Cross-posted from the Apache Mesos blog.
High availability is one of the key features of Mesos. For example, a typical Mesos cluster in production involves 3 to 5 masters with one acting as leader and the rest on standby. When a leading master fails due to a crash or goes offline for an upgrade, a standby master automatically becomes the leader without causing any disruption to running tasks. Leader election is currently performed by using Apache ZooKeeper.
A recoverable slave is critical for running services in production on Mesos for several reasons:
In a typical production environment there are stateful services (e.g., caches) running in the cluster. It is not uncommon for these services to have a high startup time (e.g., cache warm up time of a few hours). Even in the analytics world, there are cases where a single task is responsible for doing work that takes hours to complete. In such cases a restart of the slave (e.g., crash or upgrade) will have a huge impact on the service. While sharding the service wider mitigates this impact it is not always possible to do so (e.g., legacy services, data locality, application semantics). Such stateful applications would benefit immensely from running under a more resilient slave.
It is important for clusters to frequently upgrade their infrastructure to stay up-to-date with the latest features and bug fixes. A typical Mesos slave update involves stopping it, updating its libraries and starting it back. In production environments there is always tension between upgrading the infrastructure frequently and the need to not impact long running services. With respect to Mesos upgrades, if upgrading the slave binary has no impact on underlying services, then it is a win for both cluster operators and service owners.
While upgrading the slaves the most often the reason for restarting slaves, there might be other causes for a slave to fail. A slave crash could happen due to a bug in the slave code or due to external factors like a bug in the kernel or ZooKeeper client library. Typically such crashes are temporary and a restart of the slave is enough to correct the situation. If such slave restarts do not affect applications running on the slave it is a huge win for the applications.
Slave recovery works by having the slave checkpoint enough information (e.g., task information, executor information, status updates) about the running tasks and executors to local disk. Once the slave and the framework(s) enable checkpointing, any subsequent slave restarts would recover the checkpointed information and reconnect with the executors. When a checkpointing slave process goes down, both the leading master and the executors running on the slave host wait for the slave to come back up and reconnect. A nice thing about slave recovery is that frameworks and their executors/tasks are oblivious to a slave restart.
As part of this feature, the executor driver has also been improved to make it more resilient in the face of a slave failure. As an example, status updates sent by the executor while the slave is down are cached by the driver and sent to the slave when it reconnects with the restarted slave. Since this is all handled by the executor driver, framework and executor writers do not have to worry about it. The executors can keep sending status updates for their tasks while remaining oblivious to the slave being up or down.
Another benefit of slave checkpointing the status updates is that now updates are more reliably delivered to frameworks in the face of failures. Before slave recovery if the slave fails at the same time that a master is failing over, no TASK_LOST updates for tasks running on the slave were sent to the framework. This is partly because the Mesos master is stateless. A failed over master reconstructs the cluster state from the information given to it by re-registering slaves and frameworks. With slave recovery, status updates and tasks are no longer lost when slaves fail. Rather, the slave recovers tasks, status updates and reconnects with the running executors. Even if an executor does terminate when the slave is down, a recovered slave knows about it from its checkpointed state and reliably sends TASK_LOST updates to the framework.
For more information about how to enable slave recovery in your cluster, please refer to the documentation.
In a similar vein to how slave recovery makes upgrading a Mesos cluster easy, we would like to enable frameworks to upgrade their executors/tasks as well. Currently the only way to upgrade an executor/task is to kill the old executor/task and launch the new upgraded executor/task. For the same reasons as we have discussed earlier this is not ideal for stateful services. We are currently investigating proper primitives for enabling frameworks to do such upgrades, so that not every framework have to (re-)implement that logic.
While slave recovery greatly improves the reliability of delivering status updates, there are still some rare cases where updates could be lost. For example, if a slave crashes when a master is failing over and never comes back then the new leading master doesn’t know about the lost slave and executors/tasks running on it. In addition to status updates, any driver methods (e.g., launchTasks, killTask) invoked when the master is failing over are silently dropped. Currently, frameworks are responsible for reconciling their own task state using the reconciliation API. We are currently investigating ways to provide better guarantees around reconciliation.
Currently, updating Mesos involves a cluster operator to manually upgrade the master and slave binaries and roll them in a specific manner (e.g., masters before slaves). But what if Mesos could update itself? It is not hard to imagine a future where Mesos masters can orchestrate the upgrade of slaves and maybe also upgrade one another! This would also help making rollbacks easy incase an upgrade doesn’t work because it would be much easier for Mesos to check if various components are working as expected.
So what are you waiting for? Go ahead and give Mesos a whirl and let us know what you think via Twitter at @ApacheMesos. Also, Mesos is an independent open source community that’s organized by its members, including you! Whether you’re running or writing a framework, or hacking the core, there are opportunities for you to get in touch and ask questions (via the mailing list), get involved locally or contribute back.