Skip to content
This repository


Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
Fetching contributors…

Cannot retrieve contributors at this time

file 156 lines (112 sloc) 7.603 kb
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156

Redis Cluster is currenty a work in progress, however there are a few things
that you can do already with it to see how it works.

The following guide show you how to setup a three nodes cluster and issue some
basic command against it.


3) Show link status detection with CLUSTER NODES
4) Show how to add slots with CLUSTER ADDSLOTS
5) Show redirection
6) Show cluster down



*** WARNING: all the following problably has some meaning only for
*** me (antirez), most info are not updated, so please consider this file
*** as a private TODO list / brainstorming.

- disconnect FAIL clients after some pong idle time.


* Majority rule: the cluster con continue when there are all the hash slots covered AND when there are the majority of masters.
* Shutdown on request rule: when a node sees many connections closed or even a timeout longer than usual on almost all the other nodes, it will usually wait for the normal timeout before to change the state, unless it receives a query from a client: in such a case it will put itself into error status.


* When asked for a key that is not in a node's business it will reply:

    -ASK (in case we want the client to ask just one time)
    -MOVED <slotid> (in case the hash slot is permanently moved)

So with -ASK a client should just retry the query against this new node, a single time.

With -MOVED the client should update its hash slots table to reflect the fact that now the specified node is the one to contact for the specified hash slot.

* Nodes communicate using a binary protocol.

* Node failure detection.

    1) Every node contains information about all the other nodes:
        - If this node is believed to work ok or not
        - The hash slots for which this node is responsible
        - If the node is a master or a slave
        - If it is a slave, the slave of which node
        - if it is a master, the list of slave nodes
        - The slaves are ordered for "<ip>:<port>" string from lower to higher
          ordered lexicographically. When a master is down, the cluster will
          try to elect the first slave in the list.

    2) Every node also contains the unix time where every other node was
       reported to work properly (that is, it replied to a ping or any other
       protocol request correctly). For every node we also store the timestamp
       at which we sent the latest ping, so we can easily compute the current

    3) From time to time a node pings a random node, selected among the nodes
       with the least recent "alive" time stamp. Three random nodes are selected
       and the one with lower alive time stamp is pinged.

    4) The ping packet contains also information about a few random nodes
       alive time stamp. So that the receiver of the ping will update the
       alive table if the received alive timestamp is more recent the
       one present in the node local table.

       In the ping packet every node "gossip" information is somethig like



    5) The node replies to ping with a pong packet, that also contains a random
       selections of nodes timestamps.

A given node thinks another node may be in a failure state once there is a
ping timeout bigger than 30 seconds (configurable).

When a possible failure is detected the node performs the following action:

    1) Is the average between all the other nodes big? For instance bigger
       than 30 seconds / 2 = 15 seconds? Probably *we* are disconnected.
       In such a case we don't trust our lag data, and reset all the
       timestamps of sent ping to zero. This way when we'll reconnect there
       is no risk that we'll claim many nodes are down, taking inappropriate

    2) Messages from nodes marked as failed are *always* ignored by the other
       nodes. A new node needs to be "introduced" by a good online node.

    3) If we are well connected (that is, condition "1" is not true) and a
       node timeout is > 30 seconds, we mark the node as POSSIBLE_FAILURE
       (a flat in the cluster node structure). Every time we sent a ping
       to another node we inform this other nodes that we detected this
       condition, as already stated.

    4) Once a node receives a POSSIBLE_FAILURE status for a node that is
       already marked as POSSIBLE_FAILURE locally, it sends a message
       to all the other nodes of type NODE_FAILURE_DETECTED, communicating the
       ip/port of the specified node.

       All the nodes need to update the status of this node setting it into

    5) If the computer in FAILURE state is a master node, what is needed is
       to perform a Slave Election.


    1) The slave election is performed by the first slave (with slaves ordered
       lexicographically). Actually it is the first functioning slave, so if
       the first slave is marked as failing the next slave will perform the
       election and so forth. Such a slave is called the "Successor".

    2) The Successor starts checking that all the nodes in the cluster already
       marked the master in FAILURE state. If at least one node does not agree
       no action is performed.

    3) If all the nodes agree that the master is failing, the Successor does
       the following:

       a) It will send a SUCCESSION message to all the other nodes, that will
          upgrade the hash slot tables accordingly. It will make sure that all
          the nodes are updated and if some node did not received the message
          it will keep trying.
       b) Once all nodes not marked as FAILURE accepted the SUCCESSION message
          it will update his own table and will start acting as a master
          accepting write queries.
       c) Every node receiving the succession message, if not already informed
          of the change will broadcast the same message to other three random
          nodes. No action is performed if the specified host was already marked
          as the master node.
       d) A node that was a slave of the original master that failed will
          switch master to the new one once the SUCCESSION message is received.


    1) When selecting a slave, the system will try to pick one with an IP different than the master and other slaves, if possible.

    2) The PING packet also contains information about the local configuration checksum. This is the SHA1 of the current configuration, without the bits that normally change form one node to another (like latest ping reply, failure status of nodes, and so forth). From time to time the local config SHA1 is checked against the list of the other nodes, and if there is a mismatch between our configuration and the most common one that lasts for more than N seconds, the most common configuration is asked and retrieved from another node. The event is logged.

    3) Every time a node updates its internal cluster configuration, it dumps such a config in the cluster.conf file. On startup the configuration is reloaded.
    Nodes can share the cluster configuration when needed (for instance if SHA1 does not match) using this exact same format.


    - Clients may be configured to use slaves to perform reads, when read-after-write consistency is not required.
Something went wrong with that request. Please try again.