A Haskell implementation of distributed hash tables with distributed two-phase commit. The views can hold locks, and have replicas that they always communicate with to achieve quorum-based consensus.
Written as a project for COMP 360 Distributed Systems, Fall 2016, Prof. Jeff Epstein, Wesleyan University.
Make sure you have Stack installed.
Clone the repository and run stack install. The executable files
dht-client, dht-server, dht-view-leader must now be ~/.local/bin, which
is probably in your path variable.
You can run the server without any command line arguments: dht-server.
For dht-client and dht-view-leader, you can see all the options using the --help flag.
The executable dht-client take optional arguments --server (or -s) and
--viewleader (or -l) that specifies which host to connect , such a call
would be of the form
dht-client --server 127.0.0.1 setr "lang" "Türkçe"
The executable dht-server also takes the --viewleader optional argument, to
be able to perform heartbeats.
If the optional argument --server isn't provided, it is assumed to be the local computer name.
The optional argument --viewleader (or with its short form -l) can be passed multiple times for each view replica, such as:
dht-view-leader -l mycomputer:39000 -l mycomputer:39001 -l mycomputer:39002 -l mycomputer:39003 -l mycomputer:39004
If the optional argument --viewleader isn't provided, it is assumed that there are 3 default inputs: your computer name with the ports 39000, 39001 and 39002.
If a connection to a host name and a port, or a bind to a port doesn't happen in 10 seconds, you will get a timeout error and the program will try the next available port number.
If a server fails to send a heartbeat for 30 seconds, but then later succeeds, it will be told by the view leader that it expired. In that case, the server terminates.
When there is a request for a lock that receives the retry message, that client is added to the queue for that lock. Even though the client stops asking, it will get the lock when the clients that are waiting in front of it in the queue get and release the lock.
The view leader assumes that a server crashed if it sends no heartbeat for 30
seconds. In that case, if that server holds any locks (that have the requester
identifier format :38000, where the number is the port) are cancelled
automatically.
When the server has to return retry for a lock get request, it checks if there is a deadlock in the system by building a directed graph and looking for cycles. If there is, it logs the requesters and locks involved. Right now, it repeatedly logs the same deadlock information as long as it keeps getting the request. Since the client keeps retrying by default, it will not stop logging the deadlock unless the client is stopped. If a requester ID is used for multiple lock requests at the same time, this can cause some weird behavior.
Deadlock detection is currently deactivated, will be reactivated after refactoring for consensus.
We assume that our hash function evenly distributes strings to integers. We take the UUID of every server to be the string for its hash, i.e. the hash value of the server. Therefore we assume that our servers more or less evenly split the number of keys we want to hold. The primary bucket to hold a key is the bucket that has the server that has the lowest hash value that still is strictly greater than the hash of the key string. We hold 2 more replicas as backup, in servers that have the next 2 lowest hash value.
If some servers are removed, then each server will check if the keys they currently hold used to reside in one of the servers that are removed. If that's the case, then it will make a request to the new server that is suppose to hold the data, to save the data.
When there are new servers, each server will check what keys they have. If a server has a key that should reside in a different server now, it will make a request to that different server and then delete that key from itself.