fast-downward
fast-downward copied to clipboard
circuithub
Prevent planner from locking up on large problems.
In some scenarios, the downward executable was producing a lot of
output which was not being read by the library. This would eventually
cause the downward process to hang, waiting for someone to consume its
output. This change drains both output handles before calling wait, and
thus preventing the deadlock.
Other solution would be to not pipe the outputs seeing as they aren't being used.
This is interesting, I'm not sure I've observed this at all in years of running this in production. Do you have any way to reproduce this?
Set the list length as required to make downward spit out more than whatever your OS is allowing it to buffer. This patch will make it work.
import FastDownward
import FastDownward.Exec as Exec
data Add = Add Int
main :: IO ()
main = do
let
initial :: [Int]
initial = [0..500]
res <- runProblem $ do
vars <- mapM newVar initial
solve Exec.bjolp
(map (\i -> modifyVar (vars !! i) negate) initial)
(zipWith (?=) vars (map negate initial))
print res
You can tell it's hung by this (and not because the problem is too large) because your CPU will not be busy, as all processes are sleeping, downward waiting for your app to read from the pipes, and your app is waiting for downward to finish.
Did you manage to reproduce? 90 was enough to overflow the buffers on macOS and on my NixOS machine. That is, 90 vars, with 90 effects.
I've stuck the complete example in a gist.
https://gist.github.com/luke-clifton/29df0f6cc0664a3eaa64e6e433506cc5
Technically, one should really be reading from both of those handles at the same time. If the stderr handle fills up, and we are still trying to read the stdout handle, it would block as well, a fact that can be witnessed by swapping the order that we read the inputs in in this patch and seeing that it will lock up again.
To combat that we could possible use Control.Concurrent.Async.concurrently
(stderr, stdout) <- concurrently
(Data.Text.IO.hGetContents stderrHandle)
(Data.Text.IO.hGetContents stdoutHandle)
Which would be totally safe at the cost of adding async as a direct dependency. (It is already a transitive one though).