DELETE SQL with correlated subquery for table with 42 million rows?

问题

I have a table cats with 42,795,120 rows.

Apparently this is a lot of rows. So when I do:

/* owner_cats is a many-to-many join table */
DELETE FROM cats
WHERE cats.id_cat IN (
SELECT owner_cats.id_cat FROM owner_cats
WHERE owner_cats.id_owner = 1)

the query times out :(

(edit: I need to increase my CommandTimeout value, default is only 30 seconds)

I can't use TRUNCATE TABLE cats because I don't want to blow away cats from other owners.

I'm using SQL Server 2005 with "Recovery model" set to "Simple."

So, I thought about doing something like this (executing this SQL from an application btw):

DELETE TOP (25) PERCENT FROM cats
WHERE cats.id_cat IN (
SELECT owner_cats.id_cat FROM owner_cats
WHERE owner_cats.id_owner = 1)

DELETE TOP(50) PERCENT FROM cats
WHERE cats.id_cat IN (
SELECT owner_cats.id_cat FROM owner_cats
WHERE owner_cats.id_owner = 1)

DELETE FROM cats
WHERE cats.id_cat IN (
SELECT owner_cats.id_cat FROM owner_cats
WHERE owner_cats.id_owner = 1)

My question is: what is the threshold of the number of rows I can DELETE in SQL Server 2005?

Or, if my approach is not optimal, please suggest a better approach. Thanks.

This post didn't help me enough:

SQL Server Efficiently dropping a group of rows with millions and millions of rows

EDIT (8/6/2010):

Okay, I just realized after reading the above link again that I did not have indexes on these tables. Also, some of you have already pointed out that issue in the comments below. Keep in mind this is a fictitious schema, so even id_cat is not a PK, because in my real life schema, it's not a unique field.

I will put indexes on:

cats.id_cat
owner_cats.id_cat
owner_cats.id_owner

I guess I'm still getting the hang of this data warehousing, and obviously I need indexes on all the JOIN fields right?

However, it takes hours for me to do this batch load process. I'm already doing it as a SqlBulkCopy (in chunks, not 42 mil all at once). I have some indexes and PKs. I read the following posts which confirms my theory that the indexes are slowing down even a bulk copy:

SqlBulkCopy slow as molasses
What’s the fastest way to bulk insert a lot of data in SQL Server (C# client)

So I'm going to DROP my indexes before the copy and then re CREATE them when it's done.

Because of the long load times, it's going to take me awhile to test these suggestions. I'll report back with the results.

UPDATE (8/7/2010):

Tom suggested:

DELETE
FROM cats c
WHERE EXISTS (SELECT 1
FROM owner_cats o
WHERE o.id_cat = c.id_cat
AND o.id_owner = 1)

And still with no indexes, for 42 million rows, it took 13:21 min:sec versus 22:08 with the way described above. However, for 13 million rows, took him 2:13 versus 2:10 my old way. It's a neat idea, but I still need to use indexes!

Update (8/8/2010):

Something is terribly wrong! Now with the indexes on, my first delete query above took 1:9 hrs:min (yes an hour!) versus 22:08 min:sec and 13:21 min:sec versus 2:10 min:sec for 42 mil rows and 13 mil rows respectively. I'm going to try Tom's query with the indexes now, but this is heading in the wrong direction. Please help.

Update (8/9/2010):

Tom's delete took 1:06 hrs:min for 42 mil rows and 10:50 min:sec for 13 mil rows with indexes versus 13:21 min:sec and 2:13 min:sec respectively. Deletes are taking longer on my database when I use indexes by an order of magnitude! I think I know why, my database .mdf and .ldf grew from 3.5 GB to 40.6 GB during the first (42 mil) delete! What am I doing wrong?

Update (8/10/2010):

For lack of any other options, I have come up with what I feel is a lackluster solution (hopefully temporary):

Increase timeout for database connection to 1 hour (CommandTimeout=60000; default was 30 sec)
Use Tom's query: DELETE FROM WHERE EXISTS (SELECT 1 ...) because it performed a little faster
DROP all indexes and PKs before running delete statement (???)
Run DELETE statement
CREATE all indexes and PKs

Seems crazy, but at least it's faster than using TRUNCATE and starting over my load from the beginning with the first owner_id, because one of my owner_id takes 2:30 hrs:min to load versus 17:22 min:sec for the delete process I just described with 42 mil rows. (Note: if my load process throws an exception, I start over for that owner_id, but I don't want to blow away previous owner_id, so I don't want to TRUNCATE the owner_cats table, which is why I'm trying to use DELETE.)

Anymore help would still be appreciated :)

回答1:

There is no practical threshold. It depends on what your command timeout is set to on your connection.

Keep in mind that the time it takes to delete all of these rows is contingent upon:

The time it takes to find the rows of interest
The time it takes to log the transaction in the transaction log
The time it takes to delete the index entries of interest
The time it takes to delete the actual rows of interest
The time it takes to wait for other processes to stop using the table so you can acquire what in this case will most likely be an exclusive table lock

The last point may often be the most significant. Do an sp_who2 command in another query window to make sure that there isn't lock contention going on, preventing your command from executing.

Improperly configured SQL Servers will do poorly at this type of query. Transaction logs which are too small and/or share the same disks as the data files will often incur severe performance penalties when working with large rows.

As for a solution, well, like all things, it depends. Is this something you intend to be doing often? Depending on how many rows you have left, the fastest way might be to rebuild the table as another name and then rename it and recreate its constraints, all inside a transaction. If this is just an ad-hoc thing, make sure your ADO CommandTimeout is set high enough and you can just bear the cost of this big delete.

回答2:

If the delete will remove "a significant number" of rows from the table, this can be an alternative to a DELETE: put the records to keep somewhere else, truncate the original table, put back the 'keepers'. Something like:

SELECT *
INTO #cats_to_keep
FROM cats
WHERE cats.id_cat NOT IN (    -- note the NOT
SELECT owner_cats.id_cat FROM owner_cats
WHERE owner_cats.id_owner = 1)

TRUNCATE TABLE cats

INSERT INTO cats
SELECT * FROM #cats_to_keep

回答3:

Have you tried no Subquery and use a join instead?

DELETE cats 
FROM
 cats c
 INNER JOIN owner_cats oc
 on c.id_cat = oc.id_cat
WHERE
   id_owner =1

And if you have have you also tried different Join hints e.g.

DELETE cats 
FROM
 cats c
 INNER HASH JOIN owner_cats oc
 on c.id_cat = oc.id_cat
WHERE
   id_owner =1

回答4:

If you use an EXISTS rather than an IN, you should get much better performance. Try this:

DELETE
  FROM cats c
 WHERE EXISTS (SELECT 1
                 FROM owner_cats o
                WHERE o.id_cat = c.id_cat
                  AND o.id_owner = 1)

回答5:

There's no threshold as such - you can DELETE all the rows from any table given enough transaction log space - which is where your query is most likely falling over. If you're getting some results from your DELETE TOP (n) PERCENT FROM cats WHERE ... then you can wrap it in a loop as below:

SELECT 1
WHILE @@ROWCOUNT <> 0
BEGIN
 DELETE TOP (somevalue) PERCENT FROM cats
 WHERE cats.id_cat IN (
 SELECT owner_cats.id_cat FROM owner_cats
 WHERE owner_cats.id_owner = 1)
END

回答6:

As others have mentioned, when you delete 42 million rows, the db has to log 42 million deletions against the database. Thus, the transaction log has to grow substantially. What you might try is to break up the delete into chunks. In the following query, I use the NTile ranking function to break up the rows into 100 buckets. If that is too slow, you can expand the number of buckets so that each delete is smaller. It will help tremendously if there is an index on owner_cats.id_owner, owner_cats.id_cats and cats.id_cat (which I assumed the primary key and numeric).

Declare @Cats Cursor
Declare @CatId int  --assuming an integer PK here
Declare @Start int
Declare @End int
Declare @GroupCount int

Set @GroupCount = 100

Set @Cats = Cursor Fast_Forward For
    With CatHerd As
        (
        Select cats.id_cat
            , NTile(@GroupCount) Over ( Order By cats.id_cat ) As Grp
        From cats
            Join owner_cats
                On owner_cats.id_cat = cats.id_cat
        Where owner_cats.id_owner = 1
        )
        Select Grp, Min(id_cat) As MinCat, Max(id_cat) As MaxCat
        From CatHerd
        Group By Grp
Open @Cats
Fetch Next From @Cats Into @CatId, @Start, @End

While @@Fetch_Status = 0
Begin
    Delete cats
    Where id_cat Between @Start And @End

    Fetch Next From @Cats Into @CatId, @Start, @End
End 

Close @Cats
Deallocate @Cats

The notable catch with the above approach is that it is not transactional. Thus, if it fails on the 40th chunk, you will have deleted 40% of the rows and the other 60% will still exist.

回答7:

Might be worth trying MERGE e.g.

MERGE INTO cats 
   USING owner_cats
      ON cats.id_cat = owner_cats.id_cat
         AND owner_cats.id_owner = 1
WHEN MATCHED THEN DELETE;

回答8:

<Edit> (9/28/2011)
My answer performs basically the same way as Thomas' solution (Aug 6 '10). I missed it when I posted my answer because it he uses an actual CURSOR so I thought to myself "bad" because of the # of records involved. However, when I reread his answer just now I realize that the WAY he uses the cursor is actually "good". Very clever. I just voted up his answer and will probably use his approach in the future. If you don't understand why, take a look at it again. If you still can't see it, post a comment on this answer and I will come back and try to explain in detail. I decided to leave my answer because someone may have a DBA who refuses to let them use an actual CURSOR regardless of how "good" it is. :-)
</Edit>

I realize that this question is a year old but I recently had a similar situation. I was trying to do "bulk" updates to a large table with a join to a different table, also fairly large. The problem was that the join was resulting in so many "joined records" that it took too long to process and could have led to contention problems. Since this was a one-time update I came up with the following "hack." I created a WHILE LOOP that went through the table to be updated and picked 50,000 records to update at a time. It looked something like this:

DECLARE @RecId bigint
DECLARE @NumRecs bigint
SET @NumRecs = (SELECT MAX(Id) FROM [TableToUpdate])
SET @RecId = 1
WHILE @RecId < @NumRecs
BEGIN
    UPDATE [TableToUpdate]
    SET UpdatedOn = GETDATE(),
        SomeColumn = t2.[ColumnInTable2]
    FROM    [TableToUpdate] t
    INNER JOIN [Table2] t2 ON t2.Name = t.DBAName 
        AND ISNULL(t.PhoneNumber,'') = t2.PhoneNumber 
        AND ISNULL(t.FaxNumber, '') = t2.FaxNumber
    LEFT JOIN [Address] d ON d.AddressId = t.DbaAddressId 
        AND ISNULL(d.Address1,'') = t2.DBAAddress1
        AND ISNULL(d.[State],'') = t2.DBAState
        AND ISNULL(d.PostalCode,'') = t2.DBAPostalCode
    WHERE t.Id BETWEEN @RecId AND (@RecId + 49999)
    SET @RecId = @RecId + 50000
END

Nothing fancy but it got the job done. Because it was only processing 50,000 records at a time, any locks that got created were short lived. Also, the optimizer realized that it did not have to do the entire table so it did a better job of picking an execution plan.

<Edit> (9/28/2011)
There is a HUGE caveat to the suggestion that has been mentioned here more than once and is posted all over the place around the web regarding copying the "good" records to a different table, doing a TRUNCATE (or DROP and reCREATE, or DROP and rename) and then repopulating the table.

You cannot do this if the table is the PK table in a PK-FK relationship (or other CONSTRAINT). Granted, you could DROP the relationship, do the clean up, and re-establish the relationship, but you would have to clean up the FK table, too. You can do that BEFORE re-establishing the relationship, which means more "down-time", or you can choose to not ENFORCE the CONSTRAINT on creation and clean up afterwards. I guess you could also clean up the FK table BEFORE you clean up the PK table. Bottom line is that you have to explicitly clean up the FK table, one way or the other.

My answer is a hybrid SET-based/quasi-CURSOR process. Another benefit of this method is that if the PK-FK relationship is setup to CASCADE DELETES you don't have to do the clean up I mention above because the server will take care of it for you. If your company/DBA discourage cascading deletes, you can ask that it be enabled only while this process is running and then disabled when it is finished. Depending on the permission levels of the account that runs the clean up, the ALTER statements to enable/disable cascading deletes can be tacked onto the beginning and the end of the SQL statement. </Edit>

回答9:

Bill Karwin's answer to another question applies to my situation also:

"If your DELETE is intended to eliminate a great majority of the rows in that table, one thing that people often do is copy just the rows you want to keep to a duplicate table, and then use DROP TABLE or TRUNCATE to wipe out the original table much more quickly."

Matt in this answer says it this way:

"If offline and deleting a large %, may make sense to just build a new table with data to keep, drop the old table, and rename."

ammoQ in this answer (from the same question) recommends (paraphrased):