Locks and condition variables go together. synchronized (when used properly) makes sure that a group of reads and writes isn’t interleaved with the actions of another thread. notifyAll and wait let threads wait for a condition to become true.

retry inside atomic fills the same role as wait inside synchronized, but retry is safer and more powerful. With transactional waiting the STM automatically identifies the Ref changes that should lead to wakeup, so you don’t have to call notifyAll. This means that there is no chance of lost wakeup. It also means that you can wait for any condition to become true, not just those that were identified ahead of time.

To explain how retry works (and why it’s called retry), it is helpful to think of optimistic concurrency control as a search with backtracking. As an atomic block executes it is searching forward. At each step, if the STM discovers that the current search can’t proceed then it backs up (rolls back) and tries again. In this analogy the transaction is attempting to find a path of Ref reads that avoid the Ref writes performed by other threads.

To make this more concrete, consider the following two atomic blocks that each access a pair of Ref-s. The sum method returns their sum, and the transfer method increments one and decrements the other. transfer should not affect the value returned from sum.

val (x, y) = (Ref(10), Ref(0))

def sum = atomic { implicit txn => 
  val a = x()
  val b = y()
  a + b

def transfer(n: Int) {
  atomic { implicit txn =>
    x -= n
    y += n

If sum and transfer(2) are run simultaneously and sum is unlucky, it might look like this

// sum                                // transfer(2)
atomic                                atomic
|  begin txn attempt                  |  begin txn attempt
|  |  read x -> 10                    |  |  read x -> 10
|  |     :                            |  |  write x <- 8
|  |                                  |  |  read y -> 0
|  |     :                            |  |  write y <- 2
|  |                                  |  commit
|  |  read y -> x read is invalid     +-> ()
|  roll back
|  begin txn attempt
|  |  read x -> 8
|  |  read y -> 2
|  commit
+-> 10

The first time that sum tries to read y, the STM detects that the value previously read from x is no longer correct. This means that the transaction attempt (search) has reached a dead end and the system should roll back (backtrack) and try again. On the second attempt to compute the sum both of the reads are consistent and so the transaction can commit (the search is successful).


When an atomic block calls retry it is signalling that the current search is a dead end, even though all of the reads and writes are consistent. The STM will backtrack and then try again. If some of the Ref-s read by the transaction have changed before the new attempt, then the atomic block may take a different path that avoids the call to retry. The condition that retry waits for is implicitly embedded in the control path of the atomic block.

For example, to wait until x is greater than 10, just write if (x() <= 10) retry. The code that updates x doesn’t need to know ahead of time what conditions are interesting to its users, which decreases coupling. The awaited condition might include inputs from several Ref-s. For example, to wait until x, y or z is non-zero write if (x() == 0 && y() == 0 && z() == 0) retry.

This sounds like busy-waiting, so what about efficiency? It would be very wasteful to roll back and reexecute continuously. Fortunately, there is no point in trying again until one of the accessed Ref-s has changed, and the STM knows which Ref-s were accessed. Internally retry is implemented using blocking constructs, so there is no busy-waiting.

Note that retry has a return type of Nothing, which indicates that it does not return normally.


While retry lets you tell the STM about a dead end in the backtracking search, orAtomic1 lets you give multiple search paths. orAtomic lets you chain atomic blocks that are alternative solutions to the search. If the first alternative calls retry then the second will be tried, if the second calls retry then the third will be tried, and so on. (Alternatives are only considered after an explicit dead end using retry, not when the transaction’s reads are inconsistent.) You can chain alternatives directly using orAtomic, or you can pass a collection of atomic blocks to atomic.oneOf.

The following code waits until one of x, y or z is non-zero, subtracts one from the first of the list that is non-zero, and records what was done in the String msg. The second block will only be attempted if the first block calls retry, and the third block will only be attempted if the second block calls retry. If all of the blocks call retry then the thread is blocked until one of the Ref-s is changed, and then the chain of alternatives will be retried.

val msg = atomic { implicit txn =>
  if (x() == 0)
  x -= 1
  "took one from x"
} orAtomic { implicit txn =>
  if (y() == 0)
  y -= 1
  "took one from y"
} orAtomic { implicit txn =>
  if (z() == 0)
  z -= 1
  "took one from z"


Composable timeouts are an innovative feature of ScalaSTM, so we’re especially interested in your constructive criticism.

There are several reasons why you might want retry to time out. Perhaps error logging or handling code should be invoked; perhaps the waiting thread should shut down if it doesn’t receive any work for a while; or perhaps you are using the STM to implement a higher-level interface that includes timeouts in its specification.

There are two ways to limit the amount of time that retry will wait. If you would like to set a policy for all calls to retry, you can use a modififed TxnExecutor that will cause timed-out retries to throw an InterruptedException. If timeout is not an exceptional behavior you can supply a timeout when waiting by using the retryFor operation, which falls through if the timeout has expired.

Using a custom TxnExecutor

The atomic keyword defined in the scala.concurrent.stm package object is actually just a method that returns a TxnExecutor instance. The apply(block) method on the returned instance is actually responsible for actually executing the atomic block. TxnExecutor also has a withRetryTimeout method that returns a new executor that will apply the retry timeout to all of the atomic block it executes. You can apply the retry timeout directly when you call atomic, you can capture and reuse the new executor with a new name, or you can use TxnExecutor.transformDefault to change the system-wide default executor returned by atomic.

atomic.withRetryTimeout(1000) { implicit txn =>
  // any retries in this atomic block will wait for at most 1000 milliseconds

val myAtomic = atomic.withRetryTimeout(1, TimeUnit.SECONDS)
myAtomic { implicit txn =>
  // this atomic block has a timeout of 1 seconds
myAtomic { ... }

TxnExecutor.transformDefault( _.withRetryTimeout(1000) )
atomic { implicit txn =>
  // all atomic blocks now default to a 1 second timeout

Using retryFor

If timeouts are part of the normal behavior, it may be convenient to handle the timeout without an exception. retryFor(timeout) acts like retry if the timeout has not yet expired, otherwise it returns immediately. Note that retryFor and retry have different return types (Unit and Nothing, respectively), because retryFor can return while retry always triggers rollback.

Assuming that pool is a transactional pool, the following code waits up to 100 milliseconds for a pooled instance to be available, after which the pool size is increased.

val instance = atomic { implicit txn =>
  if (!pool.hasAvailable) {

Waiting for Views

If the condition you’re waiting only involves a single Ref you can also block using Ref.View.await(pred). This method blocks until the predicate becomes true, then returns. Ref.View.tryAwait(duration)(pred) returns true if the predicate became true within the specified duration, otherwise it returns false.

Ref await and tryAwait are part of 0.3. In version 0.2 the functionality of await was called retryUntil, and there was no equivalent to tryAwait.

1 ScalaSTM’s orAtomic is referred to as orElse in the other STMs that support it. We changed the name to avoid confusion with Option.orElse.