and a production copy of most of my databases. They are on two
different machines. The databases used to tuned the same way,
however one of the machines has more processing power and one has
more memory. When we retuned the databases to take advantage of the
machines strong points, it decreased the time it took to run some
queries by 400%.
Carol
P.S. If I understand your process, and your deleting the records as
you go, that would make me really nervous. As soon as you start, you
no longer have an intact table that has all the data in it. While
modern databases engines do a lot to protect your data, there is
always some quirk that can happen. If you have enough space, you
might consider running the delete after the tables are created.
On Sep 19, 2008, at 11:07 AM, Howard Eglowstein wrote:
> There are a lot of issues at work here. The speed of the machine,
> the rest of the machine's workload, the database configuration,
> etc. This machine is about 3 years old and not as fast as a test
> machine I have at my desk. It's also running three web services and
> accepting new data into the current year's tables at the rate of
> one set of rows every few seconds. The database when I started
> didn't have any indices applied. I indexed a few columns which
> seemed to help tremendously (a factor of 10 at least) and perhaps a
> few more might help.
>
> Considering that searching the tables now with the data split into
> 3 rows takes a minute or more to search the whole database, I
> suspect that there's still organizational issues that could be
> addressed to speed up all PG operations. I'm far more concerned
> with robustness and I'm not too keen on trying too many experiments
> until I get the data broken up and backed up again.
>
> I doubt this machine could perform 7 SQL operations on 1.5 million
> rows in each of 3 tables in a few seconds or minutes on a good day,
> with the wind, rolling down hill. I'd like to be proven wrong
> though...
>
> Howard
>
> Sean Davis wrote:
>> On Fri, Sep 19, 2008 at 10:48 AM, Howard Eglowstein
>> <howard@yankeescientific.com> wrote:
>>
>>> Absolutely true, and if the data weren't stored on the same
>>> machine which is
>>> running the client, I would have worked harder to combine
>>> statements. In
>>> this case though, the server and the data are on the same machine
>>> and the
>>> client application doing the SELECT, INSERT and DELETEs is also
>>> on the same
>>> machine.
>>>
>>> I'd like to see how to have done this with combined statements if
>>> I ever
>>> have to do it again in a different setup, but it is working well
>>> now. It's
>>> moved about 1/2 million records so far since last night.
>>>
>>
>> So the 150ms was per row? Not to belabor the point, but I have done
>> this with tables with tens-of-millions of rows in the space of
>> seconds
>> to minutes (for the entire move, not per row), depending on the exact
>> details of the table(s). No overnight involved. The network is one
>> issue (which you have avoided by being local), but the encoding and
>> decoding overhead to go to a client is another one that is entirely
>> avoided. When you have some free time, do benchmark, as I think the
>> difference could be substantial.
>>
>> Sean
>>
>>
>>> Sean Davis wrote:
>>>
>>>> On Fri, Sep 19, 2008 at 7:42 AM, Howard Eglowstein
>>>> <howard@yankeescientific.com> wrote:
>>>>
>>>>
>>>>> So you'd agree then that I'll need 7 SQL statements but that I
>>>>> could
>>>>> stack
>>>>> the INSERT and the first SELECT if I wanted to? Cool. That's
>>>>> what I ended
>>>>> up
>>>>> with in C code and it's working pretty well. I did some
>>>>> indexing on the
>>>>> database and got the whole transaction down to about 150ms for the
>>>>> sequence.
>>>>> I guess that's as good as it's going to get.
>>>>>
>>>>>
>>>> Keep in mind that the INSERT [...] SELECT [...] is done server-
>>>> side,
>>>> so the data never goes over the wire to the client. This is very
>>>> different than doing the select, accumulating the data, and then
>>>> doing
>>>> the insert and is likely to be much faster, relatively. 150ms is
>>>> already pretty fast, but the principle of doing as much on the
>>>> server
>>>> as possible is an important one when looking for efficiency,
>>>> especially when data sizes are large.
>>>>
>>>> Glad to hear that it is working.
>>>>
>>>> Sean
>>>>
>>>>
>>>>
>>>>> Sean Davis wrote:
>>>>>
>>>>>
>>>>>> On Thu, Sep 18, 2008 at 7:28 PM, Howard Eglowstein
>>>>>> <howard@yankeescientific.com> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>>> What confuses me is that I need to do the one select with all
>>>>>>> three
>>>>>>> tables
>>>>>>> and then do three inserts, no? The results is that the 150
>>>>>>> fields I get
>>>>>>> back
>>>>>>> from the select have to be split into 3 groups of 50 fields
>>>>>>> each and
>>>>>>> then
>>>>>>> written into three tables.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>> You do the insert part of the command three times, once for
>>>>>> each new
>>>>>> table, so three separate SQL statements. The select remains
>>>>>> basically
>>>>>> the same for all three, with only the column selection changing
>>>>>> (data_a.* when inserting into new_a, data_b.* when inserting into
>>>>>> new_b, etc.). Just leave the ids the same as in the first set of
>>>>>> tables. There isn't a need to change them in nearly every
>>>>>> case. If
>>>>>> you need to add a new ID column, you can do that as a serial
>>>>>> column in
>>>>>> the new tables, but I would stick to the original IDs, if
>>>>>> possible.
>>>>>>
>>>>>> Sean
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>> What you're suggesting is that there is some statement which
>>>>>>> could do
>>>>>>> the
>>>>>>> select and the three inserts at once?
>>>>>>>
>>>>>>> Howard
>>>>>>>
>>>>>>> Sean Davis wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>> You might want to look at insert into ... select ...
>>>>>>>>
>>>>>>>> You should be able to do this with 1 query per new table (+ the
>>>>>>>> deletes, obviously). For a few thousand records, I would
>>>>>>>> expect that
>>>>>>>> the entire process might take a few seconds.
>>>>>>>>
>>>>>>>> Sean
>>>>>>>>
>>>>>>>> On Thu, Sep 18, 2008 at 6:39 PM, Howard Eglowstein
>>>>>>>> <howard@yankeescientific.com> wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>> Somewhat empty, yes. The single set of 'data_' tables
>>>>>>>>> contains 3
>>>>>>>>> years
>>>>>>>>> worth
>>>>>>>>> of data. I want to move 2 years worth out into the 'new_'
>>>>>>>>> tables.
>>>>>>>>> When
>>>>>>>>> I'm
>>>>>>>>> done, there will still be 1 year's worth of data left in
>>>>>>>>> the original
>>>>>>>>> table.
>>>>>>>>>
>>>>>>>>> Howard
>>>>>>>>>
>>>>>>>>> Carol Walter wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>> What do you want for your end product? Are the old tables
>>>>>>>>>> empty
>>>>>>>>>> after
>>>>>>>>>> you
>>>>>>>>>> put the data into the new tables?
>>>>>>>>>>
>>>>>>>>>> Carol
>>>>>>>>>>
>>>>>>>>>> On Sep 18, 2008, at 3:02 PM, Howard Eglowstein wrote:
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>> I have three tables called 'data_a', 'data_b' and
>>>>>>>>>>> 'data_c' which
>>>>>>>>>>> each
>>>>>>>>>>> have 50 columns. One of the columns in each is 'id' and
>>>>>>>>>>> is used to
>>>>>>>>>>> keep
>>>>>>>>>>> track of which data in data_b and data_c corresponds to a
>>>>>>>>>>> row in
>>>>>>>>>>> data_a. If
>>>>>>>>>>> I want to get all of the data in all 150 fields for this
>>>>>>>>>>> month (for
>>>>>>>>>>> example), I can get it with:
>>>>>>>>>>>
>>>>>>>>>>> select * from (data_a, data_b, data_c) where
>>>>>>>>>>> data_a.id=data_b.id
>>>>>>>>>>> AND
>>>>>>>>>>> data_a.id = data_c.id AND timestamp >= '2008-09-01
>>>>>>>>>>> 00:00:00' and
>>>>>>>>>>> timestamp
>>>>>>>>>>> <= '2008-09-30 23:59:59'
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>
>>>>>>>>>>> What I need to do is execute this search which might
>>>>>>>>>>> return several
>>>>>>>>>>> thousand rows and write the same structure into 'new_a',
>>>>>>>>>>> 'new_b'
>>>>>>>>>>> and
>>>>>>>>>>> 'new_c'. What i'm doing now in a C program is executing
>>>>>>>>>>> the search
>>>>>>>>>>> above.
>>>>>>>>>>> Then I execute:
>>>>>>>>>>>
>>>>>>>>>>> INSERT INTO data_a (timestamp, field1, field2 ...[imagine
>>>>>>>>>>> 50 of
>>>>>>>>>>> them])
>>>>>>>>>>> VALUES ('2008-09-01 00:00:00', 'ABC', 'DEF', ...);
>>>>>>>>>>> Get the ID that was assigned to this row since 'id' is a
>>>>>>>>>>> serial
>>>>>>>>>>> field
>>>>>>>>>>> and
>>>>>>>>>>> the number is assigned sequentially. Say it comes back as
>>>>>>>>>>> '1'.
>>>>>>>>>>> INSERT INTO data_b (id, field1, field2 ...[imagine 50 of
>>>>>>>>>>> them])
>>>>>>>>>>> VALUES
>>>>>>>>>>> ('1', 'ABC', 'DEF', ...);
>>>>>>>>>>> INSERT INTO data_c (id, field1, field2 ...[imagine 50 of
>>>>>>>>>>> them])
>>>>>>>>>>> VALUES
>>>>>>>>>>> ('1', 'ABC', 'DEF', ...);
>>>>>>>>>>>
>>>>>>>>>>> That moves a copy of the three rows of data form the
>>>>>>>>>>> three tables
>>>>>>>>>>> into
>>>>>>>>>>> the three separate new tables.
>>>>>>>>>>> From the original group of tables, the id for these rows
>>>>>>>>>>> was, let's
>>>>>>>>>>> say,
>>>>>>>>>>> '1234'. Then I execute:
>>>>>>>>>>>
>>>>>>>>>>> DELETE FROM data_a where id='1234';
>>>>>>>>>>> DELETE FROM data_b where id='1234';
>>>>>>>>>>> DELETE FROM data_c where id='1234';
>>>>>>>>>>>
>>>>>>>>>>> That deletes the old data.
>>>>>>>>>>>
>>>>>>>>>>> This works fine and gives me exactly what I wanted, but
>>>>>>>>>>> is there a
>>>>>>>>>>> better
>>>>>>>>>>> way? This is 7 SQL calls and it takes about 3 seconds per
>>>>>>>>>>> moved
>>>>>>>>>>> record
>>>>>>>>>>> on
>>>>>>>>>>> our Linux box.
>>>>>>>>>>>
>>>>>>>>>>> Any thoughts or suggestions would be appreciated.
>>>>>>>>>>>
>>>>>>>>>>> Thanks,
>>>>>>>>>>>
>>>>>>>>>>> Howard
>>>>>>>>>>>
>>>>>>>>>>> --
>>>>>>>>>>> Sent via pgsql-novice mailing list (pgsql-
>>>>>>>>>>> novice@postgresql.org)
>>>>>>>>>>> To make changes to your subscription:
>>>>>>>>>>> http://www.postgresql.org/mailpref/pgsql-novice
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>> -------------------------------------------------------------
>>>>>>>>>> -----------
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> No virus found in this incoming message.
>>>>>>>>>> Checked by AVG - http://www.avg.com Version: 8.0.169 / Virus
>>>>>>>>>> Database:
>>>>>>>>>> 270.6.21/1678 - Release Date: 9/18/2008 9:01 AM
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>> --
>>>>>>>>> Sent via pgsql-novice mailing list (pgsql-
>>>>>>>>> novice@postgresql.org)
>>>>>>>>> To make changes to your subscription:
>>>>>>>>> http://www.postgresql.org/mailpref/pgsql-novice
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>> ---------------------------------------------------------------
>>>>>>>> ---------
>>>>>>>>
>>>>>>>>
>>>>>>>> No virus found in this incoming message.
>>>>>>>> Checked by AVG - http://www.avg.com Version: 8.0.169 / Virus
>>>>>>>> Database:
>>>>>>>> 270.6.21/1678 - Release Date: 9/18/2008 9:01 AM
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>> -----------------------------------------------------------------
>>>>>> -------
>>>>>>
>>>>>>
>>>>>> No virus found in this incoming message.
>>>>>> Checked by AVG - http://www.avg.com Version: 8.0.169 / Virus
>>>>>> Database:
>>>>>> 270.7.0/1679 - Release Date: 9/18/2008 5:03 PM
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>> -------------------------------------------------------------------
>>>> -----
>>>>
>>>>
>>>> No virus found in this incoming message.
>>>> Checked by AVG - http://www.avg.com Version: 8.0.169 / Virus
>>>> Database:
>>>> 270.7.0/1679 - Release Date: 9/18/2008 5:03 PM
>>>>
>>>>
>>>>
>> >
>> ---------------------------------------------------------------------
>> ---
>>
>>
>> No virus found in this incoming message.
>> Checked by AVG - http://www.avg.com Version: 8.0.169 / Virus
>> Database: 270.7.0/1679 - Release Date: 9/18/2008 5:03 PM
>>
>>
>
>
> --
> Sent via pgsql-novice mailing list (pgsql-novice@postgresql.org)
> To make changes to your subscription:
> http://www.postgresql.org/mailpref/pgsql-novice
--
Sent via pgsql-novice mailing list (pgsql-novice@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-novice
No comments:
Post a Comment