newFastLogger1 ensures the ordering of logs #207
Conversation
| writer buffd@BufFD{..} tvar ref mvar = loop (0 :: Int) | ||
| where | ||
| loop cnt = do | ||
| cnt' <- atomically $ do |
There was a problem hiding this comment.
I'm wondering if you're already using TVar for spin-waiting, wouldn't it be easier to just use TQueue and a poison pill instead of TVar, IORef and get rid of MVar for signalling shutdown?
P.S. It took me a while to decipher the meaning of tvar , ref, mvar. Maybe it would be a good idea to use more meaningful names?
There was a problem hiding this comment.
STM is heavier than atomicModifyIORef'.
In this design, STM transactions run only when the builder reaches the limit.
Yes, MVar Buffer is not necessary. I should think whether or not we can remove it from SingleLogger easily.
|
@carbolymer @khibino Now |
|
Merged. I will release a new major version! |
|
Released! |
@carbolymer @khibino This should fix #203.
The old logic cannot ensure the ordering of logs even if the number of internal builders are 1.
This PR calls the old one
MultiLoggerand introduces a new one calledSingleLogger.SingleLoggeruses a queue of builder in addition to an internal builder.The internal builder gets empty and the queue glows atomically.
The builders of the queue are written by a single dedicated thread, thus the ordering is maintained.
A test case is also provided.