My system does some pretty heavy processing, and I've been attacking the performance in order to give me the ability to run more test runs in shorter times.
I have quite a few cases where a UDF has to get called on say, 5 million rows (and I pretty much thought there was no way around it).
Well, it turns out, there is a way to work around it and it gives huge performance improvements when UDFs are called over a set of distinct parameters somewhat smaller than the total set of rows.
Consider a UDF that takes a set of inputs and returns a result based on complex logic, but for the set of inputs over 5m rows, there are only 100,000 distinct inputs, say, and so it will only produce 100,000 distinct result tuples (my particular cases vary from interest rates to complex code assignments, but they are all discrete - the fundamental point with this technique is that you can simply determine if the trick will work by running the SELECT DISTINCT).
I found that by doing something like this:
INSERT INTO PreCalcs
SELECT param1
,param2
,dbo.udf_result(param1, param2) AS result
FROM (
SELECT DISTINCT param1, param2 FROM big_table
)
When PreCalcs is suitably indexed, the combination of that with:
SELECT big_table.param1
,big_table.param2
,PreCalcs.result
FROM big_table
INNER JOIN PreCalcs
ON PreCalcs.param1 = big_table.param1
AND PreCalcs.param2 = big_table.param2
You get a HUGE boost in performance. Apparently, just because something is deterministic, it doesn't mean SQL Server is caching the past calls and re-using them, as one might think.
The only thing you have to watch out for is where NULL are allowed, then you need to fix up your joins carefully:
SELECT big_table.param1
,big_table.param2
,PreCalcs.result
FROM big_table
INNER JOIN PreCalcs
ON (
PreCalcs.param1 = big_table.param1
OR COALESCE(PreCalcs.param1, big_table.param1) IS NULL
)
AND (
PreCalcs.param2 = big_table.param2
OR COALESCE(PreCalcs.param2, big_table.param2) IS NULL
)
Hope this helps and any similar tricks with UDFs, or refactoring queries for performance are welcome.
I guess the question is, why is manual caching like this necessary - isn't that the point of the server knowing that the function is deterministic? And if it makes such a big difference, and if UDFs are so expensive, why doesn't the optimizer just do it in the execution plan?
Yes, the optimizer will not manually memoize UDFs for you. Your trick is very nice in the cases where you can collapse the output set down in this way.
Another technique that can improve performance if your UDF's parameters are indices into other tables, and the UDF selects values from those tables to calculate the scalar result, is to rewrite your scalar UDF as a table-valued UDF that selects the result value over all your potential parameters.
I've used this approach when the tables we based the UDF query on were subject to a lot of inserts and updates, the involved query was relatively complex, and the number of rows the original UDF had to be applied to were large. You can achieve some great improvement in performance in this case, as the table-values UDF only needs to be run once and can run as an optimized set-oriented query.
How would SQL Server know that you have 100,000 discrete combinations within 5 million rows?
By using the PreCalcs table, you are simply running the udf over 100k rows rather that 5 million rows, before expanding back out again.
No optimiser in existence would be able to divine this useful information.
The scalar udf is a black box.
For a more practical solution, I'd use a computed, persisted columns that does the udf call.
So it's available in all queries can be indexed/included.
This suits OLTP more, maybe... I query a table to get trading cash and positions in real time in many different ways so this approach suits me to avoid the udf math overhead every time.
Related
I have a hard time figuring out what is best, or if there is difference at all,
however i have not found any material to help my understanding of this,
so i will ask this question, if not for me, then for others who might end up in the same situation.
Aggregating a sub-query before or after a join, in my specific situation the sub-query is rather slow due to fragmented data and bad normalization procedure,
I got a main query that is highly complex and a sub-query that is built from 3 small queries that is combined using union (will remove duplicate records)
i only need a single value from this sub-query (for each line), so at some point i will end up summing this value, (together with grouping the necessary control data with it so i can join)
what will have the greatest impact?
To sum sub-query before the join and then join with the aggregated version
To leave the data raw, and then sum the value together with the rest of the main query
remember there are thousands of records that will be summed for each line,
and the data is not native but built, and therefore may reside in memory,
(that is just a guess from the query optimizers perspective)
Usually I keep the group-by inside the subquery (referred as "inline view" in Oracle lingo).
This way the query is much more simple and clear.
Also I believe the execution plan is more efficient, because the data set to be aggregated is smaller and the resulting set of join keys is also smaller.
This is not a definitive answer though. If the row source that you are joining to the inline view has few matching rows, you may find that a early join reduces the aggregation effort.
The right anwer is: benchmark the queries for your particular data set.
I think in such a general way there is no right or wrong way to do it. The performance from a query like the one that you describe depends on many different factors:
what kind of join are you actually doing (and what algorithm is used in the background)
is the data to be joined small enough to fit into the memory of the machine joining it?
what query optimizations are you using, i.e. what DBMS (Oracle, MsSQL, MySQL, ...)
...
For your case I simply suggest benchmarking. I'm sorry if that does not seem like a satisfactory answer, but it is the way to go in many performance questions...
So set up a simple test using both your approaches and some test data, then pick whatever is faster.
This is a never-ending topic for me and I'm wondering if I might be overlooking something. Essentially I use two types of SQL statements in an application:
Regular queries with a "fallback" limit
Sorted and paged queries
Now, we're talking about some queries against tables with several million records, joined to 5 more tables with several million records. Clearly, we hardly want to fetch all of them, that's why we have the above two methods to limit user queries.
Case 1 is really simple. We just add an additional ROWNUM filter:
WHERE ...
AND ROWNUM < ?
That's quite fast, as Oracle's CBO will take this filter into consideration for its execution plan and probably apply a FIRST_ROWS operation (similar to the one enforced by the /*+FIRST_ROWS*/ hint.
Case 2, however is a bit more tricky with Oracle, as there is no LIMIT ... OFFSET clause as in other RDBMS. So we nest our "business" query in a technical wrapper as such:
SELECT outer.* FROM (
SELECT * FROM (
SELECT inner.*, ROWNUM as RNUM, MAX(ROWNUM) OVER(PARTITION BY 1) as TOTAL_ROWS
FROM (
[... USER SORTED business query ...]
) inner
)
WHERE ROWNUM < ?
) outer
WHERE outer.RNUM > ?
Note that the TOTAL_ROWS field is calculated to know how many pages we will have even without fetching all data. Now this paging query is usually quite satisfying. But every now and then (as I said, when querying 5M+ records, possibly including non-indexed searches), this runs for 2-3minutes.
EDIT: Please note, that a potential bottleneck is not so easy to circumvent, because of sorting that has to be applied before paging!
I'm wondering, is that state-of-the-art simulation of LIMIT ... OFFSET, including TOTAL_ROWS in Oracle, or is there a better solution that will be faster by design, e.g. by using the ROW_NUMBER() window function instead of the ROWNUM pseudo-column?
The main problem with Case 2 is that in many cases the whole query result set has to be obtained and then sorted before the first N rows can be returned - unless the ORDER BY columns are indexed and Oracle can use the index to avoid a sort. For a complex query and a large set of data this can take some time. However there may be some things you can do to improve the speed:
Try to ensure that no functions are called in the inner SQL - these may get called 5 million times just to return the first 20 rows. If you can move these function calls to the outer query they will be called less.
Use a FIRST_ROWS_n hint to nudge Oracle into optimising for the fact that you will never return all the data.
EDIT:
Another thought: you are currently presenting the user with a report that could return thousands or millions of rows, but the user is never realistically going to page through them all. Can you not force them to select a smaller amount of data e.g. by limiting the date range selected to 3 months (or whatever)?
You might want to trace the query that takes a lot of time and look at its explain plan. Most likely the performance bottleneck comes from the TOTAL_ROWS calculation. Oracle has to read all the data, even if you only fetch one row, this is a common problem that all RDBMS face with this type of query. No implementation of TOTAL_ROWS will get around that.
The radical way to speed up this type of query is to forego the TOTAL_ROWS calculation. Just display that there are additional pages. Do your users really need to know that they can page through 52486 pages? An estimation may be sufficient. That's another solution, implemented by google search for example: estimate the number of pages instead of actually counting them.
Designing an accurate and efficient estimation algorithm might not be trivial.
A "LIMIT ... OFFSET" is pretty much syntactic sugar. It might make the query look prettier, but if you still need to read the whole of a data set and sort it and get rows "50-60", then that's the work that has to be done.
If you have an index in the right order, then that can help.
It may perform better to run two queries instead of trying to count() and return the results in the same query. Oracle may be able to answer the count() without any sorting or joining to all the tables (join table elimination based on declared foreign key constraints). This is what we generally do in our application. For performance important statements, we write a separate query that we know will return the correct count as we can sometimes do better than Oracle.
Alternatively, you can make a tradeoff between performance and recency of the data. Bringing back the first 5 pages is going to be nearly as quick as bringing back the first page. So you could consider storing the results from 5 pages in a temporary table along with an expiry date for the information. Take the result from the temporary table if valid. Put a background task in to delete the expired data periodically.
Say tableA has 1 row to be returned but will have 100 columns returned while tableB has 100 rows to be returned but only one column from each. TableB has a foreign key for table A.
Will a left join of tableA to tableB return 100*100 cells of data while 2 separate queries return 100 + 100 cells of data or 50 times less data or is that a misunderstanding of how it works?
Is it ever more efficient to use many simple queries rather than fewer more complex ones?
First and foremost, I would question a table with 100 columns, and suggest that there is a possibly a better design for your schema. In the real world, this number of columns is less common, so typically the difference in the amount of data returned with one query vs. two becomes less significant. 100 columns in a table is not necessarily bad, just a flag that it shold be considered.
However, assuming your numbers are what they are to make clear the question, there are a few important variables to consider:
1 - What is the speed of the link between the db server and the application server? If it is very slow, then you are probably better off minimizing the amount data returned vs. the number of queries you run. If it is not slow, then you will likely expend more time in the execution of two queries than you would returning the increased payload. Which is better can only be determined by testing in your own environment.
2 - How efficient is the transport protocol itself? Perhaps there is some kind of compression of the data, or an even more clever algorithm that knows column 2 through 101 are duplicate for every row, so it only passes them once. Strategies like this in the transport protocol would mitigate any of your concerns. Again, this is why you need to test in your own envionment to know for sure.
As others have pointed out, you also need to consider what will be done with the data once you get it (e.g., JOINs, GROUPing, etc), but I am limiting my response to the specifics of your question around query count vs. payload size.
What is best at joining? A database engine or client code? Saying that, I use both techniques: it depends on the client and how data will be used.
Where the data requires some processing to, say, render on a web page I'd probably split header and details recordsets. We do use this because we have some business logic between DB and HTML
Where it's consumed simply and linearly, I'd join in the database to avoid unnecessary processing. For example, simple reports or exports
It depends, if you only take into account the SQL efficiency obviusly several simpler and smaller result queries will be more efficient.
But you need to take into account the whole process if the join will be made otherwise on the client or you need to filter results after the join, then probably the DBM will be more efficient that doing it on your code.
Coding is always a tradeoff between diferent systems, DB vs Client, RAM vs CPU... you need to be conscious about this and try to find the perfect solution.
In this case probably 2 queries outperform 1 but that is not a general solution.
I think that your question basically is about database normalization. In general, it is advisable to normalize a database into multiple tables (using primary and foreign keys) and to join them as needed upon queries. This is better for insert/update performance and for keeping the data consistent, and usually results in smaller database sizes as well.
As for the row numbers returned, only a cross join would actually return 100*100 rows; any inner or outer join will not create all combinations, but rather tie together rows on the given conditions, and for outer joins preserve rows which could not be matched. Wikipedia has some samples in its JOIN article.
For very query-intense applications, the performance may be better when using less normlized tables. However, as always with optimizations, I'd only consider going into that direction after seeing real measurable problems (e.g. with a profiling tool).
In general, try to keep the number of roundtrips to the database low; a large number of single simple queries will suffer from the overhead of talking to the DB engine (network etc.). If you need to execute complex series of statements, consider using stored procedures.
Generally fewer queries makes for better performance, as long as the queries return data that is actually related. There is no point in trying to put unrelated data into the same query just to reduce the number or queries.
There are of course exceptions, and your example may be one of them. However, it depends on more than the number of fields returnes, like what the fields actually return, i.e. the actual amount of data.
As an example of how the number of queries affects performance, I can mention a solution that I have (sadly enough) seen many times. In that solution the programmer would first get a number of records from one table, then loop through the records and run another query for each record to get the related records from another table. This clearly results in a lot of queries, and a solution having either one or two queries would be much more efficient.
“Is it ever more efficient to use many simple queries rather than fewer more complex ones?”
The query that requires the least amount of data to traverse, and gives you no more than what you need is the more efficient one. Beyond this, there can be RDBMS specific conditions that can be more efficient on one RDBMS system than another. At the very low level, when you deal with less data, then your results can be retrieved much quicker, so efficient queries are queries that only work with the least amount of data needed to get you the result you are looking for.
I have a query that looks something like this:
select xmlelement("rootNode",
(case
when XH.ID is not null then
xmlelement("xhID", XH.ID)
else
xmlelement("xhID", xmlattributes('true' AS "xsi:nil"), XH.ID)
end),
(case
when XH.SER_NUM is not null then
xmlelement("serialNumber", XH.SER_NUM)
else
xmlelement("serialNumber", xmlattributes('true' AS "xsi:nil"), XH.SER_NUM)
end),
/*repeat this pattern for many more columns from the same table...*/
FROM XH
WHERE XH.ID = 'SOMETHINGOROTHER'
It's ugly and I don't like it, and it is also the slowest executing query (there are others of similar form, but much smaller and they aren't causing any major problems - yet). Maintenance is relatively easy as this is mostly a generated query, but my concern now is for performance. I am wondering how much of an overhead there is for all of these case expressions.
To see if there was any difference, I wrote another version of this query as:
select xmlelement("rootNode",
xmlforest(XH.ID, XH.SER_NUM,...
(I know that this query does not produce exactly the same, thing, my plan was to move the logic for handling the renaming and xsi:nil attribute to XSL or maybe to PL/SQL)
I tried to get execution plans for both versions, but they are the same. I'm guessing that the logic does not get factored into the execution plan. My gut tells me the second version should execute faster, but I'd like some way to prove that (other than writing a PL/SQL test function with timing statements before and after the query and running that code over and over again to get a test sample).
Is it possible to get a good idea of how much the case-when will cost?
Also, I could write the case-when using the decode function instead. Would that perform better (than case-statements)?
Just about anything in your SELECT list, unless it is a user-defined function which reads a table or view, or a nested subselect, can usually be neglected for the purpose of analyzing your query's performance.
Open your connection properties and set the value SET STATISTICS IO on. Check out how many reads are happening. View the query plan. Are your indexes being used properly? Do you know how to analyze the plan to see?
For the purposes of performance tuning you are dealing with this statement:
SELECT *
FROM XH
WHERE XH.ID = 'SOMETHINGOROTHER'
How does that query perform? If it returns in markedly less time than the XML version then you need to consider the performance of the functions, but I would astonished if that were the case (oh ho!).
Does this return one row or several? If one row then you have only two things to work with:
is XH.ID indexed and, if so, is the index being used?
does the "many more columns from the same table" indicate a problem with chained rows?
If the query returns several rows then ... Well, actually you have the same two things to work with. It's just the emphasis is different with regards to indexes. If the index has a very poor clustering factor then it could be faster to avoid using the index in favour of a full table scan.
Beyond that you would need to look at physical problems - I/O bottlenecks, poor interconnects, a dodgy disk. The reason why your scope for tuning the query is so restricted is because - as presented - it is a single table, single column read. Most tuning is about efficient joining. Now if XH transpires to be a view over a complex query then it is a different matter.
You can use good old tkprof to analyze statistics. One of the many forms of ALTER SESSION that turn on stats gathering. The DBMS_PROFILER package also gathers statistics if your cursor is in a PL/SQL code block.
What techniques can be applied effectively to improve the performance of SQL queries? Are there any general rules that apply?
Use primary keys
Avoid select *
Be as specific as you can when building your conditional statements
De-normalisation can often be more efficient
Table variables and temporary tables (where available) will often be better than using a large source table
Partitioned views
Employ indices and constraints
Learn what's really going on under the hood - you should be able to understand the following concepts in detail:
Indexes (not just what they are but actually how they work).
Clustered indexes vs heap allocated tables.
Text and binary lookups and when they can be in-lined.
Fill factor.
How records are ghosted for update/delete.
When page splits happen and why.
Statistics, and how they effect various query speeds.
The query planner, and how it works for your specific database (for instance on some systems "select *" is slow, on modern MS-Sql DBs the planner can handle it).
The biggest thing you can do is to look for table scans in sql server query analyzer (make sure you turn on "show execution plan"). Otherwise there are a myriad of articles at MSDN and elsewhere that will give good advice.
As an aside, when I started learning to optimize queries I ran sql server query profiler against a trace, looked at the generated SQL, and tried to figure out why that was an improvement. Query profiler is far from optimal, but it's a decent start.
There are a couple of things you can look at to optimize your query performance.
Ensure that you just have the minimum of data. Make sure you select only the columns you need. Reduce field sizes to a minimum.
Consider de-normalising your database to reduce joins
Avoid loops (i.e. fetch cursors), stick to set operations.
Implement the query as a stored procedure as this is pre-compiled and will execute faster.
Make sure that you have the correct indexes set up. If your database is used mostly for searching then consider more indexes.
Use the execution plan to see how the processing is done. What you want to avoid is a table scan as this is costly.
Make sure that the Auto Statistics is set to on. SQL needs this to help decide the optimal execution. See Mike Gunderloy's great post for more info. Basics of Statistics in SQL Server 2005
Make sure your indexes are not fragmented. Reducing SQL Server Index Fragmentation
Make sure your tables are not fragmented. How to Detect Table Fragmentation in SQL Server 2000 and 2005
Use a with statment to handle query filtering.
Limit each subquery to the minimum number of rows possible.
then join the subqueries.
WITH
master AS
(
SELECT SSN, FIRST_NAME, LAST_NAME
FROM MASTER_SSN
WHERE STATE = 'PA' AND
GENDER = 'M'
),
taxReturns AS
(
SELECT SSN, RETURN_ID, GROSS_PAY
FROM MASTER_RETURNS
WHERE YEAR < 2003 AND
YEAR > 2000
)
SELECT *
FROM master,
taxReturns
WHERE master.ssn = taxReturns.ssn
A subqueries within a with statement may end up as being the same as inline views,
or automatically generated temp tables. I find in the work I do, retail data, that about 70-80% of the time, there is a performance benefit.
100% of the time, there is a maintenance benefit.
I think using SQL query analyzer would be a good start.
In Oracle you can look at the explain plan to compare variations on your query
Make sure that you have the right indexes on the table. if you frequently use a column as a way to order or limit your dataset an index can make a big difference. I saw in a recent article that select distinct can really slow down a query, especially if you have no index.
The obvious optimization for SELECT queries is ensuring you have indexes on columns used for joins or in WHERE clauses.
Since adding indexes can slow down data writes you do need to monitor performance to ensure you don't kill the DB's write performance, but that's where using a good query analysis tool can help you balanace things accordingly.
Indexes
Statistics
on microsoft stack, Database Engine Tuning Advisor
Some other points (Mine are based on SQL server, since each db backend has it's own implementations they may or may not hold true for all databases):
Avoid correlated subqueries in the select part of a statement, they are essentially cursors.
Design your tables to use the correct datatypes to avoid having to apply functions on them to get the data out. It is far harder to do date math when you store your data as varchar for instance.
If you find that you are frequently doing joins that have functions in them, then you need to think about redesigning your tables.
If your WHERE or JOIN conditions include OR statements (which are slower) you may get better speed using a UNION statement.
UNION ALL is faster than UNION if (And only if) the two statments are mutually exclusive and return the same results either way.
NOT EXISTS is usually faster than NOT IN or using a left join with a WHERE clause of ID = null
In an UPDATE query add a WHERE condition to make sure you are not updating values that are already equal. The difference between updating 10,000,000 records and 4 can be quite significant!
Consider pre-calculating some values if you will be querying them frequently or for large reports. A sum of the values in an order only needs to be done when the order is made or adjusted, rather than when you are summarizing the results of 10,000,000 million orders in a report. Pre-calculations should be done in triggers so that they are always up-to-date is the underlying data changes. And it doesn't have to be just numbers either, we havea calculated field that concatenates names that we use in reports.
Be wary of scalar UDFs, they can be slower than putting the code in line.
Temp table tend to be faster for large data set and table variables faster for small ones. In addition you can index temp tables.
Formatting is usually faster in the user interface than in SQL.
Do not return more data than you actually need.
This one seems obvious but you would not believe how often I end up fixing this. Do not join to tables that you are not using to filter the records or actually calling one of the fields in the select part of the statement. Unnecessary joins can be very expensive.
It is an very bad idea to create views that call other views that call other views. You may find you are joining to the same table 6 times when you only need to once and creating 100,000,00 records in an underlying view in order to get the 6 that are in your final result.
In designing a database, think about reporting not just the user interface to enter data. Data is useless if it is not used, so think about how it will be used after it is in the database and how that data will be maintained or audited. That will often change the design. (This is one reason why it is a poor idea to let an ORM design your tables, it is only thinking about one use case for the data.) The most complex queries affecting the most data are in reporting, so designing changes to help reporting can speed up queries (and simplify them) considerably.
Database-specific implementations of features can be faster than using standard SQL (That's one of the ways they sell their product), so get to know your database features and find out which are faster.
And because it can't be said too often, use indexes correctly, not too many or too few. And make your WHERE clauses sargable (Able to use indexes).