List:General Discussion« Previous MessageNext Message »
From:mos Date:March 31 2011 4:20pm
Subject:Re: A common request
View as plain text  
At 07:29 AM 3/31/2011, you wrote:
>Thanks for your insight! But I'm still worried about the performance of IN 
>( big list of values ). Can you tell me how it is implemented?
>
>Suppose I have SELECT a FROM b WHERE c IN (1, 4, 5, 117, 118, 119, ..., 
>387945)
>
>1) If I put 200 values there, does it do 200 individual SELECTs 
>internally, and union them?

No. It uses one Select statement.

>Or does it notice that c has a "UNIQUE" index and thus at most one row can 
>be returned per SELECT, and does them all at once?

The IN() clause is very inefficient because MySQL will NOT use the index. 
It will have to traverse the entire table looking for these values. That is 
why a table join will be much faster than using IN().


>2) If I want to get just the primary key, or join with another table based 
>on just the primary key, does this query ever touch the disk (assuming the 
>index is in memory, which I think it always is -- correct me if I'm wrong 
>about that).

It will get the information from the index and not have to access the 
record data from disk. If the index is stored in memory, then it won't have 
to go to disk (unless you also have a sort). That is why the query cache is 
so important.


>The way I would recommend doing it (for BTREE indexes, anyway) is to sort 
>the values in ascending order, and do the search in one pass through the 
>index. The index is already in memory, and it would be straightforward to 
>modify a binary search algorithm to find the rows corresponding to 
>monotonically ascending values of the primary key, all in one pass.
>
>Even if the binary search algorithm is run 200 or 2000 times for a list, 
>it would still be faster than hitting the disk. (Even though the CPU cache 
>performance would be worse.)
>
>Can you let me know the specifics of it, and especially how I can avoid 
>hitting the I/O bottlenecks?

Use a table join and make sure you have the indexes loaded into memory. See 
http://dev.mysql.com/doc/refman/5.0/en/load-index.html. If using InnoDb 
then its index cache scheme is quite good.


Mike


>Thank you,
>Greg
>
>On 3/29/11 4:17 PM, Peter Brawley wrote:
>> > Why not optimize the IN ( ... ) to do the same type of thing?
>>
>>If the argument to IN() is a list of values, it'll be OK. If it's a 
>>SELECT, in 5.0 it will be slower than molasses (see "The unbearable 
>>slowness of IN()" at http://www.artfulsoftware.com/queries.php.
>>
>> > I always tried to avoid joins because I am planning to horizontally 
>> partition my data.
>>
>>A severe & unfortunate constraint. Can't help you there.
>>
>>PB
>>
>>-----
>>
>>On 3/29/2011 1:27 PM, Gregory Magarshak wrote:
>>>Yes, this would be fine. But often, the list of friends is obtained from 
>>>a social network like facebook, and is not stored internally. Basically, 
>>>I obtain the friend list in a request to facebook, and then see which of 
>>>those users have created things. So would I have to create a temporary 
>>>table and insert all those uids just to make a join? Why not optimize 
>>>the IN ( ... ) to do the same type of thing?
>>>
>>>There is also a second problem: I want to use MySQL Cluster, because I 
>>>expect to have many users. Would it be efficient to use JOIN between the 
>>>friends table and the articles table? Both tables are partitioned by 
>>>user_id as the primary key, so the join would have to hit many different 
>>>nodes. I always tried to avoid joins because I am planning to 
>>>horizontally partition my data. But if MySQL cluster can handle this 
>>>join transparently and split it up based on the partition, then that's 
>>>fine. Do you have any info on this?
>>>
>>>Greg
>>>
>>>On 3/29/11 2:10 PM, Peter Brawley wrote:
>>>> > How can I quickly find all the articles written by this user's 
>>>> friends, and not just random articles?
>>>>
>>>>Taking the simplest possible case, with table friends(userID,friendID) 
>>>>where each friendID refers to a userID in another row, the friends of 
>>>>userID u are ...
>>>>
>>>>select friendID from user where userID=u;
>>>>
>>>>so articles by those friends of u are ...
>>>>
>>>>select a.* from article a join (  select friendID from user where 
>>>>userID=u ) f on a.userID=f.friendID;
>>>>
>>>>PB
>>>>
>>>>-----
>>>>
>>>>On 3/29/2011 12:50 PM, Gregory Magarshak wrote:
>>>>>Hey there. My company writes a lot of social applications, and there 
>>>>>is one operation that is very common, but I don't know if MySQL 
>>>>>supports it in a good way. I thought I'd write to this list for two
> reasons:
>>>>>
>>>>>     1) Maybe MySQL has a good way to do this, and I just don't know 
>>>>> about it
>>>>>
>>>>>     2) Propose to MySQL developers a simple algorithm which would 
>>>>> greatly improve MySQL support for social networking apps.
>>>>>
>>>>>     Here is the situation. Let's say I have built a social networking
> 
>>>>> application where people create and edit some item (article, photo, 
>>>>> music mix, whatever). Now, a typical user logs in, and this user has
> 
>>>>> 3000 friends. How can I quickly find all the articles written by this
> 
>>>>> user's friends, and not just random articles?
>>>>>
>>>>>     Ideally, I would want to write something like this:
>>>>>
>>>>>     SELECT * FROM article WHERE user_id IN (345789, 324875, 398, ...,
> 
>>>>> 349580)
>>>>>
>>>>>     basically, execute a query with a huge IN ( ... ). Maybe if this
> 
>>>>> would exceed the buffer size for the MySQL wire protocol, I would 
>>>>> break up the list into several lists, and execute several queries, 
>>>>> and union the results together myself.
>>>>>
>>>>>     But my point is, this is very common for social networking apps.
> 
>>>>> Every app wants to show "the X created by your friends", or "friends
> 
>>>>> of yours (given some list from a social network) who have taken
> action X".
>>>>>
>>>>>     Here is how I would do it if I had raw access to the MySQL index
> 
>>>>> in memory:
>>>>>
>>>>>     a) Sort the list of entries in the IN, in ascending order.
>>>>>
>>>>>     b) Do *ONE* binary search through the index (assuming it's a 
>>>>> BTREE index) and get them all in one pass. If it's a HASH index or 
>>>>> something, I would have to look up each one individually.
>>>>>
>>>>>     The benefits of this approach would be that this common operation
> 
>>>>> would be done extremely quickly. If the index fits entirely in 
>>>>> memory, and I just want to get the primary keys (i.e. get the list of
> 
>>>>> friends who did X), the disk isn't even touched. In addition, for 
>>>>> BTREE indexes, I would just need ONE binary search, because the 
>>>>> entries have been sorted in ascending order.
>>>>>
>>>>>     Does MySQL have something like this? And if not, perhaps you can
> 
>>>>> add it in the next version? It would really boost MySQL's support for
> 
>>>>> social networking apps tremendously. Alternative, how can I add this
> 
>>>>> to my MySQL? Any advice would be appreciated.
>>>>>
>>>>>Sincerely,
>>>>>Gregory Magarshak
>>>>>Qbix
>>>
>
>
>--
>MySQL General Mailing List
>For list archives: http://lists.mysql.com/mysql
>To unsubscribe:    http://lists.mysql.com/mysql?unsub=1

Thread
A common requestGregory Magarshak29 Mar
  • Re: A common requestPeter Brawley29 Mar
    • Re: A common requestGregory Magarshak29 Mar
      • Re: A common requestPeter Brawley29 Mar
        • Re: A common requestGregory Magarshak31 Mar
          • Re: A common requestGregory Magarshak31 Mar
            • Re: A common requestJohan De Meersman31 Mar
          • Re: A common requestmos31 Mar
            • Re: A common requestJohan De Meersman31 Mar
            • Re: A common requestmos31 Mar
              • Re: A common requestWm Mussatto31 Mar
  • Re: A common requestSander de Bruijne29 Mar