for a column holding monetary values may have an implied decimal; 9990 would be inserted as 99.90. In SQL*Loader, you could insert this by performing the calculation during the data load:
money_amount position (20:28) external decimal(9) ":tax_amount/100"
See the "SQL*Loader Case Studies" of the Oracle9i Utilities Guide for additional SQL*Loader examples and sample control files.
Managing Data Loads
Loading large data volumes is a batch operation. Batch operations should not be performed concurrently with the small transactions prevalent in many database applications. If you have many concurrent users executing small transactions against a table, you should schedule your batch operations against that table to occur at a time when no users are accessing the table.
Oracle maintains read consistency
for users' queries. If you execute the SQL*Loader job against the table at the same time that other users are querying the table, Oracle will internally maintain undo entries to enable those users to see their data as it existed when they first queried the data. To minimize the amount of work Oracle must perform to maintain read consistency (and to minimize the associated performance degradation caused by this overhead), schedule your long-running data load jobs to be performed when few other actions are occurring in the database. In particular, avoid contention with other accesses of the same table.
Design your data load processing to be easy to maintain and reuse. Establish guidelines for the structure and format of the input datafiles. The more standardized the input data formats are, the simpler it will be to reuse old control files for the data loads. For repeated scheduled loads into the same table, your goal should be to reuse the same control file each time. Following each load, you will need to review and move the log, bad, data, and discard files so they do not accidentally get overwritten.
Within the control file, use comments to indicate any special processing functions being performed. To create a comment within the control file, begin the line with two dashes, as shown in the following example:
-- Limit the load to LA employees:
If you have properly commented your control file, you will increase the chance that it can be reused during future loads. You will also simplify the maintenance of the data load process itself, as described in the next section.
Repeating Data Loads
Data loads do not always work exactly as planned. Many variables are involved in a data load, and not all of them will always be under your control. For example, the owner of the source data may change its data formatting, invalidating part of your control file. Business rules may change, forcing additional changes. Database structures and space availability may change, further affecting your ability to load the data.
In an ideal case, a data load will either fully succeed or fully fail. However, in many cases, a data load will partially succeed, making the recovery process more difficult. If some of the records have been inserted into the table, then attempting to reinsert those records should result in a primary key violation. If you are generating the primary key value during the insert (via the sequence
option), then those rows may not fail the second timeand will be inserted twice.
To determine where a load failed, use the log file. The log file will record the commit points as well as the errors encountered. All of the rejected records should be in either the bad file or the discard file. You can minimize the recovery effort by forcing the load to fail if many errors are encountered. To force the load to abort before a large number of errors is encountered, use the errors
keyword of the SQLLDR
command. You can also use the discardmax
keyword to limit the number of discarded records permitted before the load aborts.
If you set errors
to 0, the first error will cause the load to fail. What if that load fails after 100 records have been inserted? You will have two options: identify and delete the inserted records and reapply the whole load, or skip the successfully inserted records. You can use the skip
keyword of SQLLDR
to skip the first 100 records during its load processing. The load will then continue with record 101 (which, we hope, has been fixed prior to the reload attempt). If you cannot identify the rows that have just been loaded into the table, you will need to use the skip
option during the restart process.
The proper settings for errors
depend on the load. If you have full control over the data load process, and the data is properly "cleaned" before being extracted to a load file, you may have very little tolerance for errors and discards. On the other hand, if you do not have control over the source for the input datafile, you need to set errors
high enough to allow the load to complete. After the load has completed, you need to review the log file, correct the data in the bad file, and reload the data using the original bad file as the new input file. If rows have been incorrectly discarded, you need to do an additional load using the original discard file as the new input file.
After modifying the errant CategoryName value, you can rerun the BOOKSHELF table load example using the original bookshelf.dat file. During the reload, you have two options when using the original input datafile:
- Skip the first row by specifying skip=1 in the SQLLDR command line.
- Attempt to load both rows, whereby the first row fails because it has already been loaded (and thus causes a primary key violation).
Alternatively, you can use the bad file as the new input datafile and not worry about errors and skipped rows.
Tuning Data Loads
In addition to running the data load processes at off-peak hours, you can take other steps to improve the load performance. The following steps all impact your overall database environment, and must be coordinated with the database administrator. The tuning of a data load should not be allowed to have a negative impact on the database or on the business processes it supports.
First, batch data loads may be timed to occur while the database is in NOARCHIVELOG mode. While in NOARCHIVELOG mode, the database does not keep an archive of its online redo log files prior to overwriting them. Eliminating the archiving process improves the performance of transactions. Since the data is being loaded from a file, you can re-create the loaded data at a later time by reloading the datafile rather than recovering it from an archived redo log file.
However, there are significant potential issues with disabling NOARCHIVELOG mode. You will not be able to perform a point-in-time recovery of the database unless archiving is enabled. If there are non-batch transactions performed in the database, you will probably need to run the database in ARCHIVELOG mode all the time, including during your loads. Furthermore, switching between ARCHIVELOG and NOARCHIVELOG modes requires you to shut down the instance. If you switch the instance to NOARCHIVELOG mode, perform your data load, and then switch the instance back to ARCHIVELOG mode, you should perform a backup of the database (see Chapter 40) immediately following the restart.
Instead of running the entire database in NOARCHIVELOG mode, you can disable archiving for your data load process by using the unrecoverable
keyword within SQL*Loader. The unrecoverable
option disables the writing of redo log entries for the transactions within the data load. You should only use this option if you will be able to re-create the transactions from the input files during a recovery. If you follow this strategy, you must have adequate space to store old input files in case they are needed for future recoveries. The unrecoverable
option is only available for Direct Path loads, as described in the next section.
Rather than control the redo log activity at the load process level, you can control it at the table or partition level. If you define an object as nologging
, then block-level inserts performed by SQL*Loader Direct Path loading and the insert /*+ APPEND */
command will not generate redo log entries.
If your operating environment has multiple processors, you can take advantage of the CPUs by parallelizing the data load. The parallel
option of SQLLDR
, as described in the next section, uses multiple concurrent data load processes to reduce the overall time required to load the data.
In addition to these approaches, you should work with your database administrator to make sure the database environment and structures are properly tuned for data loads. Tuning efforts should include the following:
- Preallocate space for the table, to minimize dynamic extensions during the loads.
- Allocate sufficient memory resources to the shared memory areas, including the log buffer area.
- Streamline the data writing process by creating multiple database writer (DBWR) processes for the database.
- Remove any unnecessary triggers during the data loads. If possible, disable or remove the triggers prior to the load, and perform the trigger operations on the loaded data manually after it has been loaded.
- Remove or disable any unnecessary constraints on the table. You can use SQL*Loader to dynamically disable and re-enable constraints.
- Remove any indexes on the tables. If the data has been properly cleaned prior to the data load, then uniqueness checks and foreign key validations will not be necessary during the loads. Dropping indexes prior to data loads significantly improves performance.
If you leave indexes on during a data load, Oracle must manage and rebalance the index with each inserted record. The larger your data load is, the more work Oracle will have to do to manage the associated indexes. If you can, you should consider dropping the indexes prior to the load and then re-creating them after the load completes. The only time indexes do not cause a penalty for data load performance is during a Direct Path load, as described in the next section.
Direct Path Loading
SQL*Loader, when inserting records, generates a large number of insert statements. To avoid the overhead associated with using a large number of inserts, you may use the Direct Path option in SQL*Loader. The Direct Path option creates preformatted data blocks and inserts those blocks into the table. As a result, the performance of your load can dramatically improve. To use the Direct Path option, you must not be performing any functions on the values being read from the input file.
Any indexes on the table being loaded will be placed into a temporary DIRECT LOAD state (you can query the index status from USER_INDEXES). Oracle will move the old index values to a temporary index it creates and manages. Once the load has completed, the old index values will be merged with the new values to create the new index, and Oracle will drop the temporary index it created. When the index is once again valid, its status will change to VALID. To minimize the amount of space necessary for the temporary index, presort the data by the indexed columns. The name of the index for which the data is presorted should be specified via a sorted indexes
clause in the control file.
To use the direct path option, specify
as a keyword on the SQLLDR command line or include this option in the control file.
If you use the Direct Path option, you can use the unrecoverable
keyword to improve your data load performance. This instructs Oracle not to generate redo log entries for the load. If you need to recover the database at a later point, you will need to re-execute the data load in order to recover the table's data. All conventional path loads are recoverable, and all Direct Path loads are recoverable by default.
Direct Path loads are faster than conventional loads, and unrecoverable
Direct Path loads are faster still. Since performing unrecoverable
loads impacts your recovery operations, you need to weigh the costs of that impact against the performance benefit you will realize. If your hardware environment has additional resources available during the load, you can use the parallel
Direct Path load option to divide the data load work among multiple processes. The parallel Direct Path operations may complete the load job faster than a single Direct Path load.
Instead of using the parallel
option, you could partition the table being loaded (see Chapter 18). Since SQL*Loader allows you to load a single partition, you could execute multiple concurrent SQL*Loader jobs to populate the separate partitions of a partitioned table. This method requires more database administration work (to configure and manage the partitions), but it gives you more flexibility in the parallelization and scheduling of the load jobs.
As of Oracle9i, you can take advantage of multithreaded loading functionality for Direct Path loads to convert column arrays to stream buffers and perform stream buffer loading in parallel. Use the streamsize
parameter and multithreading
flag to enable this feature.
Direct Path loading may impact the space required for the table's data. Since Direct Path loading inserts blocks of data, it does not follow the usual methods for allocating space within a table. The blocks are inserted at the end of the table, after its high-water mark,
which is the highest block into which the table's data has ever been written. If you insert
100 blocks worth of data into a table and then delete
all of the rows, the high-water mark for the table will still be set at 100. If you then perform a conventional SQL*Loader data load, the rows will be inserted into the already allocated blocks. If you instead perform a Direct Path load, Oracle will insert new blocks of data following block 100, potentially increasing the space allocation for the table. The only way to lower the high-water mark for a table is to truncate
it (which deletes all rows and cannot be rolled back) or to drop and re-create it. You should work with your database administrator to identify space issues prior to starting your load.
Additional Oracle9i Enhancements
In addition to features noted earlier in this chapter, SQL*Loader features support for Unicode and expanded datatypes. As of Oracle9i, SQL*Loader can load integer and zoned/packed decimal datatypes across platforms with different byte ordering and accept EBCDIC-based zoned or packed decimal data encoded in IBM format. SQL*Loader also offers support for loading XML columns, loading object types with subtypes (see Chapter 30), and Unicode (UTF16 character set). SQL*Loader also provides native support for the new Oracle9i date, time, and interval-related datatypes (see Chapter 9).
If a SQL*Loader job fails, you may be able to resume it where it failed using the resumable
, and resumable_timeout
options. For example, if the segment to which the loader job was writing could not extend, you can disable the load job, fix the space allocation problem, and resume the job. Your ability to perform these actions depends on the configuration of the database; work with your DBA to make sure the resumable features are enabled and adequate undo history is maintained for your purposes.
As of Oracle9i, you can access external files as if they are tables inside the database. This "external table" feature, described in Chapter 25, allows you to potentially avoid loading large volumes of data into the database. The syntax for external table definitions very closely resembles that of the SQL*Loader control file. Although they are limited in some significant ways (you cannot perform DML on external tables, for example), you should consider external tables as alternatives to data loads. See Chapter 25 for implementation details.