maria-discuss team mailing list archive

[Justin Swanhart <greenlion@xxxxxxxxx>]

Hi,

You can't fix that with a flag that says an index is unique.  It is a big
problem.  Domas opened a bug on a variety of cases of it recently.

On Mon, Jun 22, 2015 at 8:03 PM, Roberto Spadim <roberto@xxxxxxxxxxxxx>
wrote:

> the other solution is  "force index()" but i'm trying to understand why it
> choose a 'bad' index
>
> 2015-06-23 0:01 GMT-03:00 Roberto Spadim <roberto@xxxxxxxxxxxxx>:
>
>> ehhe that's the point, how to remove perverse order by, today i'm using
>> temporary table and after ordering, but i don't like this "solution"
>>
>> 2015-06-22 23:50 GMT-03:00 Justin Swanhart <greenlion@xxxxxxxxx>:
>>
>>> If you have the PK in your query (unless the db does something perverse
>>> with order by (which is common)), then the PK will be used.  Sometimes the
>>> order by thing bites in other direction, and PK is used for desc or asc
>>> access to table when another index could be used.  Hash indexes are only on
>>> memory, and are preferred over b-tree unless ranges are performed because
>>> hash is O(1) and b-tree is O(log(n)).  rtree and fulltext indexes can only
>>> be used with particular functions.  non-clustered indexes are used in all
>>> other cases, sometimes with ICP.
>>>
>>> When you have more than one index to choose that could be effective, the
>>> histograms (statistics) for the tables are compared and the database tries
>>> to decide which index will perform best (cost based optimization).
>>>
>>> --Justin
>>>
>>> On Mon, Jun 22, 2015 at 7:47 PM, Roberto Spadim <roberto@xxxxxxxxxxxxx>
>>> wrote:
>>>
>>>> yeap
>>>>
>>>> at optimizer part of mariadb, is there some check about
>>>> clustered/nonclustered/rtree/hash index to select best index or not?
>>>>
>>>> 2015-06-22 23:45 GMT-03:00 Justin Swanhart <greenlion@xxxxxxxxx>:
>>>>
>>>>> you are probably thinking of b-trees which are unbalanced.  b+tree is
>>>>> o(log(n)) based on depth of tree and you always have to go to bottom to get
>>>>> leaf.
>>>>>
>>>>> On Mon, Jun 22, 2015 at 7:44 PM, Justin Swanhart <greenlion@xxxxxxxxx>
>>>>> wrote:
>>>>>
>>>>>> No, you have to look in the leaf!
>>>>>>
>>>>>> On Mon, Jun 22, 2015 at 7:43 PM, Roberto Spadim <
>>>>>> roberto@xxxxxxxxxxxxx> wrote:
>>>>>>
>>>>>>> yeap, i just want know if the unique allow stop the conquer at first
>>>>>>> result (unique key) or if it continue (non unique key)
>>>>>>>
>>>>>>> 2015-06-22 23:42 GMT-03:00 Justin Swanhart <greenlion@xxxxxxxxx>:
>>>>>>>
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> b+tree search is binary search, which is divide and conquer.
>>>>>>>> Saving scanning a key is like sneezing in a hurricane, you still are
>>>>>>>> o(log(n)) in.
>>>>>>>>
>>>>>>>> On Mon, Jun 22, 2015 at 7:39 PM, Justin Swanhart <
>>>>>>>> greenlion@xxxxxxxxx> wrote:
>>>>>>>>
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> If you do a code inspection you will find that uniqueness is
>>>>>>>>> checked on insertion, and the database cares not about it after that.
>>>>>>>>>
>>>>>>>>> --Justin
>>>>>>>>>
>>>>>>>>> On Mon, Jun 22, 2015 at 7:20 PM, Roberto Spadim <
>>>>>>>>> roberto@xxxxxxxxxxxxx> wrote:
>>>>>>>>>
>>>>>>>>>> and there's optimization with this flag? i see that equal could
>>>>>>>>>> be optimized cause if we found 1 row we can return to user that row and
>>>>>>>>>> stop select
>>>>>>>>>>
>>>>>>>>>> in other words, if i see a index with primary key or unique
>>>>>>>>>> columns, could i change the index to unique and get some kind of
>>>>>>>>>> performace? today i think not ,cause unique key check if it's really unique
>>>>>>>>>> (for example at insert/update/replace), i'm wrong?
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> --
>>>>>>> Roberto Spadim
>>>>>>> SPAEmpresarial - Software ERP
>>>>>>> Eng. Automação e Controle
>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>
>>>>
>>>> --
>>>> Roberto Spadim
>>>> SPAEmpresarial - Software ERP
>>>> Eng. Automação e Controle
>>>>
>>>
>>>
>>
>>
>> --
>> Roberto Spadim
>> SPAEmpresarial - Software ERP
>> Eng. Automação e Controle
>>
>
>
>
> --
> Roberto Spadim
> SPAEmpresarial - Software ERP
> Eng. Automação e Controle
>