Speeding Up SQL Query: Grouping & Joining

[BWV]

I have a T-SQL query (MS SQL Server) that runs slowly due to the six inner joins I need to get the data I want. The seven tables of data I am working with consist of four of category data and three of quantitative - and the query, which is not important to the question, just makes calculations based on classification and filtering of the category data

question is - the four category tables are all the same key, they just have different information I need on them. So do you all think I could materially speed up the query by first grouping the four category tables into a single one with only the data I want then joining that to the quant data?

Not sure if I am permissioned to create a table, so may have to use a nested query to do this

 

[FactChecker]

I think it should help a lot. That being said, computer speed issues can be very tricky and what I think will be faster can be slower in fact. You should conduct a timing test. If you do, I would be interested. If the result is so significant that no timing is needed, that is enough to know.

 

[BWV]

Will give it a try. FWIW one minor tweak I made after some research was replacing = with IN in the WHERE and some CASE WHEN statements and it cut a 10:54 runtime by a few seconds to 10:47

 

[BWV]

Negligible difference - 10:47 down to 10:41 using a nested query. the nested query on its own runs too fast for mgmt studio to time, so maybe be the case statements in SELECT or some other issue

 

[WWGD]

Have you tried doing a trace (Profiler, IIRC) ? Maybe translating to Relational Algebra?Execution Plan
Activity Monitor?

 

[jedishrfu]

Have you built indexes for the tables in question? that is one way to speed up a join operation.

https://support.microsoft.com/en-us...formance-0a8e2aa6-735c-4c3a-9dda-38c6c4f1a0ce

 

[BWV]

Thanks,

Have not messed with profiler

The tables are actually views on a third party app, they are indexed, but I have no ability to alter them

 

[pbuk]

I think I'd probably denormalize the data into 3 or fewer tables and query against that. If it is practical to do this each time you want to run the report then you have a solution, if not then you have an idea of the minimum time you are aiming at with optimizations.

Also, if the slow times are driven by the volume of data rather than the complexity of calculations think about how you can parallelize the process - or even run serially in smaller batches to make more use of faster cache levels.

 

[jack action]

Have you tried the EXPLAIN command?

 

[FactChecker]

The tables are actually views on a third party app, they are indexed, but I have no ability to alter them

Are you saying that you don't have permission to join (inner, outer, left, or right as appropriate) the four tables together with the common key and save them before the query? In that case, I wouldn't be surprised if there was no timing improvement if the time includes the join.
I thought that you were talking about joining them once to get a larger table that could be saved and used in many queries later. In that case, I would expect the time required for a single query to be much less than the time required for four separate queries.

 

[BWV]

Appreciate all the responses, several good avenues to explore for a hack programmer

Ultimately the thinking in my OP comes down to misunderstanding that the JOIN operator is associative so nesting the joins should have no effect on run time

Also my nested query of categorical data took literally 0 time to run, so that can’t be the issue

The quant data tables are very large, in one to many with the cat data and many to many with one another, so that likely is the source of the slow run time

 

[pbuk]

The quant data tables are very large, in one to many with the cat data and many to many with one another, so that likely is the source of the slow run time

No doubt. That is going to be particularly slow if the common fields on the many to many are not all indexed, so that's the first thing to try on a snapshot before denormalizing.

 

[pasmith]

Appreciate all the responses, several good avenues to explore for a hack programmer

Ultimately the thinking in my OP comes down to misunderstanding that the JOIN operator is associative so nesting the joins should have no effect on run time

Also my nested query of categorical data took literally 0 time to run, so that can’t be the issue

The quant data tables are very large, in one to many with the cat data and many to many with one another, so that likely is the source of the slow run time

It sounds like you are describing fact tables in a star schema, and you never want to join those to each other: Firstly this is very slow, and secondly it might result in a single row of a fact table being repeated multiple times in the joined table, which will make all of your aggregations inaccurate.

Instead you should aggregate and filter the fact tables individually, and then outer join the results with the category tables on the common category keys.

 

[BWV]

Tried some pre-filtered views on the large (~10^7 records) data tables, thought by pre-filtering them with a trailing 180 date window it would speed up some of the stuff I run, but the runtime is essentially unchanged. Now messing around with trying to create an indexed view with CREATE UNIQUE CLUSTERED INDEX, but running into permission issues, so will likely just give up. You can't schemabind any of the views in this DB to create an index so need to go to the main tables, but run into permissioning issues as its a 3rd party app and don't want to mess with the core data tables too much

 

[pbuk]

Tried some pre-filtered views on the large (~10^7 records) data tables

Doesn't sound that large, have you tried just loading the lot into memory and using your programming tool of choice?

 

[BWV]

Doesn't sound that large, have you tried just loading the lot into memory and using your programming tool of choice?

it is manageable, but reducing the current 10-20 minute runtime would improve workflows

 

[pbuk]

I don't understand, are you saying you are currently loading 10^7 records into some suitable data structure in e.g Python and it is taking 10-20 minutes for Python to run the computations?

 

[BWV]

No just talking about the query time in SSMS

 

[jack action]

I wish I could at least see the query. Seeing the table definitions would be a plus.

What you are trying to do - rewriting your query to eliminate the JOINs - is most likely pointless. The program is always using the most optimized way to solve your query. Thinking you could do it better than the program is most likely not going to happen.

You have to make sure that the query is written such that it works well with the built-in optimizer AND that your data is well stored (most simple data type and useful indexing).

 

[pbuk]

No just talking about the query time in SSMS

Yes, so I am saying avoid the query altogether, just send the tables over and let Python or whatever do the heavy lifting.

 

[FactChecker]

If you can not do a join once and save the result for later use multiple times, I think there is very little improvement that you can make. If you do not have permission to create and save a joined table, you should consult the data administrator to see what he will suggest and allow. There might be serious issues with keeping a separate, joined table up to date. There also might be storage space issues. It sounds to me as though the database might be poorly designed, with the same key repeated many times in separate, single-data-column tables. If so, you might be out of luck. One reason for designing the database that way might be to have unique rules of access for each data type.

 

[FactChecker]

Yes, so I am saying avoid the query altogether, just send the tables over and let Python or whatever do the heavy lifting.

If speed is the issue, I would have serious doubts about Python. It is relatively slow. A professional database program should be orders of magnitude faster.

 

[PeterDonis]

If speed is the issue, I would have serious doubts about Python.

Python has third-party libraries for all of the popular database engines that use C extension modules to do all the heavy lifting. There isn't really a speed impact on DB operations in Python if you use these.

 

[FactChecker]

Python has third-party libraries for all of the popular database engines that use C extension modules to do all the heavy lifting. There isn't really a speed impact on DB operations in Python if you use these.

There is no reason that it would be faster than the professional database program if it uses the same algorithms in the same language. The process of pulling the data out for any Python manipulations would add time and might be slow.

 

[pbuk]

If speed is the issue, I would have serious doubts about Python. It is relatively slow. A professional database program should be orders of magnitude faster.

No, no, and no.

I can't blame you for repeating the often quoted myth that "Python is slow", because it is often repeated, but it is also a myth.

And "professional database programs" are only fast when (i) doing things that "professional database programs" are good at, and (ii) they are properly structured for the task at hand. In this case it seems that neither criterion is fulfilled.

 

[FactChecker]

No, no, and no.

I can't blame you for repeating the often quoted myth that "Python is slow", because it is often repeated, but it is also a myth.

It's the result of timing tests.

 

[pbuk]

It's the result of timing tests.

What timing tests are relevant to the problem here, which is not even known?

 

[FactChecker]

What timing tests are relevant to the problem here, which is not even known?

Of course, I am talking in general. In general, Python is a fairly slow language.
If Python can do everything through calls to more efficient subroutines and never has to bring the data into the Python environment, then fine. But in that case, it should be doing the same thing that the original database program is doing. I wouldn't call that having Python "do the heavy lifting".

 

[PeterDonis]

There is no reason that it would be faster than the professional database program if it uses the same algorithms in the same language.

I'm not sure what you mean by a "professional database program", but the C libraries that the Python wrappers use for all of the common database engines are the same libraries that the command line clients for those engines use, and those libraries are provided by the database engine developers. So the Python wrappers should be running the same C code "under the hood" that the database engine developers use.

 

[pbuk]

Of course, I am talking in general. In general, Python is a fairly slow language.
If Python can do everything through calls to more efficient subroutines and never has to bring the data into the Python environment, then fine. But in that case, it should be doing the same thing that the original database program is doing. I wouldn't call that having Python "do the heavy lifting".

There seems to be a big misunderstanding here.

We have a query over 10^7 records (or is it 3x10^7 records) that is taking 10-20 minutes to run. That can only be becuase the data is not structured and indexed in an appropriate way for this particular computation. We cannot, apparently, change the structure of the database in MS SQL Server because the database is proprietory.

What we can do, however, is pull the (lets assume 10GB of) data over to a workstation running (any language that has an effective hash table implementation) so that we can structure it in a way that is appropriate for this particular computation.

The fact that Python is slowed by a small constant factor due to running a P-code virtual machine is insignificant compared to the exponential gains that can be made by structuring and indexing the data in an appropriate way.