12 breaking releases
0.13.0 | Nov 20, 2022 |
---|---|
0.12.0 | May 23, 2022 |
0.11.0 | May 14, 2022 |
0.10.0 | Feb 22, 2022 |
0.2.0 | Jun 7, 2020 |
#1021 in Network programming
23 downloads per month
500KB
11K
SLoC
paxakos
Paxakos is a pure Rust implementation of a distributed consensus algorithm based on Leslie Lamport's Paxos. It enables distributed systems to consistently modify shared state across their network, even in the presence of failures.
Usage
In order to use Paxakos, the traits LogEntry
, State
, NodeInfo
and Communicator
need to be implemented. The first two
describe what state will be replicated across the network and which
operations on it are available. The latter describe the nodes in your
network and how to communicate between them.
Below are two partial implementations of LogEntry
and State
. The other
two traits are more abstract and won't be illustrated here. Please refer to
the examples to get a fuller picture.
use std::collections::HashSet;
use std::convert::Infallible;
use paxakos::{LogEntry, State};
use uuid::Uuid;
#[derive(Clone, Copy, Debug, serde::Deserialize, serde::Serialize)]
pub enum CalcOp {
Add(f64, Uuid),
Div(f64, Uuid),
Mul(f64, Uuid),
Sub(f64, Uuid),
}
impl LogEntry for CalcOp {
type Id = Uuid;
fn id(&self) -> Self::Id {
match self {
CalcOp::Add(_, id)
| CalcOp::Div(_, id)
| CalcOp::Mul(_, id)
| CalcOp::Sub(_, id) => {
*id
}
}
}
}
#[derive(Clone, Debug)]
pub struct CalcState {
applied: HashSet<Uuid>,
value: f64,
}
impl State for CalcState {
type Context = ();
type LogEntry = CalcOp;
type Outcome = f64;
type Effect = f64;
type Error = Infallible;
fn apply(
&mut self,
log_entry: &Self::LogEntry,
_context: &mut (),
) -> Result<(Self::Outcome, Self::Effect), Self::Error> {
if self.applied.insert(log_entry.id()) {
match log_entry {
CalcOp::Add(v, _) => {
self.value += v;
}
CalcOp::Div(v, _) => {
self.value /= v;
}
CalcOp::Mul(v, _) => {
self.value *= v;
}
CalcOp::Sub(v, _) => {
self.value -= v;
}
}
}
Ok((self.value, self.value))
}
fn freeze(&self, _context: &mut Self::Context) -> Self::Frozen {
self.clone()
}
}
Motivation
Rust is a great language with which to implement efficient and truly reliable, fault-tolerant services. And while there already are several Rust implementations of consensus algorithms, they are either rudimentary or incomplete or their API was not approachable enough for this author.
Priorities
The project's priorities are as follows.
-
Correctness
The highest priority is correctness, which, in the context of consensus, requires stability, consistency and liveness.
- Stability implies that once a node learns that a log entry
a
has been appended to the distributed log, it will never learn that a different entryb
belongs in its place. - Consistency means that all nodes in the Paxakos network agree on the contents of their shared log. While nodes may temporarily fall behind, i.e. their log may be shorter than other nodes', their logs must be otherwise equivalent.
- Liveness means that the system won't get stuck, i.e. progress is always eventually made (assuming a there is no contention/some degree of cooperation).
- Stability implies that once a node learns that a log entry
-
Simplicity
Paxakos aims to be simple by providing few orthogonal primitives. To be clear, the goal is not to arbitrarily limit complexity. The goal is to have unentangled primitves; providing the same features with the least amount of complexity possible.
-
Ergonomics
Using Paxakos should be as pleasant as possible without sacrificing correctness or simplicity. The biggest challenge in this area are, at present, build times. If you have other concerns, please open an issue.
Features
Paxakos is a Multi-Paxos implementation. It supports membership changes, concurrency, as well as taking and installing snapshots.
Membership Changes
The State
trait exposes the cluster_at
method. By
implementing it, arbitrary membership changes may be made. No restrictions
are imposed and it is up to users and implementors to make sure that any
transition is safe.
Concurrency
Multi-Paxos allows for any number of rounds to be settled concurrently. This
can be exploited by implementing State
's concurrency
method.
Please note that when concurrency is enabled, gaps may appear in the log. These must be closed before entries after them can be applied to the state. This is typically done by appending no-op entries. A utility for doing so automatically is provided, but its API is far from stable.
Snapshots
A node may take a snapshot of its current state. The result is a combination
of the application specific State
as well as pakakos specifc information.
These snapshots may be used for graceful shutdown and restart, to bootstrap
nodes which have joined the cluster or to catch up nodes that have fallen
behind.
Decorations
Implementations of the Decoration
trait
can provide reusable functionality than can but need not be used. Paxakos
comes with several decorations (see below).
Heartbeats (heartbeats
flag)
The heartbeats decoration sends a heartbeat message at regular intervals.
Autofill (autofill
flag)
From time to time gaps will appear in the distributed log. For example due
to concurrency or dropped messages. When that happens, the
autofill
decoration will fill the gap, implicitly catching the node up or
making sure that queued log entries may be applied.
Track Leadership (track-leadership
flag)
The leadership tracking decoration infers which nodes are leading the cluster.
Ensure Leadership (experimental, ensure-leadership
flag)
Similar to heartbeats, the ensure leadership decoration will append a log entry after none has been for a certain amount of time. However the goal of this decoration is to ensure there is a leader.
Automatic Delegation (experimental, delegation
flag)
The Delegate
decoration allows automatic delegation of appends, ordinarily
to the leader node.
Automatic Catch-Up (experimental, catch-up
flag)
When a node is started or just installed a snapshot it is likely lagging.
The CatchUp
decoration will poll other nodes until the node is caught up.
Verification of Consistency (experimental, verify
flag)
The Verify
decoration will periodically check that all nodes are
consistent with each other. Nodes that have become inconsistent with the
majority eject their state to reduce the chance of propagation.
Leases
A cluster will often have shared resources which must be locked before they may be accessed. To account for node failures, locks time out unless they are refreshed. Such locks are commonly referred to as "leases".
Leaser (experimental, leaser
flag)
Leaser
makes taking, refreshing and releasing a lease as convenient as
calling take_lease
. The lease
is refreshed as long as the returned value is held onto.
Releaser (experimental, releaser
flag)
The releaser decoration makes sure that leases that have timed out are cleared away, releasing the underlying resource.
Master Leases (experimental, master-leases
flag)
Master leases are a mechanism to allow passive local reads. A comprehensive description may be found in Paxos Made Live.
Protocol
This section describes the Paxakos protocol. It is, for the most part, a typical Multi-Paxos protocol. Multi-Paxos generalizes Paxos to be run for multiple rounds, where each round represents a slot in the distributed log. Nodes may propose log entries to place in those slots. The liveness property guarantees that the cluster will, for each round, eventually converge on one of the proposed entries.
A central component of the protocol are coordination numbers. These are usually called "proposal numbers". However, because they are used throughout the protocol, Paxakos uses the more abstract term.
Appending an entry to the distributed log takes the following five steps.
-
Prepare
(RoundNum, CoordNum)
In order for a node to append an entry to the distributed log, it must first become leader for the given round. If it already believes itself leader for the round, it will skip to step 3.
To become leader for a round the node will send a prepare message containing the round number and a coordination number. The coordination number is chosen so that it is
- higher than any previously encountered coordination number and
- guaranteed not to be used by another node.
The former is important for liveness. The latter is required for stability and consistency and is achieved by exploiting the deterministic order of nodes returned by
cluster_at
. -
Vote
When a node receives a prepare request, it checks that it hasn't accepted a previous such request with a coordination number that's higher than the given one. If it has, it will reply with a conflict. If it hasn't, the node will either abstain, i.e., choose not to give its vote, or it sends back a promise not to accept any more proposals with a coordination number that's less the given one.
-
Promise
(Vec<(RoundNum, CoordNum, LogEntry)>)
The promise is a set of triples, each consisting of a round number, a coordination number and a log entry. It can be thought to mean "I acknowledge your bid to become leader and give you my vote. However, in return you must respect these previous commitments I've made."
-
Conflict
(CoordNum, Option<(CoordNum, LogEntry)>)
A rejection is sent with the greatest observed coordination number so far. For the special case that the round has already converged and the node still has it available, it will send it along as well.
-
Abstention
A
The node chose to abstain. By default nodes will never abstain. This can be changed by providing a custom
Voter
implementation. The argument typeA
is defined byCommunicator::Abstain
andVoter::Abstain
.
-
-
Propose
(RoundNum, CoordNum, LogEntry)
When a node sent a
prepare(r, c)
request and received a quorum or more of promises in return (counting its own), it will believe itself to be leader for all roundsr..
. It may now start proposing log entries for any of these rounds, usingc
as the coordination number.The only restriction is that it must respect the promises it has received. If multiple promises contain a triple with the same round number, the one with the greatest coordination number wins. (Triples with the same round and coordination number will have the same log entry as well.)
-
Vote
When a node receives a proposal, it will check whether it has made any conflicting promises. If it has, it responds with a conflict. If it hasn't, it may choose to accept or reject the proposal and reply accordingly.
-
Acceptance
Y
/ RejectionN
By default nodes accept all proposals with
Y = ()
. This can be changed by providing a customVoter
implementation. The argument typesY
andN
are defined byCommunicator::Yea
,Communicator::Nay
,Voter::Yea
andVoter::Nay
. -
Conflict
(CoordNum, Option<(CoordNum, LogEntry)>)
See 2.2.
-
-
Commit
(RoundNum, CoordNum, LogEntry::Id)
/ CommitById(RoundNum, CoordNum, LogEntry)
After having received a quorum of acceptances, the round has converged on the proposed entry. The leader node will commit the entry locally and inform other nodes as well. Nodes who sent an acceptance will only be sent the log entry's id, others will receive the full entry.
Project Status
This section examines different aspects of paxakos, how far along they are and what's missing or in need of improvement.
☀️ Consensus Implementation ☀️
The core algorithm of paxakos is well-tested and has been exercised a lot. There is reason to be confident in its correctness.
⛅ Passive Mode ⛅
Passive mode is implemented and superficially tested. Thorough testing is still needed.
⛅ Serialization ⛅
Snapshot serialization is serde based and still maturing.
⛅ API Stability ⛅
The API has been fairly stable and there is no reason to expect big changes.
The decorations may be reworked so that their configuration can be moved
into the State
.
⛅ Efficiency ⛅
Paxakos supports concurrency and has a (for now rudimentary) implementation of master leases. Assuming a scheme to delegate to the current master, this combination is highly efficient.
Future Direction
This is a side project. I will work on the following as my time allows (in no particular order).
- Tests
- Adding comments and documentation
- Rounding off existing decorations
- Implement decoration for read delegation
- Find way to detect and react to when a node was left behind
- Update playground
- Allow configuration of
CatchUp
andVerify
- Visualize
CatchUp
- Add master-lease decoration
- Add delegate decoration
- Allow configuration of
License: GPL-3.0-only
Dependencies
~1.6–4.5MB
~79K SLoC