JS7 JobScheduler

Space shortcuts

Page tree

You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 10 Next »

Introduction

Use of a JS7 - Controller Cluster provides high availability and is a feature subject to the JS7 - License.

  • Fail-over is an automated operation that occurs when the Primary Controller is aborted or killed. Fail-over is applied in case of abnormal termination only.
  • Switch-over is an operation that is caused by user intervention in JOC Cockpit or by use of the JS7 - REST Web Service API. The procedure to switch-over does not require termination of an Active Controller Instance, instead it shifts the active role to the second Controller instance.

For fail-over and switch-over a dedicated Standalone Agent acting as a Cluster Watch Agent is required.

For command line references see the JS7 - Controller - Command Line Operation article.

Cluster Roles

Controller Cluster

The documentation frequently indicates a Primary Controller Instance and a Secondary Controller Instance. The names suggest that one Controller Instance is primarily used and one is for backup purposes.

  • The wording in cluster terms suggests to indicate the Active Controller Instance and the Standby Controller Instance independently from the fact if the Primary or Secondary Controller Instance is active.
  • A Controller implements an active-passive cluster, however, the term passive is misleading as the Standby Controller Instance is not passive at all but records any order state transitions occurring in the Active Controller Instance. Both Controller instances hold a journal of order state transitions that is actively synchronized. Fail-over and switch-over will occur only if both Controller instances' journals are in sync.
  • The Cluster presents itself as a single unit to the outside world, i.e. to JOC Cockpit and to Agents.
    • Any operations performed in JOC Cockpit are automatically applied to the Active Controller Instance.
    • At any point in time only one Controller instance is active and the other instance is in standby mode.

Cluster Watch Agent

Primary and Secondary Controller instances require a dedicated Standalone Agent to be available that acts as an arbitrator in case of fail-over and switch-over.

Start-up of Controller Cluster

On start-up both Primary and Secondary Controller instances connect to the Cluster Watch Agent.

  • The Cluster Watch Agent casts its vote which Controller instance owns the leading journal.
  • Start-up of a Controller Cluster is not possible with the Cluster Watch Agent being unavailable.
  • Start-up of the previously Active Controller Instance without the Standby Controller Instance being available is possible if the Cluster Watch Agent is active.

Failure of the Active Controller Instance

In case of failure of the Active Controller Instance the Cluster Watch Agent plays its role as an arbitrator:

  • The Cluster Watch Agent knows immediately when the Active Controller Instance is down due to a connection loss from this instance.
  • The Standby Controller Instance holds a connection to the Active Controller Instance and knows immediately when this connection is lost.
  • Failure of the Active Controller Instance is the point in time when the Standby Controller Instance and the Cluster Watch Agent check to find common ground about a cluster fail-over operation: They determine if they should declare the Active Controller Instance being inoperable and after a short period of 1-2s they proceed and cast their votes if the Standby Controller Instance should now become the Active Controller Instance.
  • As a prerequisite for fail-over both the Cluster Watch Agent and the Standby Controller Instance have to confirm that the Standby Controller Instance's journal was in sync with the Active Controller Instance at the point in time of failure.

Operation of Cluster Watch Agent

Above explanations suggest that a Cluster Watch Agent may never be running on the hosts that the Primary and Secondary Controller instances are operated for.

  • If the Cluster Watch Agent is terminated at the same time as a failed Active Controller Instance then no fail-over can occur.
  • If the Cluster Watch Agent is terminated at the same time as one of the Controller instances then the Controller Cluster cannot start up as this requires operational readiness of the Cluster Watch Agent.
  • A Cluster Watch Agent that is started after failure of the Active Controller Instance is disqualified from casting its vote as it has no knowledge if the Controller instances' journals are  in sync.

Cluster Operations

Cluster operations include an automated fail-over and a manual switch-over of the Active Controller Instance.

Fail-over

Fail-over occurs when the Active Controller Instance is terminated abnormally. 

Fail-over can be invoked by the following actions:

  • The Active Controller Instance is killed, for example
    • for Unix with a SIGKILL signal corresponding to the command: kill -9
    • for Windows with the command: taskkill /F
  • The operating system crashes.
  • In the JS7 - Dashboard the user performs one of the operations: 
    • Active Controller Instance action menu: Abort -> With fail-over
    • Active Controller Instance action menu: Abort and restart -> With fail-over
  • From the command line the user performs one of the operations:
    • controller.sh | .cmd abort
    • controller.sh | .cmd kill

No fail-over occurs when

  • the Active Controller Instance is stopped normally from the command line:
    • controller.sh | .cmd stop
  • the operating system is shut down and systemd / init.d or a Windows Service are in place to stop the Controller normally.

Fail-over happens within a short period of time, typically in 2-3s.

Switch-over

Switch-over occurs exclusively when invoked by user intervention.

Switch-over can be invoked by the following actions:

  • In the JS7 - Dashboard the user performs one of the operations: 
    • Active Controller Instance action menu: Abort -> With fail-over
    • Active Controller Instance action menu: Abort and restart -> With fail-over

No switch-over occurs when

  • the Active Controller Instance is stopped normally from the command line:
    • controller.sh | .cmd stop
  • the operating system is shut down and systemd / init.d or a Windows Service are in place to stop the Controller normally.

Switch-over happens within a short period of time, typically in 2-3s.

A Warning to Users trying to implement their own Clustering Mechanism

Users might be tempted to implement their own clustering with Standalone Controller Instances, for example

  • using tools for virtual machine management such as VMware®,
  • using Microsoft® Windows Server Cluster or similar cluster solutions.

The best advice is not to apply such clustering mechanisms. Reasons include but are not limited to the following issues:

  • A Controller Cluster guarantees high availability when used with a JS7 - Agent Cluster. Use of Standalone Agents limits high availability.
  • The cluster has to guarantee that only one of both Controller instances is started at any point in time.
    • Should this rule not be observed then both Controller instances will instruct Agents to execute the same workflows and jobs which will result in double job execution.
    • Controller journals will be messed up with the same orders in different state transitions.
    • In this situation the only solution is to drop both Controller instance's journals that are available from the state sub-directory, to accept that any orders are lost and to redeploy scheduling objects.
  • There is no simple way to determine if a Controller instance is not in perfect shape to manage orders.
    • Performing PID file checks is of limited use: this can prove the unavailability of a Controller instance. However, a positive PID file check does not prove that a Controller instance works.
    • Log file analysis is pointless. Controllers are heavily making use of asynchronous operations when it comes to Agents. Occurrence of error messages in log files does not prevent a situation to be recovered within the next few seconds.

Further Resources


  • No labels