2

I have table that suffers from racing condition and I don't know how to fix it. I can't solve it 100% in code, so I need the server to help me.

The scenario

Admins of a system may accept or reject an appointment. This can be done by email (via token etc) or by logging in the system and doing it via admin panel.

The problem appears when two admin do it at the same time:

t = 0:
admin A clicks accept from the system email;
admin B clicks reject from the admin panel;

t = 1:
back-end checks was appointment already handled?
This happens at the same time for both admins, so they both pass the test -> appointment is waiting for approval...

t = 2:
back-end A sends query "update status to accepted"
back-end B sends query "update status to rejected"

t = 3: both sqls hit the server.
The server than changes the appointment to accepted, then proceeds to execute the second query, and changes the appointment to rejected.

This table has heavy traffic, so locking it is not an option.

What I think could solve my problem, is to:
first: have a query lock the row;
second: check the the status column to see if it is already set;
third: if not set, then set the value;
fourth: release the lock;


meanwhile, the second query would keep trying to acquire the lock. After it succeeded, it would then proceed to check whether the value was already set, and then it would find that indeed it was, and would return an error or something.

How can I go about doing something like this on Sql Server?

The flow would look something like this:

while (can't read) {

    try to acquire lock;
}

// lock acquired.
read appointment from table;
if value is null, set the value;
release lock;

I figured out my framework is using READ_COMMITTED_SNAPSHOT as default.

0

2 Answers 2

4

We had a similar requirement and so added an extra where clause to the query. E.g.

UPDATE AppointmentTable 
SET STatus = [Approved/Rejected] 
WHERE AppointmentID = @ID AND Status = Pending

Then afterwards, check how many rows were updated. If there were 1 rows updated then report back to the user that their action was recorded. If 0 rows were updated then report back to the user that someone else has already done the work.

Isn't there a chance the queries will be executed "at the same time"

Nope. If they execute "at the same time" one of them will still get the update lock first. The other will have to wait for the update lock. When the 2nd gets the lock it still has to check the where clause to validate it's still true. If it's no longer true then it skips that row and no update happens (@@rowcount = 0). The key is to have the old value in the where clause.

0
0

As @Kin indicated, the solution hinges on what transaction isolation level you're using. Your DBA should be able to answer that question. If you're using plain old out-of-the-box READ_COMMITTED (the default) and no SNAPSHOT isolation, you should be able to:

  • BEGIN TRANSACTION
  • UPDATE AppointmentTable SET AppointmentStatus = AppointmentStatus WHERE ID=@ID (This will create an update lock on the row, but you haven't altered anything yet because you're updating to the same value that's already there. If someone else is updating this row, your transaction will wait for their update to complete)
  • SELECT AppointmentStatus FROM AppointmentTable (If someone else is updating this row, your transaction will wait for their update to complete.
  • If the AppointmentStatus IN ('Accept','Reject'), you can rollback your work, even though nothing was really updated so far.
  • If the AppointmentStatus NOT IN ('Accept','Reject'), UPDATE AppointmentTable SET AppointmentStatus= 'either Accept or Reject'
  • Commit transaction
1
  • Thanks for your input, it seems simple enough. I figured out my framework is using READ_COMMITTED_SNAPSHOT as default.
    – victor
    Commented Sep 6, 2016 at 20:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.