Re: Change in Index costing because of blevel change

For the quick and dirty workaround I would adjust the blevel on the partition and at the table level. Any time anything triggers a stats collection on the table you MIGHT find that the blevel would be recalculated, but if you've set the blevel on the partition to 2 then the problem could only re-appear if the stats on that partition were re-collected (and the collection was set to cascade to the indexes). This happens because the global (table level) blevel is set to be the maximum of the partition blevel values (See:
https://jonathanlewis.wordpress.com/2010/11/14/local-indexes-2/ )

Given this behaviour it would probably be sensible at some point to split the over-large partition into two pieces so that an unexpected stats gather won't accidentally re-introduced the higher blevel).

Regards
Jonathan Lewis

On Tue, 22 Jun 2021 at 06:20, Lok P <loknath.73_at_gmail.com> wrote:

> Thanks a lot.
>
> In regards to get/set_index_stats, I am thinking of the stats
> gathering(which I believe is currently happening in an INCREMENTAL way)
> which has turned the blevel at global level to -3 because of that one
> partition mess. So if i forcibly set the BLEVEL for the 15th jun index
> partition to -2 manually(whose value is currently-3) using set_index_stats,
> will the subsequent stats gather revert the BLEVEL of the index at global
> level i.e. in dba_indexes automatically to -2?
> Or it will do the opposite, i.e. it will regather the stats on that 15th
> jun index partition and make it -3 along with global blevel to -3?
> Or are you pointing towards forcibly get/set_index_stats the blevel
> to-2 for both local i.e the 15th jun partition along with global blevel for
> that index?
>
> Regards
> Lok
>
> On Tue, Jun 22, 2021 at 12:50 AM Jonathan Lewis <jlewisoracle_at_gmail.com>
> wrote:
>
>>
>> You might get away with two calls to dbms_stats, one to "get_index_stats"
>> the other to "set_index_stats". modifying the variable holding the blevel
>> between the calls.
>> Example here: https://jonathanlewis.wordpress.com/2010/01/06/copy-stats/
>>
>> As far as the split partition is concerned - if the 15th June is now in a
>> read-only stats you could do the following steps (to hide the time, rather
>> than getting rid of it).
>> create table temp14 as select * from15th June partition where data = 14th
>> june, create indexes
>> create table temp15 as select * from15th June partition where data = 15th
>> june, create indexes
>>
>> truncate 15th June partition
>> split partition into 14th june and 15th june
>> exchange partition for 14th june and temp14 including indexes
>> exchange partition for 15th june and temp15 including indexes
>>
>> The users would only see brief interruptions on the exchange - unless you
>> had some global indexes to maintain as well.
>>
>> Regards
>> Jonathan Lewis
>>
>>
>>
>>
>> On Mon, 21 Jun 2021 at 19:55, Lok P <loknath.73_at_gmail.com> wrote:
>>
>>> Hi All, It's version 11.2.0.4 of Oracle.
>>>
>>> We are having a situation in which multiple queries suddenly changed the
>>> execution path and deviated from the local indexed access path to the FTS
>>> path(partition range all) for a daily range partitioned table. And then
>>> after debugging into the issue we found that the COST of the index access
>>> actually has been increased because of the increase in BLEVEL of the index.
>>> And the BLEVEL has been increased from 2 to 3 in global level(dba_indexes)
>>> because of one of the missing daily partitions. For e.g. 14th jun partition
>>> was missing , so 15th jun partition now holds both 14th june and 15th jun
>>> data, so its twice as bigger as a normal daily partition.
>>>
>>> The missing day range partition caused the subsequent day partitions to
>>> actually hold two partitions/days worth of data and thus the blevel of that
>>> specific local index increased from 2 to -3 for that index partition. and
>>> since then, in global level(dba_indexes) the blevel is staying as-3 for
>>> that index, costing that index access higher and thus making plan change
>>> for many queries.
>>>
>>> My question is how should we handle such a situation ? Should we go back
>>> and break the specific partition(15th Jun partition) into two daily range
>>> partitions i.e. 14th and 15th (which i believe is going to take lot of time
>>> and resources considering this table has ~8 different local indexes in it)
>>> and gather stats again or is there any other way to just tweak the stats
>>> for that one big partition such that the global index BLEVEl can be back to
>>> -2 , and get this issue fixed?
>>>
>>> Regards
>>> Lok
>>>
>>

--
http://www.freelists.org/webpage/oracle-l