On Wednesday 5th December, I attended Tech 17. While it is a pain to organise getting there and back, I found in previous years that it was useful for me to go and find out what is happening in the industry.

Cost Based Optimisation - The Panel Session

Panelists:

Slow Start

This took questions from the internet and the audience, so it is a bit variable. The question has to match something the panel can talk about. The first questions about the Autonomous database was disappointing in this regard as very little information has been released about it, and the Oracle employees had to stick to the company line, so couldn’t speculate.

Favourite 12.2 feature.

Nigel Bayliss said he liked the new cost based OR transformation, where an OR can be replaced by a UNION ALL, and the statements in the union can be run in parallel. I leaned more about this later.

Another panelist mentioned they liked the fact that you can now reorganise a table online, including indexes. This means that the cluster factor of tables can be changed to match the index for important queries, making index use more likely for these queries.

Clustering Factor

There was some discussion of clustering factor after another question on tuning. This is a number calculated for an index when gathering statistics which the optimizer uses to judge whether it is worth using an index or a full table scan. It counts the number of blocks that would need to be visited in the table to read it in the order of the index. So the best is the number of blocks in the table, where the table is ordered exactly as the index. The worst is the number of rows in the table, where every row is in another block.

The problem occurs when the rows are actually clustered together, but subsequent rows are in different blocks (e.g. to read rows in order you need to skip between two blocks. Apparently in some circumstances a parallel load of data can cause this to happen. This means that the data is clustered, but the optimiser thinks it isn’t.

The solution is to set the table_cache_blocks preference on the stats gathering process to something higher. Jonathan Lewis preferred 8, and Richard Foote prefers 42. The idea is that if the row is in a block that has been read in the last few blocks, it is not counted as another read, as the block will have been cached. This makes it more likely that an index will be selected.

We should be cautious of this though, as other parameters such as optimizer_index_cost_adj and optimizer_index_caching may already have been set to solve this problem. A query on a poorly clustered table with all these parameters set may end up using the index when it is not appropriate, therefore generate more IO than a table scan, so as with many changes it should be done cautiously.

Index Monitoring

Also mentioned was that 12.2 now monitors indexes by default making it more easy to understand whether it is being used.

Histograms

A questioned asked Maria Colgan and Jonathan Lewis about their differing view of histograms. They both agree they are useful in some circumstances and not others. Maria Colgan takes the view that you should allow the optimiser to create them and remove the ones that are not helpful. It is difficult to know when a histogram will be helpful, so this allows you to get the benefit. Jonathan Lewis takes the view that histograms can cause problems with plan instability, and that they should only be used when they are needed to avoid these problems. However, he seemed to be of the opinion that the Top N histograms don’t have this sort of problem, though the height balanced and hybrid histograms still need to be treated with caution.

Performance

There was some discussion on proactive SQL monitoring. To see where the optimiser is going wrong, you can look at the different between the expected and actual rows, and at the first point in the plan (by execution) where these numbers are different, look into why the optimiser generated in incorrect estimate.

SQL Plan directives were automatically generated and used in 12.1 for queries with similar predicates, which caused problems in some cases, particularly for OLTP and mixed workloads. This is turned off by default in 12.2 with the optimizer_dynamic_statistics parameter.