SQL atomic increment and locking strategies - is this safe?

Solution 1:

UPDATE query places an update lock on the pages or records it reads.

When a decision is made whether to update the record, the lock is either lifted or promoted to the exclusive lock.

This means that in this scenario:

s1: read counter for image_id=15, get 0, store in temp1
s2: read counter for image_id=15, get 0, store in temp2
s1: write counter for image_id=15 to (temp1+1), which is 1 
s2: write counter for image_id=15 to (temp2+1), which is also 1

s2 will wait until s1 decides whether to write the counter or not, and this scenario is in fact impossible.

It will be this:

s1: place an update lock on image_id = 15
s2: try to place an update lock on image_id = 15: QUEUED
s1: read counter for image_id=15, get 0, store in temp1
s1: promote the update lock to the exclusive lock
s1: write counter for image_id=15 to (temp1+1), which is 1 
s1: commit: LOCK RELEASED
s2: place an update lock on image_id = 15
s2: read counter for image_id=15, get 1, store in temp2
s2: write counter for image_id=15 to (temp2+1), which is 2

Note that in InnoDB, DML queries do not lift the update locks from the records they read.

This means that in case of a full table scan, the records that were read but decided not to update, will still remain locked until the end of the transaction and cannot be updated from another transaction.

Solution 2:

If the locking is not done properly it certainly is possible to get this type race condition, and the default locking mode (read committed) does allow it. In this mode, the reads only place a shared lock on the record, so they can both see 0, increment it and write 1 out to the database.

In order to avoid this race condition, you need to set an exclusive lock on the read operation. 'Serializable' and 'Repeatable Read' concurrency modes will do this, and for an operation on a single row they are pretty much equivalent.

To make it completely atomic you have to:

• Set an appropriate transaction isolation level such as Serializable. Normally you can do this from a client library or explicilty in SQL.
• Begin the transaction
• Read the data
• Update it
• Commit the transaction.

You can also force an exclusive lock on the read with a HOLDLOCK (T-SQL) or equivalent hint, depending on your SQL dialect.

A single update query will do this atomically but you can't split the operation (perhaps to read the value and return it to the client) without ensuring that the reads take out an exclusive lock. You will need to get the value out atomically in order to implement a sequence, so the update by itself is probably not quite all you need. Even with the atomic update, you still have a race condition to read the value after the update. The read will still have to take place within a transaction (storing what it got in a variable) and issue an exclusive lock during the read.

Note that to do this without creating a hot spot your database needs to have proper support for autonomous (nested) transactions within a stored procedure. Note that sometimes 'nested' is used to refer to chaining transactions or save points, so the term can be a bit confusing. I've edited this to refer to autonomous transactions.

Without autonomous transactions your locks are inherited by the parent transaction, which can roll back the whole lot. This means they will be held until the parent transaction commits, which can turn your sequence into a hot spot that serialises all transactions using that sequence. Anything else trying to use the sequence will block until the whole parent transaction commits.

IIRC Oracle supports autonomous transactions but DB/2 didn't until fairly recently and SQL Server doesn't. Off the top of my head I don't know whether InnoDB supports them, but Grey and Reuter go on at some length about how difficult they are to implement. In practice I'd guess it's quite likely that it might not. YMMV.