When optimizing queries, the first thing I do with a slow query is figure out what it's trying to do. You can't fully optimize a query unless you know how to consider alternative ways to write it, and you can't do that unless you know what the query "means." I frequently run into a situation where I'm forced to stop and ask what they were trying to do with the query. This is database-agnostic, not related to Oracle.
Dont optimize if you dont need to
- Use Oracle Grid Control to find the top queries and only work on them.
- Check Oracle Grid Alerts to see if you are doing more I/O or more CPU or insufficient SGA is allocated, if queries are doing more sorts or using sort joins than a bigger sort area can help, similarly a bigger hash area can help for hash joins.
- Avoid Union and use Union All if possible: - Union will require elimination of duplicates and if you know that the sets to union does not have duplicates than use Union all.
- Avoid correlated subqueries: - If possible use denormalized fields. Like Weight and Volume were denormalized on Movement to avoid subqueries for each execution of report.
- If operating on a large dataset than avoid filtering on multiple tables and than joining the result. In Transaction report we are filtering on shipment first to reduce the datset and than do all the big joins.
- Avoid Sorting if possible: - this will consume CPU/memory resources on server and sometimes if sorting is on disk it will be I/O bound.
- Avoid OR clause if possible: - This will cause the dataset to be iterated more than once depending on if the OR clause is the primary index.
- Reduce the size of union set on temporary views: - The number of fields in the select clause should be reduced if we are doing a union. This will reduce the dataset size for union.
- Use clustered indexes: - Clustered indexes will improve performance over non-clustered indexes as the data to filter is already in the index leading to less I/O.
- Use function based indexes for leading wild card search: - We are using Reverse function based indexes for this.
- Reduce the size of Dataset by adding Date filters if you can (e.g some page shows last 90 days and some pages shows last 5 days worth of data as thats what customer is mostly interested in) a
- Use less logic in sql and if possible put it on app layer: - Previously cost component link logic was in sql causing unneccessay joins. We improved it by adding a temporary redirect jsp to have this logic.
- Look at explain plan and besides cost look at the cardinality of dataset and try to reduce the no of bytes in operation
- Look at CPU cost and see if a better index lookup can reduce it: - Sometimes queries can be optimized by adding a clustered index on _id and start_date, earlier we had single column indexes leading to bitmap conversion of individual indexes and than a bitmap join which was cpu intensive for 3 million rows.
- Denormalize fields if you can at write time if read is costly remove joins
- Sometimes you can't do much with the query than a better way is to step back and think if we need to design the tables in a different way or we need to do some logic in the app layer.
- Use a covering index to reduce table lookup by rowId. Sometimes you join to a table only to fetch two to three columns and adding extra columns at the end of index might help the query as all data can be fetched from index instead of going to the table. Use these kind of indexes only for large dataset operations. For smaller dataset queries which access only 1000 or so rows dont use this trick.
- Join condition with mismatched data types will never use an index. Using a cast to convert both to same datatype will do the trick
If Deletes or update on tables are slow and Oracle Grid control shows TM Contention as the wait event than this happening due to missing foreign key indexes. Use FindMissingFKIndex link to find missing foreign key indexes and create them.
Avoid Index full scans. For example a query like
(SELECT SYS_ORG_ID FROM ORGANIZATION WHERE NVL(ORG_NAME, '$null') = ? AND NVL(ENT_NAME, '$null') = ? AND NVL(VC_ID, '-128') = ? )
will use a full index scan whereas a query like
(SELECT SYS_ORG_ID FROM ORGANIZATION WHERE ORG_NAME = ? AND ENT_NAME= ? AND VC_ID = ? )will use a unique scan.