title: Apache Mesos - Agent Recovery layout: documentation

Agent Recovery

If the mesos-agent process on a host exits (perhaps due to a Mesos bug or because the operator kills the process while upgrading Mesos), any executors/tasks that were being managed by the mesos-agent process will continue to run. When mesos-agent is restarted, the operator can control how those old executors/tasks are handled:

  1. By default, all the executors/tasks that were being managed by the old mesos-agent process are killed.
  2. If a framework enabled checkpointing when it registered with the master, any executors belonging to that framework can reconnect to the new mesos-agent process and continue running uninterrupted.

Hence, enabling framework checkpointing enables tasks to tolerate Mesos agent upgrades and unexpected mesos-agent crashes without experiencing any downtime.

Agent recovery works by having the agent checkpoint information (e.g., Task Info, Executor Info, Status Updates) about the tasks and executors it is managing to local disk. If a framework enables checkpointing, any subsequent agent restarts will recover the checkpointed information and reconnect with any executors that are still running.

Note that if the operating system on the agent is rebooted, all executors and tasks running on the host are killed and are not automatically restarted when the host comes back up.

Framework Configuration

A framework can control whether its executors will be recovered by setting the checkpoint flag in its FrameworkInfo when registering with the master. Enabling this feature results in increased I/O overhead at each agent that runs tasks launched by the framework. By default, frameworks do not checkpoint their state.

Agent Configuration

Three configuration flags control the recovery behavior of a Mesos agent:

  • strict: Whether to do agent recovery in strict mode [Default: true].

    • If strict=true, all recovery errors are considered fatal.
    • If strict=false, any errors (e.g., corruption in checkpointed data) during recovery are ignored and as much state as possible is recovered.

  • recover: Whether to recover status updates and reconnect with old executors [Default: reconnect].

    • If recover=reconnect, reconnect with any old live executors, provided the executor's framework enabled checkpointing.
    • If recover=cleanup, kill any old live executors and exit. Use this option when doing an incompatible agent or executor upgrade!

    NOTE: If no checkpointing information exists, no recovery is performed and the agent registers with the master as a new agent.

  • recovery_timeout: Amount of time allotted for the agent to recover [Default: 15 mins].

    • If the agent takes longer than recovery_timeout to recover, any executors that are waiting to reconnect to the agent will self-terminate.

NOTE: If none of the frameworks have enabled checkpointing, the executors and tasks running at a agent die when the agent dies and are not recovered.

A restarted agent should re-register with master within a timeout (75 seconds by default: see the --max_agent_ping_timeouts and --agent_ping_timeout configuration flags). If the agent takes longer than this timeout to re-register, the master shuts down the agent, which in turn will shutdown any live executors/tasks. Therefore, it is highly recommended to automate the process of restarting a agent (e.g., using a process supervisor such as monit or systemd).

Known issues with systemd and POSIX isolation

There is a known issue when using systemd to launch the mesos-agent while also using only posix isolation mechanisms that prevents tasks from recovering. The problem is that the default KillMode for systemd processes is cgroup and hence all child processes are killed when the agent stops. Explicitly setting KillMode to process allows the executors to survive and reconnect.

The following excerpt of a systemd unit configuration file shows how to set the flag:

[Service]
ExecStart=/usr/bin/mesos-agent
KillMode=process

NOTE: There are also known issues with using systemd and raw cgroups based isolation, for now the suggested non-Posix isolation mechanism is to use Docker containerization.