The Nomad client and server agents collect a wide range of runtime metrics related to the performance of the system. Operators can use this data to gain real-time visibility into their cluster and improve performance. Additionally, Nomad operators can set up monitoring and alerting based on these metrics in order to respond to any changes in the cluster state.
On the server side, leaders and followers have metrics in common as well as metrics that are specific to their roles. Clients have separate metrics for the host metrics and for allocations/tasks, both of which have to be explicitly enabled. There are also runtime metrics that are common to all servers and clients.
There are three ways to obtain metrics from Nomad:
Query the /metrics API endpoint to return metrics for the current Nomad process (as of Nomad 0.7). This endpoint supports Prometheus formatted metrics.
Send the USR1 signal to the Nomad process. This will dump the current telemetry information to STDERR (on Linux).
Configure Nomad to automatically forward metrics to a third-party provider.
The recommended practice for alerting is to leverage the alerting capabilities of your monitoring provider. Nomad’s intention is to surface metrics that enable users to configure the necessary alerts using their existing monitoring systems as a scaffold, rather than to natively support alerting. Here are a few common patterns.
Export metrics from Nomad to Prometheus using the StatsD exporter, define alerting rules in Prometheus, and use Alertmanager for summarization and routing/notifications (to PagerDuty, Slack, etc.). A similar workflow is supported for Datadog.
Periodically submit test jobs into Nomad to determine if your application deployment pipeline is working end-to-end. This pattern is well-suited to batch processing workloads.
Deploy Nagios on Nomad. Centrally manage Nomad job files and add the Nagios monitor when a new Nomad job is added. When a job is removed, remove the Nagios monitor. Map Consul alerts to the Nagios monitor. This provides a job-specific alerting system.
Write a script that looks at the history of each batch job to determine whether or not the job is in an unhealthy state, updating your monitoring system as appropriate. In many cases, it may be ok if a given batch job fails occasionally, as long as it goes back to passing.
»Key Performance Indicators
The sections below cover a number of important metrics
»Consensus Protocol (Raft)
Nomad uses the Raft consensus protocol for leader election and state replication. Spurious leader elections can be caused by networking issues between the servers, insufficient CPU resources, or insufficient disk IOPS. Users in cloud environments often bump their servers up to the next instance class with improved networking and CPU to stabilize leader elections, or switch to higher-performance disks.
nomad.raft.leader.lastContact metric is a general indicator of
Raft latency which can be used to observe how Raft timing is
performing and guide infrastructure provisioning. If this number
trends upwards, look at CPU, disk IOPs, and network
nomad.raft.leader.lastContact should not get too close to
the leader lease timeout of 500ms.
nomad.raft.replication.appendEntries metric is an indicator of
the time it takes for a Raft transaction to be replicated to a quorum
of followers. If this number trends upwards, check the disk I/O on the
followers and network latency between the leader and the followers.
The details for how to examine CPU, IO operations, and networking are
specific to your platform and environment. On Linux, the
package contains a number of useful tools. Here are examples to
vmstat 1, cloud provider metrics for "CPU %"
sar -d, cloud provider metrics for "volume write/read ops" and "burst balance"
netstat -s, cloud provider metrics for interface "allowance"
nomad.raft.fsm.apply metric is an indicator of the time it takes
for a server to apply Raft entries to the internal state machine. If
this number trends upwards, look at the
nomad.nomad.fsm.* metrics to
see if a specific Raft entry is increasing in latency. You can compare
this to warn-level logs on the Nomad servers for "attempting to apply
large raft entry". If a specific type of message appears here, there
may be a job with a large job specification or dispatch payload that
is increasing the time it takes to apply raft messages.
»Federated Deployments (Serf)
Nomad uses the membership and failure detection capabilities of the Serf library
to maintain a single, global gossip pool for all servers in a federated
deployment. An uptick in
msg.suspect is a reliable indicator
that membership is unstable.
If these metrics increase, look at CPU load on the servers and network latency and packet loss for the Serf address.
The Scheduling documentation describes the workflow of how evaluations become scheduled plans and placed allocations. The following metrics, listed in the order they are emitted, allow an operator to observe changes in throughput at the various points in the scheduling process.
nomad.worker.invoke_scheduler.<type> - The time to run the scheduler of the given type. Each scheduler worker handles one evaluation at a time, entirely in-memory. If this metric increases, examine the CPU and memory resources of the scheduler.
nomad.broker.total_blocked - The number of blocked evaluations. Blocked evaluations are created when the scheduler cannot place all allocations as part of a plan. Blocked evaluations will be re-evaluated so that changes in cluster resources can be used for the blocked evaluation's allocations. An increase in blocked evaluations may mean that the cluster's clients are low in resources or that job have been submitted that can never have all their allocations placed. Nomad also emits a similar metric for each individual scheduler. For example
nomad.broker.batch_blockedshows the number of blocked evaluations for the batch scheduler.
nomad.broker.total_unacked - The number of unacknowledged evaluations. When an evaluation has been processed, the worker sends an acknowledgment RPC to the leader to signal to the eval broker that processing is complete. The unacked evals are those that are in-flight in the scheduler and have not yet been acknowledged. An increase in unacknowledged evaluations may mean that the schedulers have a large queue of evaluations to process. See the
invoke_schedulermetric (above) and examine the CPU and memory resources of the scheduler. Nomad also emits a similar metric for each individual scheduler. For example
nomad.broker.batch_unackedshows the number of unacknowledged evaluations for the batch scheduler.
nomad.plan.evaluate - The time to evaluate a scheduler plan submitted by a worker. This operation happens on the leader to serialize the plans of all the scheduler workers. This happens entirely in memory on the leader. If this metric increases, examine the CPU and memory resources of the leader.
nomad.plan.wait_for_index - The time required for the planner to wait for the Raft index of the plan to be processed. If this metric increases, refer to the Consensus Protocol (Raft) section above.
nomad.plan.submit - The time to submit a scheduler plan from the worker to the leader. This operation requires writing to Raft and includes the time from
nomad.plan.wait_for_index(above). If this metric increases, refer to the Consensus Protocol (Raft) section above.
nomad.plan.queue_depth - The number of scheduler plans waiting to be evaluated after being submitted. If this metric increases, examine the
nomad.plan.submitmetrics to determine if the problem is in general leader resources or Raft performance.
Upticks in any of the above metrics indicate a decrease in scheduler throughput.
The importance of monitoring resource availability is workload specific. Batch processing workloads often operate under the assumption that the cluster should be at or near capacity, with queued jobs running as soon as adequate resources become available. Clusters that are primarily responsible for long running services with an uptime requirement may want to maintain headroom at 20% or more. The following metrics can be used to assess capacity across the cluster on a per client basis.
»Task Resource Consumption
The metrics listed here can be used to track resource consumption on a per task basis. For user facing services, it is common to alert when the CPU is at or above the reserved resources for the task.
»Job and Task Status
We do not currently surface metrics for job and task/allocation status, although we will consider adding metrics where it makes sense.
Runtime metrics apply to all clients and servers. The following metrics are general indicators of load and memory pressure.
It is recommended to alert on upticks in any of the above, server memory usage in particular.