Optimizer questions

Question

Hello, I am pretty new to Oracle. Now I have something strange on the optimizer looking at explain plan.
I have a fact table and several dimension tables. I want to get the Max value of a field from the fact table. And it's extremely slow.
If i don't use Max function, it returns within 3 sec. If i do, it takes 5~10 minutes.
Each dim table with its condition returns the result within 1 sec. Searching the dim table keys in the fact table returns within a couple of seconds (The fact table has millions of records. May be more than that).
And the explain plan looks strange.
If I join 2 dimension tables, the Max function seems to run after all joining.
Explain plan looks like this as I expected:
Hash (GROUP BY)
         NL ...
                   Index Unique Scan (Dim table 1)
                   Index Range Scan (Dim table 2)
               Partition Range
            Partition Range
But, if i join 3 dimension tables, the Max function seems to run before all joining
Explain plan looks like this:
Temp table transformation
     Load as select
          View
               Filter predicates (Dim table 1)
               Hash Join
                    Access Predicates (ROWID = ROWID)
                    Index Range scan (Dim table 1 again?)
     Load as select
          Table access by index rowid
               index range scan (Dim table 2)
     Hash (group by) <== this runs here after joining only 1 dim table with the fact table
          Nested loops
               Hash join
                    access predicates
                    inlist iterator
                         table access by global index rowid
                              index unique scan (Dim table 3)
                    View
                         Nested loops
                              partition range (subquery)
                                   partition list (subquery)
                                        bitmap conversion (to rowids)
                                                  access predicates (FACT_TABLE.KEY_OF_DIMTABLE3 = C0 <=== what is C0?)
                                             BITMAP MERGE
                                                  BITMAP KEY ITERAION                                                                                                    
                                                       BUFFER(SORT)
                                                            TABLE ACCESS (FULL <==== why full?)
                                                       BITMAP INDEX (RANGE SCAN)
                                                            ACCESS PREDICATES (FACT_TABLE.KEY_OF_DIMTABLE2 = C0 <== why use key of dim table 2 here although the condition of dim table was calculated already)
Add just one more table shows not only the much more complex explain plan, but much slower performance. Does it really run the Max function before joining all tables so that it runs on so many rows?
If it is so, what can I do to run it after joining all tables so that it can run on the reduced number of records?
Please help me! Thank you                                                              

Answer 1

the decisions of the optimizer are cost based: there are a lot of (more or less) complex calculations used to determine the plan that seems to be the most appropriate choice to the optimizer. Depending on the statistics (on which the calculation is based) and some limitations of the calculation model this decision is good in many cases, not bad in a lot of cases - and not so good in some other cases. So an additional join and the use of a function can lead to a plan that needs much more resources and multiplies the execution time. (Since I have worked with some other RDBMS I think it's save to say that the Oracle optimizer is quite good in doing its job)
That said I would suggest that you add the complete execution plans since your representation shows the operations but not the costing. If you can also add rowsource statistics (with the gather_plan_hint or a fitting statistics_level) this would make the analysis simpler.
You could also take a look at https://blogs.oracle.com/optimizer/entry/star_transformation since your query seems to be based on a classical star schema and there are some options provided by Oracle to improve the access in such cases.