Bounty is a drop-in replacement for memcached that also supports a subset of the relational algebra. Inspired by Neha Narula’s presentation about Pequod at Ricon East 2013.
N.B. Bounty is currently highly experimental, and not even kind of complete.
-
Implemented in Haskell.
-
The front-end parser is build in Attoparsec for speed. As much as possible is done using bytestrings, again for speed.
-
All IO is handled via conduits for constant resource usage. The TCP server uses network-conduit for the same reason. UDP and Unix-domain socket styles are planned.
-
The underlying data structure is just a Map while I get the interface working. Obviously, this needs to be replaced with something more concurrency-friendly once things get going.
-
Tests are all written in HUnit for the moment.
-
After researching a number of different options, I’m exploring the possibility of using Prefix Hash Trees (PHTs) for distributed range queries. Based on that paper, I’m implementing this in a separate library.
-
I keep getting this feeling like Knuth has already solved this, but I can’t find it. Maybe Volume 5.
-
There are some space leaks around pushing data onto the STM Map in a naïve fashion. If the final product was going to store everything in a lazy btree wrapped in a TVar, I would be remiss for not explicitly copying key and value bytestrings and evaluating the btree to normal form prior to shoving it back into STM. Instead, I’m already researching more concurrency-friendly data structures (see Notes).
-
Some inputs can totally crash the server. Ouch. That’s gonna need some exception handling. I’ll add that when I reorganize how commands are implemented (it’s ugly right now).