I have made small investigation about redo generation. From early days of my career I was remember that nologging operation is very performance effective but never try to quantify this very.
Every application can theoretically be split into 4 groups of tables (I use my personal names but hopefully it has sense):
1) Regular tables – contain valuable information need to be stored for legal and functional purposes. Stored as normal tables.
2) Staging tables – contain process lifetime specific information, easily re-creatable. Used for transferring information between sessions and report generation. Stored as regular tables or materialized views.
3) Session Temporary tables– contain process lifetime specific information, easily re-creatable. Used for reporting stored as GLOBAL TEMPORARY tables ON COMMIT PRESERVE.
4) Transaction Temporary tables– contain process lifetime specific information, easily re-creatable. Used for processing optimisation stored as GLOBAL TEMPORARY tables ON COMMIT DELETE.
By default all 4 groups generate REDO logs records that can be significant amount of resources. The redo information is valuable if we:
1) Support StandBy database
2) Information inside tables is valuable and have to be safe in case of database crush.
To make the standby or backup completely usable after a nologging statement is run, a mechanism other than database recovery must be used to get or create current copies of the affected blocks. You have to drop and recreate the object with the invalidated blocks or truncate it, using the program that maintains the object. Thus extra step to manage switchover/failover to standby database process have to be introduced.
Again based on my understanding the only business requirements for logging is to keep data from “Regular tables”. The safety of the data from other groups is not such important.
The only DML operation that can be optimised in terms of REDO log generation is INSERT with APPEND hint. (MERGE is actually presentation layer above INSERT thus can be treated together) . Hint APPEND if it works have one negative issue. The data in new table is not actually available until end of transaction.Due to the following error.
ORA-12838: cannot read/modify an object after modifying it in parallel
It linked to the fact that oracle could not make consistent model of block if there is no UNDO information. This actually makes using this hint on Global Temporary tables with ON COMMIT DELETE rows unreasonable. You can insert data but never be able to use it until it would be deleted.
Another fact that I have to highlight UPDATE and DELETE always generate REDO information. Thus if the table intensively update the gains would be minimal. Avoiding this operation on a temporary tables is another skills that developers have to be used to for optimal performance of your application.
There are 5 parameters that actually affect SEGMENT logging: Database LOGGING, Database FORCE LOGGING, Tablespace LOGGING, Tablespace FORCE LOGGING (Can be switched on tablespaces with “Regular tables” and switched off on tablespaces with “Staging tables” , Table LOGGIN. Global Temporary tables actually always in NOLOGGING mode thus we can assume for table groups “Session Temporary tables” and “Transaction Temporary tables” always have all parameters equal to NO. Production databases should always be in protected mode thus the value DATABASE LOGGING should always be in YES, it takes value NO outside of investigation.
To test I have created the table TEST.BIGTABLE (column1 NUMBER) with 39999960 rows and few tables to generate INSERT as SELECT statement from BIGTABLE dataset. The results are below.
|TABLESPACE FORCE LOGGING||*||Y||N||N||N||N||N|
|DATABASE FORCE LOGGING||Y||N||N||N||N||N||N|
|Amount of redo for INSERT APPEND||501000K||501000K||457K||501000K||456K||501000K||501000K|
|Amount of redo for Standard INSERT AS SELECT||501000K||501000K||501000K||501000K||501000K||501000K||501000K|
Amount of redo for temporary tables
|Standard INSERT AS SELECT||INSERT APPEND value|
|Transaction Temp Table||110K||0.3K|
|Session Temp Table||110K||0.3K|
Hope all above have sense and can be used for good
P.S. The “redo size” values has been got from AUTOTRACE statistics.