How to partition tables and indexes in this scenario?

Question

So our situation is pretty simple. We have 3 tables.
A, B and C
the model is A->>B->>C
Currently A, B and C are range partitioned on a key created_date however it's typical that only C is every qualfied with created date. There is a foreign key from B -> A and C -> B
we have many queries where the data is identified by state that is indexed currently non partitioned on columns in A ... there are also indexes on the foreign keys that get from C -> B -> A. Again these are non partitioned indexes at this time.
It is typical that we qualify A on either account or user or both. There are indexes (non partitioned) on these
We have a problem now because many of the queries use leading wildcards ie. account like '%ACCOUNT' etc. This often results in large full table scans. Our solution has been to remove the leading wildcard but this isn't always possible.
We are wondering how we can benefit from partitioning and or sub partitioning table A. since it's partitioned on created_date but rarely qualify by that.
We are also wondering where and how we can benefit from either global partitioned index or local partitioned indexes on tables A. We suspect that the index on the foreign key from C to B could be a local partitioned index.
I am also wondering what impact pushing the state from A that's used to qualify A down to C would have any advantage.
C is the table that currently we qualify with the partition key so I figure if you also pushed down the state from A that's used to qualify the set of C's we want based on the set of B's we want based on the set of A thru qualfying on columns within A.
If we push down some of those values to C and simply use C when filtering I'm wondering what the plans will look like compared to having to work all the way up from the bottom to the top before it begins qualifying A.
Edited by: steffi on Jan 14, 2011 11:36 PM

Answer 1

We are wondering how we can benefit from partitioning and or sub partitioning table A. since it's partitioned on >created_date but rarely qualify by that. Very good question. Why did you partition on it? You will never have INSERTS on these partitions, but maybe deletes and updates? The only advantage (I can think of) would be to set these partitions in a read only tablespace to ease backup... but that's a weired reason.
we have many queries where the data is identified by state that is indexed currently non partitioned on columns in >A ... there are also indexes on the foreign keys that get from C -> B -> A. Again these are non partitioned indexes at >this time.Of course. Why should they be partitioned by Create_date?
It is typical that we qualify A on either account or user or both. There are indexes (non partitioned) on these
We have a problem now because many of the queries use leading wildcards ie. account like '%ACCOUNT' etc. This >often results in large full table scans. Our solution has been to remove the leading wildcard but this isn't always possible.I would suspect full index scan. Isn't it?
We are also wondering where and how we can benefit from either global partitioned index or local partitioned >indexes on tables A. We suspect that the index on the foreign key from C to B could be a local partitioned index.As A is not accessed by any partition, why should C and B profit? You should look to partition by the key you are using to access. But, you are looking to tune your SQLs where the access is like '%ACCOUNT' on A. Then when there is a match. ORACLE joins via your index and nested loop (right?) to B and C.
I am also wondering what impact pushing the state from A that's used to qualify A down to C would have any >advantage.Why should it. It just makes the table and indexes larger => more IO.
C is the table that currently we qualify with the partition key so I figure if you also pushed down the state from A >that's used to qualify the set of C's we want based on the set of B's we want based on the set of A thru qualfying >on columns within A.If the access from A to C would be .. AND A.CREATE_DATE =C.CREATE_DATE and c.key like '%what I want%' which does not qualifify for a FK ;-) then, as that could be resulting in a partition scan, you could "profit". But, I'm sure that's not your model.
If we push down some of those values to C and simply use C when filtering I'm wondering what the plans will look >like compared to having to work all the way up from the bottom to the top before it begins qualifying A.So you want to denormalize A,B,C and into one table? With the same access is like '%ACCOUNT' you would get a full scan on an index as before, just the objects would be larger due to redundance and harder to maintain. In the end you would have a bad and slower design.
Maybe you explain what the problem is.
Full index scan can not be avoided, but that can be made faster by e.g. parallel query, and then the join to B and C should be a "snip" if you just identify a small subset of rows in these tables.