大表的UNION性能问题

Question

I have seven large tables, that can be storing between 100 to 1 million rows at any time. I'll call them LargeTable1 , LargeTable2 , LargeTable3 , LargeTable4 ... LargeTable7 . These tables are mostly static: there are no updates nor new inserts. They change only once every two weeks or once a month, when they are truncated and a new batch of registers are inserted in each.

All these tables have three fields in common: Headquarter , Country and File . Headquarter and Country are numbers in the format '000', though in two of these tables they are parsed as int due to some other system necessities.

I have another, much smaller table called Headquarters with the information of each headquarter. This table has very few entries. At most 1000, actually.

Now, I need to create a stored procedure that returns all those headquarters that appear in the large tables but are either absent in the Headquarters table or have been deleted (this table is deleted logically: it has a DeletionDate field to check this).

This is the query I've tried:

CREATE PROCEDURE deletedHeadquarters
AS
BEGIN
    DECLARE @headquartersFiles TABLE
    (
        hq int,
        countryFile varchar(MAX)
    );

    SET NOCOUNT ON

    INSERT INTO @headquartersFiles
    SELECT headquarter, CONCAT(country, ' (', file, ')')
    FROM
    (
        SELECT DISTINCT CONVERT(int, headquarter) as headquarter,
                        CONVERT(int, country) as country,
                        file
        FROM            LargeTable1     
        UNION
        SELECT DISTINCT headquarter,
                        country,
                        file
        FROM            LargeTable2
        UNION
        SELECT DISTINCT headquarter,
                        country,
                        file
        FROM            LargeTable3
        UNION
        SELECT DISTINCT headquarter,
                        country,
                        file
        FROM            LargeTable4
        UNION
        SELECT DISTINCT headquarter,
                        country,
                        file
        FROM            LargeTable5
        UNION
        SELECT DISTINCT headquarter,
                        country,
                        file
        FROM            LargeTable6
        UNION
        SELECT DISTINCT headquarter,
                        country,
                        file
        FROM            LargeTable7
    ) TC

    SELECT  RIGHT('000' + CAST(st.headquarter AS VARCHAR(3)), 3) as headquarter,
            MAX(s.deletionDate) as deletionDate,
            STUFF
            (
                (SELECT DISTINCT ', ' + st2.countryFile
                FROM @headquartersFiles st2
                WHERE st2.headquarter = st.headquarter
                FOR XML PATH('')),
                1,
                1,
                ''
            ) countryFile
    FROM    @headquartersFiles as st
    LEFT JOIN headquarters s ON CONVERT(int, s.headquarter) = st.headquarter
    WHERE   s.headquarter IS NULL
       OR   s.deletionDate IS NOT NULL
    GROUP BY st.headquarter

END

This sp's performance isn't good enough for our application. It currently takes around 50 seconds to complete, with the following total rows for each table (just to give you an idea about the sizes):

• LargeTable1: 1516666 rows
• LargeTable2: 645740 rows
• LargeTable3: 1950121 rows
• LargeTable4: 779336 rows
• LargeTable5: 1100999 rows
• LargeTable6: 16499 rows
• LargeTable7: 24454 rows

What can I do to improve performance? I've tried to do the following, with no much difference:

• Inserting into the local table by batches , excluding those headquarters I've already inserted and then updating the countryFile field for those that are repeated
• Creating a view for that UNION query
• Creating indexes for the LargeTables for the headquarter field

I've also thought about inserting these missing headquarters in a permanent table after the LargeTables change, but the Headquarters table can change more often, and I would like not having to change its module to keep these things tidy and updated. But if it's the best possible alternative, I'd go for it.

Thanks

Answer 1

Take this filter

LEFT JOIN headquarters s ON CONVERT(int, s.headquarter) = st.headquarter
WHERE   s.headquarter IS NULL
   OR   s.deletionDate IS NOT NULL

And add it to each individual query in the union and insert into @headquartersFiles

It might seem like this makes a lot more filters but it will actually speed stuff up because you are filtering before you start processing as a union.

Also take out all your DISTINCT, it probably won't speed it up but it seems silly because you are doing a UNION and not a UNION all.

Answer 2

I'd try doing the filtering with each individual table first. You just need to account for the fact that a headquarter might appear in one table, but not another. You can do this like so:

SELECT
    headquarter
FROM
(

    SELECT DISTINCT
        headquarter,
        'table1' AS large_table
    FROM
        LargeTable1 LT
    LEFT OUTER JOIN Headquarters HQ ON HQ.headquarter = LT.headquarter
    WHERE
        HQ.headquarter IS NULL OR
        HQ.deletion_date IS NOT NULL
    UNION ALL
    SELECT DISTINCT
        headquarter,
        'table2' AS large_table
    FROM
        LargeTable2 LT
    LEFT OUTER JOIN Headquarters HQ ON HQ.headquarter = LT.headquarter
    WHERE
        HQ.headquarter IS NULL OR
        HQ.deletion_date IS NOT NULL
    UNION ALL
    ...
) SQ
GROUP BY headquarter
HAVING COUNT(*) = 5

That would make sure that it's missing from all five tables.

Answer 3

Table variables have horrible performance because sql server does not generate statistics for them. Instead of a table variable, try using a temp table instead, and if headquarter + country + file is unique in the temp table, add a unique constraint (which will create a clustered index) in the temp table definition. You can set indexes on a temp table after creating it, but for various reasons SQL Server may ignore it.

Edit: as it turns out, you can in fact create indexes on table variables, even non-unique in 2014+.

Secondly, try not to use functions in your joins or where clauses - doing so often causes performance problems.

Answer 4

Do the filtering at each step. But first, modify the headquarters table so it has the right type for what you need . . . along with an index:

alter table headquarters add headquarter_int as (cast(headquarter as int));
create index idx_headquarters_int on headquarters(headquarters_int);

SELECT DISTINCT headquarter, country, file
FROM LargeTable5 lt5
WHERE NOT EXISTS (SELECT 1
                  FROM headquarters s
                  WHERE s.headquarter_int = lt5.headquarter and s.deletiondate is not null
                 );

Then, you want an index on LargeTable5(headquarter, country, file) .

This should take less than 5 seconds to run. If so, then construct the full query, being sure that the types in the correlated subquery match and that you have the right index on the full table. Use union to remove duplicates between the tables.

Answer 5

真正的答案是为每个表创建单独的INSERT语句，但要注意的是，要插入的数据在目标表中不存在。