Chat now with support
Chat with Support

SharePlex 9.2.9 - Administration Guide

About this Guide Conventions used in this guide Overview of SharePlex Run SharePlex Run multiple instances of SharePlex Execute commands in sp_ctrl Set SharePlex parameters Configure data replication Configure replication to and from a container database Configure named queues Configure partitioned replication Configure replication to a change history target Configure a replication strategy Configure DDL replication Configure error handling Configure data transformation Configure security features Start replication on your production systems Monitor SharePlex Prevent and solve replication problems Repair out-of-sync data Tune the Capture process Tune the Post process Recover replication after Oracle failover Make changes to an active replication environment Apply an Oracle application patch or upgrade Back up Oracle data on the source or target Troubleshooting Tips Appendix A: Peer-To-Peer Diagram Appendix B: SharePlex environment variables

Before you use Compare and Repair

Before running the compare or repair commands, review these guidelines.

  • All of the SharePlex processes (Capture, Read, Export, Import, Post) must be running when you run a comparison or repair command.
  • The tables that you want to compare or repair must be part of an active configuration file.
  • Uncommitted transactions on a source table prevent the compare and repair processes from obtaining the brief locks they need to obtain read consistency. Make certain that all transactions are committed before you run a comparison or repair.
  • If a table is large, it will probably need to be sorted in the TEMP tablespace. Before running the compare or repair commands, the TEMP tablespace may need to be made larger. The size depends on the setting of the SP_DEQ_THREADS parameter or the threads option within the command syntax, both of which controls the number of processing threads used by SharePlex on the target. Each thread processes a table. At the default of two threads, the size of the tablespace should be larger than the sum of the sizes of the two largest tables. If you set the number of threads higher, then increase the size of the tablespace to accommodate a proportionate number of the largest tables. However,
  • The UNDO tablespace may also need to be increased. Based on transaction volume and the length of time it takes to compare the largest table, increase the size of the UNDO tablespace and increase the undo_retention database parameter to avoid an ORA-1555 Snapshot too old error. Tables with LOBs take much longer to compare or repair than tables without them.

The following are commonly modified compare and repair parameters. Do not increase the values unless necessary. For details about these parameters, see their documentation in the SharePlex Reference Guide.

Parameter Description
SP_DEQ_MALLOC

This parameter controls the fetch batch size. The batch size controls the number of rows that SharePlex selects at once for comparison. Larger batch sizes increase processing speed but require more memory. The value is divided equally by the number of compare threads to be used, and then the batch size is recalculated based on all column sizes added together.

SP_DEQ_PARRALLISM

This parameter manages the select statement Degree of Parallelism hint. The parallelism option of the command overrides this setting.

SP_DEQ_PART_TABLE_UPDATE This parameter controls how the repair commands work on Oracle partitioned tables, depending on whether row movement is possible.
SP_DEQ_READ_BUFFER_SIZE This parameter controls the size of the buffer that holds fetched LONG and LOB data and can be adjusted based on available system memory.
SP_DEQ_ROW_LOCK_THRESHOLD This parameter sets a threshold that controls whether SharePlex uses row-level or table-level locking when a where option is used.
SP_DEQ_SKIP_LOB

This parameter determines whether or not LOBs are included in the compare/repair processing.

  • When the parameter is set to the default of 0, the compare processes include LOBs in their processing.
  • When the parameter is set to 1, only non-LOB columns are compared and repaired. If LOBs are not modified once inserted, you can speed up processing by setting this parameter to 1.

Set this parameter on the source system.

SP_DEQ_TIMEOUT This parameter sets a queue backlog threshold. High backlogs delay the establishment of a connection between the source and target compare/repair processes. If the backlog meets or exceeds this value, any compare or repair command that is issued on the source will exit and return an error. If this happens, consider running the compare or repair when the system is less busy.

How to use the repair and compare commands

The recommended procedure for maintaining synchronized data through the comparison and repair commands is to run the compare or compare using command first, then view the results with the repair status command. This command shows any rows that are out-of-sync and the possible cause. Unless the cause of the out-of-sync condition is corrected, replication will go out of synchronization again, even if you repair the rows this time. After the problem is fixed, issue the repair or repair using command.

You can run the repair or repair using command without doing a preliminary comparison. The command performs a comparison first, to identify the out-of-sync rows, and then it repairs those rows. However, the underlying cause of the out-of-sync condition must be corrected to prevent future out-of-sync conditions.

See for causes and solutions for out-of-sync conditions.

To view the status or results of a comparison, use the compare status command in sp_ctrl.

To view the status or results of a repair, use the repair status command in sp_ctrl.

When to run a repair

The best time to repair a target table depends on its size, the cause of the problem, the extent of out-of-sync rows, and how long you are willing to tolerate users being locked out. Before you initiate a repair, consider the following:

  • Although the users of the tables are not usually affected by the brief locks that are applied when tables are compared, they are locked out of the target table for the duration of the repair process. For a small table, this might not be disruptive, but for a large table needing extensive repairs, the wait can be significant.
  • Locks on a target table can reduce posting performance if Post must wait for the repair to finish before it can apply changes to that table and move on to other tables. This increases the latency of the target data and causes operations to accumulate in the post queue. If the objects that Post needs to change are different from those being repaired, the two processes run simultaneously.
  • If you must repair a table immediately, but cannot tolerate locks or replication latency, you can use the where option to limit the repair to certain rows. An alternative is to use the key option, but this option may cause the repair to miss some out-of-sync rows.

  • If the repair can wait, correct the cause of the problem immediately and then do the repair during non-peak hours.
  • Replication latency can slow down the compare and repair processing. The message sent from the source to spawn the command processes on the target is sent through the queues along with regular replicated data. Delays caused by a data backlog will delay the spawn message and cause the process on the source to lose its read consistency, which results in errors. If possible, perform comparisons and repairs during off-peak hours.

How to run the compare and repair commands

To get additional information and syntax for the compare and repair commands, see the command documentation in the SharePlex Reference Guide.

Tune the Capture process

This chapter contains instructions for improving the performance of the Capture process to prevent Capture from losing pace with the volume of redo that an Oracle source database generates.

Contents

Disable LOB Mapping

If you have PK/UK logging enabled on the source database (recommended to support more SharePlex features and faster processing), check the setting of the SP_OCT_ENABLE_LOBMAP parameter. This parameter controls whether or not SharePlex uses a LOB map when replicating tables that contain out-of-row LOB columns. The LOB map is used by the Capture process to map LOBIDs and rows when PK/UK logging is not enabled. LOB mapping is enabled by default. The SHAREPLEX_LOBMAP table stores these mappings.Transactions with numerous LOB operations can slow down Capture because it needs to maintain and refer to the mappings. If PK/UK logging is enabled on the database, you can disable LOB mapping by setting this parameter to 0.

To disable LOB mapping during active replication

  1. Run sp_ctrl on the source system.
  2. Set SP_OCT_ENABLE_LOBMAP to 0.

    sp_ctrl> set param SP_OCT_ENABLE_LOBMAP 0

  3. Stop Capture.

    sp_ctrl> stop capture

  4. Truncate the SHAREPLEX_LOBMAP table.
  5. Restart Capture.

    sp_ctrl> start capture

Related Documents

The document was helpful.

Select Rating

I easily found the information I needed.

Select Rating