Adjoint's implementation of the Raft consensus algorithm. See original paper for further details about the protocol.
Raft proposes a strong single-leader approach to consensus. It simplifies operations, as there are no conflicts, while being more efficient than other leader-less approaches due to the high throughput achievable by the leader. In this leader-driven consensus algorithm, clients must contact the leader directly in order to communicate with the system. The system needs to have an elected leader in order to be available.
In addition to a pure core event loop, this library uses the systematic concurrency testing library dejafu to test certain properties about streams of events throughout the system. Random thread interleavings are generated in a raft network and realistic event streams are delivered to each node's event queue. We test for the absence of deadlocks and exceptions, along with checking that the convergent state of the system matches the expected results. These concurrency tests can be found here.
Ensuring valid transitions between node states
Each server in Raft can only be in one of these three states:
- Leader: Active node that handles all client interactions and send AppendEntries RPCs to all other nodes.
- Candidate: Active node that attempts to become a leader.
- Follower: Passive node that just responds to RPCs.
Temporal (e.g. ElectionTimeout) and spatial (e.g. AppendEntries or RequestVote) events cause nodes to transition from one state to another.
   ------------> Follower --------------> Candidate --------------> Leader ^ ^ | | | |  | | | |_______________________| | | | |  | |___________________________________________________| -  Starts up | Recovers -  Times out | Starts election -  Receives votes from majority of servers and becomes leader -  Discovers leader of new term | Discovers candidate with a higher term -  Discovers server with higher term
All nodes in the Raft protocol begin in the follower state. A follower will stay a follower unless it fails to hear from a leader or a candidate requesting a vote within its ElectionTimeout timer. If this happens, a follower will transition to a candidate state. These node states are illustrated in the type:
data Mode = Follower | Candidate | Leader
The volatile state a node keeps track of may vary depending on the mode that it
is in. Using
GADTs, we relate these specific node state
datatypes that contain the relevant data to the current node's mode with the
NodeState type. This way, we can enforce that the volatile state carried by a
node in mode
Follower is indeed
-- | The volatile state of a Raft node data NodeState (a :: Mode) v where NodeFollowerState :: FollowerState v -> NodeState 'Follower v NodeCandidateState :: CandidateState v -> NodeState 'Candidate v NodeLeaderState :: LeaderState v -> NodeState 'Leader v
The library's main event loop is comprised of a simple flow: Raft nodes receive
events on an STM channel, handle the event depending on the current node state,
return a list of actions to perform, and then perform those actions in the order
they were generated. The
Event type specifies the main value to which raft
nodes react to, whereas the
Action type specifies the action the raft node
performs as a result of the pairing of the current node state and received
The Raft protocol has constraints on how nodes transition from one state to another. For example, a follower cannot transition to a leader state without first transitioning to a candidate state. Similarly, a leader can never transition directly to a candidate state due to the algorithm specification. Candidates are allowed to transition to any other node state.
To adhere to the Raft specification, we make use of some type level programming to ensure that only valid transitions happen between node states.
-- | All valid state transitions of a Raft node data Transition (init :: Mode) (res :: Mode) where StartElection :: Transition 'Follower 'Candidate HigherTermFoundFollower :: Transition 'Follower 'Follower RestartElection :: Transition 'Candidate 'Candidate DiscoverLeader :: Transition 'Candidate 'Follower HigherTermFoundCandidate :: Transition 'Candidate 'Follower BecomeLeader :: Transition 'Candidate 'Leader HandleClientReq :: Transition 'Leader 'Leader SendHeartbeat :: Transition 'Leader 'Leader DiscoverNewLeader :: Transition 'Leader 'Follower HigherTermFoundLeader :: Transition 'Leader 'Follower Noop :: Transition init init
To compose the
Transition with the resulting state from the event handler, we
ResultState datatype, existentially quantifying the result state mode:
-- | Existential type hiding the result type of a transition, fixing the -- result state to the state dictated by the 'Transition init res' value. data ResultState init v where ResultState :: Transition init res -> NodeState res -> ResultState init
This datatype fixes the result state to be dependent on the transition that
occurred; as long as the allowed transitions are correctly denoted in the
Transition data constructors, only valid transitions can be specified by the
ResultState values existentially hide the result
state types, that can be accessed via pattern matching. Thus, all event
handlers, be they RPC handlers, timeout handlers, or client request handlers,
have a type signature of the form:
handler :: NodeState init -> ... relevant handler data ... -> ResultState init
ResultState will enforce that invalid transitions are not made
when writing handlers for all combinations of raft node modes and events. In the
future, this approach may be extended to limit the actions a node can emit
dependent on its current mode.
Within the Raft protocol, there is a pure core that can be abstracted without the use of global state. The protocol can be looked at simply as a series of function calls of a function from an initial node state to a result node state. However, sometimes these transitions have side effects. In this library we have elected to separate the pure and effectful layers.
The core event handling loop is a pure function that, given the current node
state and a few extra bits of global state, computes a list of
Actions for the
effectful layer to perform (updating global state, sending a message to another
node over the network, etc.).
In order to update the replicated state machine, clients contact the leader via "client requests" containing commands to be committed to the replicated state machine. Once a command is received, the current leader assesses whether it is possible to commit the command to the replicated state machine.
The replicated state machine must be deterministic such that every command committed by a leader to the state machine will eventually be replicated on every node in the network at the same index.
As the only part of the internal event loop that needs to be specified manually,
We ask users of our library to provide an instance of the state machine
typeclass. This typeclass relates a state machine type to a command type
and a single type class function
applyCmdRaftStateMachinePure, a pure function that
should return the result of applying the command to the initial state machine.
class RaftStateMachinePure sm v | sm -> v where data RaftStateMachinePureError sm v type RaftStateMachinePureCtx sm v = ctx | ctx -> sm v applyCmdRaftStateMachinePure :: RaftStateMachinePureCtx sm v -> sm -> v -> Either (RaftStateMachinePureError sm v) sm
Everything else related to the core event handling loop is not exposed to library users. All that needs to be specified is the type of the state machine, the commands to update it, and how to perform those updates.
In this Raft implementation, there are four components that need access to global state and system resources. Firstly, raft nodes must maintain some persistent state for efficient and correct recovery from network outages or partitions. Secondly, raft nodes need to send messages to other raft nodes for the network (the replicated state machine) to be operational. Next, library users must specify what event channel datastructure to use. Finally, the programmer must provide a way to fork threads and run a list of effectful actions concurrently.
The reasons for the latter two design decisions-- requiring the programmer to
provide an event channel type and new/read/writeChannel primitives, a way to
fork an effectful action to run concurrently, and a way to run a list of actions
concurrently-- is a result of property based concurrency testing that we do,
test/TestDejaFu.hs. In order to test the system as a whole, to run
several nodes concurrently and test invariants about the system such as the
absence of deadlocks, we must be able to swap out the base monad for the
ConcIO monad, which has implementations of concurrency primitives that act
deterministically. This allows us to test that the raft nodes run correctly in
a wide space of thread interleavings giving us more confidence that our code is
correct, assuming "correct" implementations of the
Each node persists data to disk, including the replicated log entries. Since persisting data is an action that programmers have many opinions and preferences regarding, we provide two type classes that abstract the specifics of writing log entries to disk as well as a few other small bits of relevant data. These are separated due to the nature in which the log entries are queried, often by specific index and without bounds. Thus, it may be desirable to store the log entries in an efficient database. The remaining persistent data is always read and written atomically, and has a much smaller storage footprint.
The actions of reading or modifying existing log entries on disk is broken down even further: we ask the user to specify how to write, delete, and read log entries from disk. Often these types of operations can be optimized via smarter persistent data solutions like modern SQL databases, thus we arrive at the following level of granularity:
-- | The type class specifying how nodes should write log entries to storage. class Monad m => RaftWriteLog m v where type RaftWriteLogError m -- | Write the given log entries to storage writeLogEntries :: Exception (RaftWriteLogError m) => Entries v -> m (Either (RaftWriteLogError m) ()) -- | The type class specifying how nodes should delete log entries from storage. class Monad m => RaftDeleteLog m v where type RaftDeleteLogError m -- | Delete log entries from a given index; e.g. 'deleteLogEntriesFrom 7' -- should delete every log entry deleteLogEntriesFrom :: Exception (RaftDeleteLogError m) => Index -> m (Either (RaftDeleteLogError m) (Maybe (Entry v))) -- | The type class specifying how nodes should read log entries from storage. class Monad m => RaftReadLog m v where type RaftReadLogError m -- | Read the log at a given index readLogEntry :: Exception (RaftReadLogError m) => Index -> m (Either (RaftReadLogError m) (Maybe (Entry v))) -- | Read log entries from a specific index onwards readLogEntriesFrom :: Exception (RaftReadLogError m) => Index -> m (Either (RaftReadLogError m) (Entries v)) -- | Read the last log entry in the log readLastLogEntry :: Exception (RaftReadLogError m) => m (Either (RaftReadLogError m) (Maybe (Entry v)))
To read and write the
PersistentData type (the remaining persistent data that
is not log entries), we ask the user to use the following
-- | The RaftPersist type class specifies how to read and write the persistent -- state to disk. class Monad m => RaftPersist m where type RaftPersistError m readPersistentState :: Exception (RaftPersistError m) => m (Either (RaftPersistError m) PersistentState) writePersistentState :: Exception (RaftPersistError m) => PersistentState -> m (Either (RaftPersistError m) ())
The other non-deterministic, effectful part of the protocol is the communication between nodes over the network. It can be unreliable due to network delays, partitions and packet loss, duplication and reordering, but the Raft consensus algorithm was designed to achieve consensus in such harsh conditions.
The actions that must be performed in the networking layer are sending RPCs to other raft nodes, receiving RPCs from other raft nodes, sending client responses to clients who have issued requests, and receiving client requests from clients wishing to update the replicated state. Depending on use of this raft library, the two pairs are not necessary symmetric and so we do not force the user into specifying a single way to send/receive messages to and from raft nodes or clients.
We provide several type classes for users to specify the networking layer
themselves. The user must make sure that the
receiveClient pairs perform complementary actions; that an RPC
sent from one raft node to another is indeed receivable via
receiveRPC on the
node to which it was sent:
-- | Interface for nodes to send messages to one -- another. E.g. Control.Concurrent.Chan, Network.Socket, etc. class RaftSendRPC m v where sendRPC :: NodeId -> RPCMessage v -> m () -- | Interface for nodes to receive messages from one -- another class RaftRecvRPC m v where type RaftRecvRPCError m v receiveRPC :: m (Either (RaftRecvRPCError m v) (RPCMessage v)) -- | Interface for Raft nodes to send messages to clients class RaftSendClient m sm v where sendClient :: ClientId -> ClientResponse sm v -> m () -- | Interface for Raft nodes to receive messages from clients class RaftRecvClient m v where type RaftRecvClientError m v receiveClient :: m (Either (RaftRecvClientError m v) (ClientRequest v))
We have written a default implementation for network sockets over TCP in src/Examples/Raft/Socket
The core of the effectful layers of this Raft implmentation is the event channel. Since different data channels have different performance, we ask the programmer to supply an implementation of such a channel via yet another type class and type family:
class Monad m => MonadRaftChan v m where type RaftEventChan v m readRaftChan :: RaftEventChan v m -> m (Event v) writeRaftChan :: RaftEventChan v m -> Event v -> m () newRaftChan :: m (RaftEventChan v m)
On spawning a raft node, the program will create a new event channel using
newRaftChan. Then, the event producers will be forked; These event producers
will use the aforementioned typeclasses like
to wait for messages from other raft nodes or clients wishing to contact the
node. Once a message is received, a message event is constructed from the
message contents and written to the main event channel via
the main thread, the core event handler will be repeatedly reading events from
the event channel using
readRaftChan and performing the correct action in
response to each event.
The last of the type class instances the programmer must provide for the monad
they are running the raft node in is a
MonadRaftFork, which provides the main
raft loop with the ability to fork a concurrent action; The raft node needs to
know how to fork actions in the monad. This is necessary for the raft node to
be able to fork its event producers, and run other actions concurrently; e.g.
a leader responding to all followers at the same time during a heartbeat RPC
broadcast. The typeclass is defined as follows:
-- | The typeclass encapsulating the concurrency operations necessary for the -- implementation of the main event handling loop. class Monad m => MonadRaftFork m where type RaftThreadId m raftFork :: RaftThreadRole -- ^ The role of the current thread being forked -> m () -- ^ The action to fork -> m (RaftThreadId m)
The implementation of this typeclass is a bit subtle, and it is advised that
programmers do not implement it from scratch themselves. A default
implementation is provided for the
IO type using standard concurrency
primitives making it easy for the programmer to simply rely on that
implementation. An example instance of this type class for a custom monad
transformer stack with
IO the bottom:
instance MonadRaftFork MyMonad where type RaftThreadId MyMonad = RaftThreadId IO raftFork threadRole myMonad = lift $ raftFork threadRole (runMyMonad myMonad)
The last thing to mention about this typeclass is the
that must be passed to invocations of the
raftFork function. In some
applications (and, noteably our concurrency testing suite) thread names can be
used for debugging and even message passing purposes. For instance of
MonadRaftFork that do not need to distinguish threads by name, simply ignore
The Raft Example (
In this library we provide a full fledged, non-production ready, example
implementation/s of monad transformers and type class instances for all type
classes necessary to run a raft node. They can be found in
src/Examples/Raft/... or in
RaftExampleT (found in
RaftSocketT (found in
RaftLogFileStoreT (found in
RaftPersistFileStoreT (found in
Programmers can use these files and implementations for references when implementing the necessary type class instances for their bespoke monads, or even use some of the monad transformers in their own stack!
We provide a complete example of the library where nodes communicate via network sockets, and they write their logs on text files. See app/Main.hs to have further insight.
Build the example executable:
$ stack build
In separate terminals, run some raft nodes:
The format of the cmd line invocation is:
$ raft-example node <fresh/existing> <file/postgres> <node-id> <peer-1-node-id> ... <peer-n-node-id>
We are going to run a network of three nodes:
- On terminal 1:
$ stack exec raft-example node fresh file localhost:3001 localhost:3002 localhost:3003
- On terminal 2:
$ stack exec raft-example node fresh file localhost:3002 localhost:3001 localhost:3003
- On terminal 3:
$ stack exec raft-example node fresh file localhost:3003 localhost:3001 localhost:3002
The first node spawned should become candidate once its election's timer times out and request votes to other nodes. It will then become the leader, once it receives a majority of votes and will broadcast messages to all nodes at each heartbeat.
Note: If you want to run a raft example node with existing persistent data, pass the
existingcommand line option to the
raft-exampleprogram instead of
$ stack exec raft-example node existing ...
Note: The example also runs using a PostgreSQL database as long as a user 'libraft_test' with password 'libraft_test' exists in your local postgresql installation. For ease of experimentation, if such a user does not exist, create it like so:
$ sudo -su postgres psql -U postgres -c "CREATE USER libraft_test WITH CREATEDB PASSWORD 'libraft_test';"
Run a client:
$ stack exec raft-example client
In the example provided, there are five basic operations:
addNode <host:port>: Add a nodeId to the set of nodeIds that the client will communicate with. Adding a single node will be sufficient, as this node will redirect the command to the leader in case he is not.
getNodes: Return all node ids that the client is aware of.
read: Return the state of the leader.
set <var> <val>: Set a variable to a specific value.
incr <var>: Increment the value of a variable.
Assuming that two nodes are run as mentioned above, a valid client workflow would be:
>>> addNode localhost:3001 >>> set testVar 4 >>> incr testVar >>> read
It will return the state of the leader's state machine (and eventually the state of all nodes in the Raft network). In our example, it will be a map of a single key
How to use this library
- Define the state machine
- Implement the networking layer
- Implement the persistent layer
- Putting it all together
Define the state machine
The only requirement for our state machine is to instantiate the state machine
-- | Interface to handle commands in the underlying state machine. Functional --dependency permitting only a single state machine command to be defined to --update the state machine. class RaftStateMachinePure sm v | sm -> v where data RaftStateMachinePureError sm v type RaftStateMachinePureCtx sm v = ctx | ctx -> sm v rsmTransition :: RaftStateMachinePureCtx sm v -> sm -> v -> Either (RaftStateMachinePureError sm v) sm
In our example we use a simple map as a store whose values can only increase.
Implement the networking layer
We leave the choice of the networking layer open to the user, as it can vary depending on the use case (E.g. TCP/UDP/cloud-haskell/etc).
We need to specify how nodes will communicate with clients and with each other.
As described above in the Networking
section, it suffices to
implement those four type classes (
Note that our datatypes will need to derive instances of
MonadConc. This allows us to test concurrent
properties of the system, using randomized thread scheduling to assert the
absence of deadlocks and exceptions.
In case of the
RaftSocketT data type used in our example:
deriving instance MonadConc m => MonadThrow (RaftSocketT v m) deriving instance MonadConc m => MonadCatch (RaftSocketT v m) deriving instance MonadConc m => MonadMask (RaftSocketT v m) deriving instance MonadConc m => MonadConc (RaftSocketT v m)
Implement the persistent layer
There are many different possibilities when it comes to persist data to disk, so we also leave the specification open to the user.
As explained in the Persistent
State section above, we
will create instances for
RaftDeleteLog to specify how we will read, write and
delete log entries, as well as
RaftPersist. There are actually several data
that must be stored on disk; 1) the data the raft paper calls "persistent data"
and 2) the log entries of the node.
We provide an implementation that stores the persistent data in a file in src/Examples/Raft/FileStore/Persistent.hs
An example of storing log entries in a single file (for ease of implementation in lieu of good performance for reads/writes) can be found in src/Examples/Raft/FileStore/Log.hs
Lastly, a more "production ready" example of log entry storage using a
PostgreSQL database can be found in src/Raft/Log/PostgreSQL.hs.
This implementation is used in our
quickcheck-state-machine model testing
module and is thus thoroughly tested.
Putting it all together
The last step is wrapping our previous data types that deal with
networking and persistent data into a single monad that also derives instances
of all the Raft type classes described (
In our example, this monad is
RaftExampleM sm v. See
Finally, we are ready to run our Raft nodes. We call the
file, together with the function we define to run the stack of monads that
derive our Raft type classes.
Test suite dependencies
The test suite depends on libfiu (commonly installed with package
which it uses to simulate network failures. In addition, the test suite also
depends on libpq-dev and postgresql. Furthermore, in order to successfully run
the model tests (which will run autmatically when executing
stack test, but
fail immediately with "Failed to spawn node), you will have to run the following
$ sudo -su postgres psql -U postgres -c "CREATE USER libraft_test WITH CREATEDB PASSWORD 'libraft_test';"
Ongaro, D., Ousterhout, J. In Search of an Understandable Consensus Algorithm, 2014
Howard, H. ARC: Analysis of Raft Consensus 2014