Oracle Anti-Join Execution plan question

Question

We have two tables like so:

Event
    id
    type
    ... a bunch of other columns

ProcessedEvent
    event_id
    process

There are indexes defined for

• Event(id) (PK)
• ProcessedEvent (event_id, process)

The first represents events in an application.

The second represents the fact that a certain event got processes by a certain process. There are many processes that need to process a certain event, so there are multiple entries in the second table for each entry in the first.

In order to find all the events that need processing we execute the following query:

select * // of course we do name the columns in the production code
from Event
where type in ( 'typeA', 'typeB', 'typeC')
and id not in (
    select event_id
    from ProcessedEvent
    where process = :1  
)

Statistics are up to date

Since most events are processed, I think the best execution plan should look something like this

• full index scan on the ProcessedEvent Index
• full index scan on the Event Index
• anti join between the two
• table access with the rest
• filter

Instead Oracle does the following

• full index scan on the ProcessedEvent Index
• full table scan on the Event table
• filter the Event table
• anti join between the two sets

With an index hint I get Oracle to do the following:

• full index scan on the ProcessedEvent Index
• full index scan on the Event Index
• table acces on the Event table
• filter the Event table
• anti join between the two sets

which is really stupid IMHO.

So my question is: what might be the reason for oracle to insist on the early table access?

---

Addition: The performance is bad. We are fixing the performance problem by selecting just the Event.IDs and then fetching the needed rows 'manually'. But of course that is just a work around.

Answer 1

your FULL INDEX SCAN will probably be faster than a FULL TABLE SCAN since the index is likely "thinner" than the table. Still, the FULL INDEX SCAN is a complete segment reading and it will be about the same cost as the FULL TABLE SCAN.

However, you're also adding a TABLE ACCESS BY ROWID step. It is an expensive step: one logical IO per row for the ROWID access whereas you get one logical IO per multi blocks (depending upon your db_file_multiblock_read_count parameter) for the FULL TABLE SCAN.

In conclusion, the optimizer computes that:

cost(FULL TABLE SCAN) < cost(FULL INDEX SCAN) + cost(TABLE ACCESS BY ROWID)

Update: The FULL TABLE SCAN also enables the filter on type sooner than in the FULL INDEX SCAN path (since the INDEX doesn't know what type an event is), therefore reducing the size of the set that will be anti-joined (yet another advantage of the FULL TABLE SCAN).