Skip to content
Go to file
Cannot retrieve contributors at this time
93 lines (68 sloc) 3.31 KB

Gryadka is a minimalistic master-master replicated consistent key-value storage based on the CASPaxos protocol. It uses Redis as a backend and makes multiple Redis instances work as a whole tolerating up to F failures out of 2F+1 nodes.

Its core has less than 500 lines of code but provides full-featured Paxos implementation supporting such advanced features as cluster membership change (ability to add/remove nodes to a cluster) and distinguished proposer optimization (using one round trip to change a value instead of two).

Is it correct?

Yes, the protocol has formal proof (see the CASPaxos paper) and TLA+ models independently written by Greg Rogers and Tobias Schottdorf:

Moreover, the implementation was heavily tested with fault injections on the network layer.

Is it production ready?

No, it's an educational project, and it was never intended to be deployed in production. The goal of the project is to build master-master replicated consistent key-value storage as simple as possible.

Even though Gryadka is an educational project, its operational characteristics surpass established databases (see the comparison).


Gryadka's core interface is a change function which takes two arguments:

  • a key
  • a update function

change gets a value associated with the key, applies update to calculate a new value, saves and returns it.

The pseudo-code:

class Paxos {
  constuctor() { = ...;
  change(key, update) {
    const value = update(;, value);
    return value;

By choosing the appropriate update functions, it's possible to customize Gryadka to fulfill different tasks. A "last write win" key/value could be implemented as:

class LWWKeyValue {
  constuctor(paxos) {
    this.paxos = paxos;
  read(key) {
    return this.paxos.change(key, x => x);
  write(key, value) {
    return this.paxos.change(key, x => value);

A key/value storage with compare-and-set support may look like:

class CASKeyValue {
  constuctor(paxos) {
    this.paxos = paxos;
  read(key) {
    return this.paxos.change(key, x => x==null ? { ver: 0, val: null} : x);
  write(key, ver, val) {
    return this.paxos.change(key, x => {
      if (x.ver != ver) throw new Error();
      return { ver: ver+1, val: val };

Dockerized example of cluster membership change

Please see for an example of:

  • Dockerized Gryadka-based cluster
  • Using Gryadka to build a HTTP key-value storage
  • Cluster membership change

Additional information about cluster membership change is in the CASPaxos paper and in the simulation tests (see 2p2k2.a3.a4 and c2p2k2.flux tests):


Testing is done by mocking the network layer and checking consistency invariants during various network fault injections such as message dropping and message reordering.

See for more information.

You can’t perform that action at this time.