If I stop a long running query, does it rollback?

Question

A query that is used to loop through 17 millions records to remove duplicates has been running now for about 16 hours and I wanted to know if the query is stopped right now if it will finalize the delete statements or if it has been deleting while running this query? Indeed, if I do stop it, does it finalize the deletes or rolls back?

I have found that when I do a

 select count(*) from myTable

That the rows that it returns (while doing this query) is about 5 less than what the starting row count was. Obviously the server resources are extremely poor, so does that mean that this process has taken 16 hours to find 5 duplicates (when there are actually thousands), and this could be running for days?

This query took 6 seconds on 2000 rows of test data, and it works great on that set of data, so I figured it would take 15 hours for the complete set.

Any ideas?

Below is the query:

--Declare the looping variable
DECLARE @LoopVar char(10)


    DECLARE
     --Set private variables that will be used throughout
      @long DECIMAL,
      @lat DECIMAL,
      @phoneNumber char(10),
      @businessname varchar(64),
      @winner char(10)

    SET @LoopVar = (SELECT MIN(RecordID) FROM MyTable)

    WHILE @LoopVar is not null
    BEGIN

      --initialize the private variables (essentially this is a .ctor)
      SELECT 
        @long = null,
        @lat = null,
        @businessname = null,
        @phoneNumber = null,
        @winner = null

      -- load data from the row declared when setting @LoopVar  
      SELECT
        @long = longitude,
        @lat = latitude,
        @businessname = BusinessName,
        @phoneNumber = Phone
      FROM MyTable
      WHERE RecordID = @LoopVar

      --find the winning row with that data. The winning row means 
      SELECT top 1 @Winner = RecordID
      FROM MyTable
      WHERE @long = longitude
        AND @lat = latitude
        AND @businessname = BusinessName
        AND @phoneNumber = Phone
      ORDER BY
        CASE WHEN webAddress is not null THEN 1 ELSE 2 END,
        CASE WHEN caption1 is not null THEN 1 ELSE 2 END,
        CASE WHEN caption2 is not null THEN 1 ELSE 2 END,
        RecordID

      --delete any losers.
      DELETE FROM MyTable
      WHERE @long = longitude
        AND @lat = latitude
        AND @businessname = BusinessName
        AND @phoneNumber = Phone
        AND @winner != RecordID

      -- prep the next loop value to go ahead and perform the next duplicate query.
      SET @LoopVar = (SELECT MIN(RecordID) 
    FROM MyTable
    WHERE @LoopVar < RecordID)
    END

Solution 1

no, sql server will not roll back the deletes it has already performed if you stop query execution. oracle requires an explicit committal of action queries or the data gets rolled back, but not mssql.

with sql server it will not roll back unless you are specifically running in the context of a transaction and you rollback that transaction, or the connection closes without the transaction having been committed. but i don't see a transaction context in your above query.

you could also try re-structuring your query to make the deletes a little more efficient, but essentially if the specs of your box are not up to snuff then you might be stuck waiting it out.

going forward, you should create a unique index on the table to keep yourself from having to go through this again.

Solution 2

Your query is not wrapped in a transaction, so it won't rollback the changes already made by the individual delete statements.

I specifically tested this myself on my own SQL Server using the following query, and the ApplicationLog table was empty even though I cancelled the query:

declare @count int
select @count = 5
WHILE @count > 0
BEGIN
  print @count
  delete from applicationlog;
  waitfor time '20:00';
  select @count = @count -1
END

However your query is likely to take many days or weeks, much longer then 15 hours. Your estimate that you can process 2000 records every 6 seconds is wrong because each iteration in your while loop will take significantly longer with 17 million rows then it does with 2000 rows. So unless your query takes significantly less then a second for 2000 rows, it will take days for all 17 million.

You should ask a new question on how you can delete duplicate rows efficiently.

Solution 3

If you don't do anything explicit about transactions then the connection will be in autocommit transactions mode. In this mode every SQL statement is considered a transaction.

The question is whether this means the individual SQL statements are transactions and are therefore being committed as you go, or whether the outer WHILE loop counts as a transaction.

There doesn't seem to be any discussion of this in the description of the WHILE construct on MSDN. However, since a WHILE statement can't directly modify the database it would seem logical that it doesn't start an auto-commit transaction.

Solution 4

Implicit transactions

If no 'Implicit transactions' has been set, then each iteration in your loop committed the changes.

It is possible for any SQL Server to be set with 'Implicit transactions'. This is a database setting (by default is OFF). You can also have implicit transactions in the properties of a particular query inside of Management Studio (right click in query pane>options), by default settings in the client, or a SET statement.

SET IMPLICIT_TRANSACTIONS ON;

Either way, if this was the case, you would still need to execute an explicit COMMIT/ROLLBACK regardless of interruption of the query execution.

---

Implicit transactions reference:

http://msdn.microsoft.com/en-us/library/ms188317.aspx

http://msdn.microsoft.com/en-us/library/ms190230.aspx

Solution 5

I inherited a system which had logic something like yours implemented in SQL. In our case, we were trying to link together rows using fuzzy matching that had similar names/addresses, etc, and that logic was done purely in SQL. At the time I inherited it we had about 300,000 rows in the table and according to the timings, we calculated it would take A YEAR to match them all.

As an experiment to see how much faster I could do it outside of SQL, I wrote a program to dump the db table into flat files, read the flat files into a C++ program, build my own indexes, and do the fuzzy logic there, then reimport the flat files into the database. What took A YEAR in SQL took about 30 seconds in the C++ app.

So, my advice is, don't even try what you are doing in SQL. Export, process, re-import.

View more solutions