On 9/10/2007 12:49 PM, Cyril SCETBON wrote:
> 
> Christopher Browne wrote:
>> Cyril SCETBON wrote:
>>>
>>>
>>> Jan Wieck wrote:
>>>> On 9/10/2007 10:07 AM, Cyril SCETBON wrote:
>>>>> Hi,
>>>>>
>>>>> We're hitting 400 requests per second and noticed that sl_log_? are 
>>>>> growing a lot. This is also due to the fact that our tables have a 
>>>>> lot of columns (more that 100). We noticed that slony stores each 
>>>>> column and its value for each request in sl_log_x. Would it not be 
>>>>> more optimized to store just the user request ? Actually, a user is 
>>>>> just updating a few column, or inserting a tuple by giving some 
>>>>> attributes but not all, however as slony stores every column with 
>>>>> its value it sl_log_x, these tables are growing very fast :-(
>>>>
>>>> Slony only logs columns where the value actually has changed. Please 
>>>> explain in detail how you propose that the log table would contain 
>>>> which columns have changed without naming them.
>>> when I execute this command on the master :
>>>
>>> insert into t1(ise) values('cyril100001');
>>>
>>> I can see in sl_log_1 :
>>>
>>> psql>select log_cmddata from sl_log_1;
>>> (ise,id,id2) values ('cyril100001','24161',NULL)
>>>
>>> dbtest=# \d t1;
>>>                                Table "public.t1"
>>> Column |         Type          |                    
>>> Modifiers                   
>>> --------+-----------------------+------------------------------------------------- 
>>>
>>> ise    | character varying(54) | not null
>>> id     | integer               | not null default 
>>> nextval('t1_id_seq'::regclass)
>>> id2    | integer               |
>>> Indexes:
>>>    "t1_pkey" PRIMARY KEY, btree (ise)
>>> Triggers:
>>>    "_CLUSTER1_logtrigger_1" AFTER INSERT OR DELETE OR UPDATE ON t1 
>>> FOR EACH ROW EXECUTE PROCEDURE "_CLUSTER1".logtrigger('_CLUSTER1', 
>>> '1', 'kvv')
>>>
>>> I agree for id which is an auto_increment but not for id2 :-(
>>>
>>> In another plateform my table has more than  100 columns as said 
>>> earlier and that's really matter :-(
>> Jan's request seems to remain perfectly good...
>>
>> "Please explain in detail how you propose that the log table would 
>> contain which columns have changed without naming them."
>>
>> It is *not* obvious how to avoid naming all of the columns.  It is not 
>> safe to simply assume "that column was null so we may omit it" - 
>> different nodes may be configured differently, and it may well be 
>> important to actually have that NULL value.
> I was just thinking of using the same way as other statement based 
> replication way like does mysql, that is to say just log the request 
> that user did. However, I forgot that slony uses triggers to log 
> requests :-( Maybe finding a way to know which current request is done 
> on the table for which the trigger has been called would be possible ?

Turning Slony into a statement based replication system would make it 
far worse. Queries depend on the exact transaction commit order with 
respect to the exact rows that are visible to them. That in addition to 
the non-deterministic behavior of sequential scans and a myriad of 
functions ... how do you expect some query like

     update mytable set foo = nextval('someseq') where foo is null;

to lead to exactly the same result on two different databases? Not even 
the order in which the rows are processed is defined.

Compared to Postgres, MySQL is a fairly limited system feature wise. Yet 
with their limited functionality the MySQL user manual already states 
that several features are not supported by their statement based 
replication. What do you think the corresponding documentation for Slony 
would look like?


Jan

-- 
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#================================================== JanWieck at Yahoo.com #