I mentioned in my previous two posts that I had tried to figure out if it would be safe to turn on force logging on a production database that does a bunch of batch processing on the weekend: post1, post2. We know that many of the tables are set to NOLOGGING and some of the inserts have the append hint. We put in force logging on Friday and the heavy weekend processing ran fine last weekend.
I used an AWR report to check the top INSERT statements from the weekend and I only found one that was significantly slower. But the table it inserts into is set for LOGGING, it does not have an append hint, and the parallel degree is set to 1. So, it is a normal insert that was slower last weekend for some other reason. Here is the output of my sqlstatsumday.sql script for the slower insert:
It averaged 197062 milliseconds last weekend but 89684 the previous one. The target table has always been set to LOGGING so FORCE LOGGING would not change anything with it.
One of the three INSERT statements that I expected to be slowed by FORCE LOGGING was faster this weekend than without FORCE LOGGING last weekend:
It ran 2421533 milliseconds this weekend and 5672230 the prior one. So clearly FORCE LOGGING did not have much effect on its overall run time.
It went so well this weekend that we decided to leave FORCE LOGGING in for now to see if it slows down the mid-week jobs and the web-based front end. I was confident on Friday, but I am even more confident now that NOLOGGING writes have minimal performance benefits on this system. But we will let it bake in for a while. Really, we might as well leave it in for good if only for the recovery benefits. Then when we configure GGS for the zero downtime upgrade it will already have been there for some time.
The lesson for me from this experience and the message of my last three posts is that NOLOGGING writes may have less benefits than you think, or your system may be doing less NOLOGGING writes than you think. That was true for me for this one database. It may be true for other systems that I expect to have a lot of NOLOGGING writes. Maybe someone reading this will find that they can safely use FORCE LOGGING on a database that they think does a lot of NOLOGGING writes, but which really does not need NOLOGGING for good performance.
Pingback: Upgrade Data Migration Choices – HP to Linux | Bobby Durrett's DBA Blog