I'm considering using a TVar to store some state in a web application (that can be recreated on restart). However, the contention aspects of TVar concern me. It seems that a frequent short running transaction can starve out longer transactions by continually interrupting them. Also, as more longer running transactions keep restarting, this would increase load on the CPU, tending to further increase the length of these transactions. Eventually I feel this could cause the server to become completely unresponsive.
Considering this, I have these questions:
(1) Can TVar (or another data type) use locks, not simultaneous attempts/retries.
(2) Can TVar (or another data type) have some different contention mechanism, i.e. "let transactions run for a second before running another transaction", or at least some guarantee that transactions will eventually complete (i.e. a contention algorithm that prevents starvation for longer running transactions).
This is only a concern if you have many cheap transactions that update data and a few expensive ones that read it. Analytics on a live-updated dataset, perhaps.
If you're actually concerned about this, consider using a flag TVar. Set it to false and check that it is false at the start of every cheap transaction, calling
retry
otherwise. Then simply set it to true before entering into your longrunning transaction, and set it to false on exit. Alternately, you can simply guard your state behind aTMVar
. Your longrunning computation takes the tmvar atomically, does what it feels like, and then returns it. The other transactions take place entirely within a single actual STM transaction.Remember also, a long running STM transaction is sort of a tricky beast. Due to laziness, you can put an expensive value into a var cheaply. You can also read a "snapshot" of data from a whole bunch of vars very quickly. To have a genuinely long-running transaction you need to read from a whole bunch of vars, then based on what you've read, compute what vars you are going to write new values into (or read values from), and that computation itself must be expensive. So likely you're not even in that scenario to begin with!