Your initial idea was not bad at all. What you can do is store the unwanted partitions with their local indexes in separate tablespaces. Use rman for the cloning but use the SKIP TABLESPACE option to not clone the tablespaces with the unwanted partitions. (assuming online backup)
After the clone, the skipped tablespaces have datafiles with status RECOVER.
see RMAN DUPLICATE DATABASE: Options
In the end you just drop the unwanted partitions. To be able to do that you first have to get rid of things like constraints and indexes that need to be re-created later on. This worked in 10gR2. Make sure that you don't drop the last partition of a table, in that case drop the table.
It is a bit of work but certainly possible. If the difference in Volume is huge or there are lots of copies, it might be worth spending some time for it.
Chained and Migrated Rows are explained in the Logical Storage Structures of the Concepts Guide.
A migrated row would be a row who's column data was completely transferred from one block to another due to an update. The original block would essentially only have a "forwarding address" stored for that row.
A chained row would have parts of its column data in multiple blocks. The original block would contain both actual column data and a forwarding address for the rest of it. (You can get rows chained to more than two blocks.)
Both are implemented the same way deep down, so they're really two aspects of the same thing.
Also note that for tables with more than 255 columns, all rows are technically chained - one "row piece" can only contain 255 column values. The chaining can happen in the same block, or with other blocks depending on space availability (and isn't particularly "bad" if all the data ends up in the same block).
The only way, as far as I know, to get accurate data on row chaining is to use:
ANALYZE TABLE your_table [partition (your_part)] LIST CHAINED ROWS
See Listing Chained Rows of Tables and Clusters.
This is potentially expensive, the whole table needs to be scanned. Statistics gathering doesn't fill the CHAIN_CNT
column of the dba_tables
view. (I think it might have at some point, but it doesn't in 11.2 at least.)
You can monitor the table fetch continued row
1 statistic (v$sysstat
) to see if a query is affected by chained or migrated rows, but I don't believe you can have that metric per-session so either you need a quiet system to measure, or the reading will be "noisy".
The Secrets of Oracle Row Chaining and Migration has interesting information about chained and migrated rows, how you measure them, and potential ways of fixing them.
As always, don't go about rebuilding tables or changing storage parameters "just because" you see chained or migrated rows. Do so only if you measure that it's actually causing you performance problems.
1 From Statistics Descriptions:
Number of times a chained or migrated row is encountered during a fetch
Retrieving rows that span more than one block increases the logical I/O by a factor that corresponds to the number of blocks than need to be accessed. Exporting and re-importing may eliminate this problem. Evaluate the settings for the storage parameters PCTFREE and PCTUSED. This problem cannot be fixed if rows are larger than database blocks (for example, if the LONG datatype is used and the rows are extremely large).
See also Table Fetch by Continued Row.
Best Answer
You have a lot of columns and a high average row length (662). Maybe you even have rows longer than the block size.
With long rows, it is inevitable to have chained rows. Example with 8K block size:
Also make sure you run
ANALYZE TABLE
, because that is what updatesCHAIN_CNT
.