Skip to content

Latest commit

 

History

History
76 lines (59 loc) · 3.62 KB

global.md

File metadata and controls

76 lines (59 loc) · 3.62 KB
Raw

Precondition

  1. Fully Connected Cluster

    Because it depends on global's name registration service. The name server must maintains a fully connected network.

    For example, if node N1 connects to node N2 (which is already connected to N3), the global name servers on the nodes N1 and N3 ensure that also N1 and N3 are connected. In other words, command-line flag -connect_all false can not be used.

  2. Same Global Configuration

    All node's global_quorum_size and global_jobs must keep the same value. This ensures that the global task manager can transfer between nodes when the network splits.

Configuration

  1. global_jobs

    the same format as local_jobs, default is []. This means only run local jobs without running global task manager and monitor processes.

  2. global_quorum_size

    ecron application live on at least global_quorum_size nodes in the same cluster, can be regarded as a healthy cluster.

    Global task manager only running on a healthy cluster.

    If you want to guarantee always no more than one global task manager even when the cluster has network split, you should set it to "half plus one". For example:

    Run on majority:

    1. ABC 3 nodes in one cluster.
    2. global_quorum_size=2.
    3. (ABC) cluster split into 2 part(AB =|= C).
    4. the global task manager would run on AB cluster(AB is the healthy cluster now).
    5. C node only running local jobs without global jobs.

    Run on majority

    1. ABC 3 nodes in one cluster.
    2. global_quorum_size=2.
    3. (ABC) cluster split into 3 part(A =|= B =|= C).
    4. every node only running local jobs without global jobs(all nodes is unhealthy).

    Run on every node if brain split.

    1. ABC nodes in one cluster.
    2. global_quorum_size=1.
    3. (ABC) cluster split into 3 part(A =|= B =|= C).
    4. the global task manager would run on every nodes(we have three healthy cluster now).
    5. But the global task manager only running one in the same cluster.

Implementation

  1. The top supervisor ecron_sup start at first.
  2. Nothing will happen if the global_jobs is empty.
  3. ecron_sup would start_link ecron_monitor worker (gen_server) if global_jobs is not empty.
  4. ecron_monitor subscribes node's up/down messages by net_kernel:monitor_nodes(true), when it initializes.
  5. Enter ecron_monitor main loop:
    • Checking if there is enough ecron process in the cluster(global_quorum_size).
    • Trying to terminate global_job manager process when cluster's ecron number less than global_quorum_size.
    • Otherwise, trying to start a global_job manager process, This gen_server register by global:register_name/2.
    • All the nodes are rushing to register this global jobs manager process, only one node will success, other node's ecron_monitor would link this process if the process already exists.
    • The ecron_monitor should receive notification, when node down/up or the global_job manager has terminated.
    • Goto step 5 again, When notified.
     NodeA             NodeB             NodeC
   supervisor        supervisor        supervisor
       |                |     |            |
    monitor          monitor  |          monitor
         \              |     |            /
          \           [link]  |           /
           \            |     |          /
           [link]-----GlobalJob-----[link]

NodeB has won the race and spawn global manager process, other node's ecron_monitor only link to this manager.