It seems to me that the Adaptive framework[1][2] is well-suited to STM transactions. Consider this snippet:

expensive_transaction :: STM Int -- does a bunch of work

example :: TVar Int -> STM Int example t = (+) <$> readTVar t <*> expensive_transaction

If the TVar passed to example is rapidly changing (perhaps it represents the current time), then this transaction will starve; it never gets a chance to complete, even if the values used by expensive_transaction aren't changing at all, we still re-evaluate the entire action each time we "retry" this transaction. But STM has a wonderful property: what we need to commit a transaction is just the result value and a transaction log; both of these depend only on the values of TVars read by that transaction. So, if "expensive_transaction" doesn't refer to the TVar "t", then even if "t" is rapidly changing, the transaction log generated by "expensive_transaction" must not change, and we can just re-generate the part of the transaction log for "readTVar t" and recompute the (+). This gets even better when considering big processes that are blocked by an explicit "retry"; if some change in input is just going to cause the transaction to retry again (and go back to sleep), we'd like that to happen as quickly as possible. Has anyone looked down this path? It seems quite promising to me. -- ryan [1] http://hackage.haskell.org/package/Adaptive [2] Acar et al, Adaptive Functional Programming, POPL 2002 http://citeseer.ist.psu.edu/old/752721.html