Monitoring Nomad
The Nomad client and server agents collect a wide range of runtime metrics. These metrics are useful for monitoring the health and performance of Nomad clusters. Careful monitoring can spot trends before they cause issues and help debug issues if they arise.
All Nomad agents, both servers and clients, report basic system and Go runtime metrics.
Nomad servers all report many metrics, but some metrics are specific to the leader server. Since leadership may change at any time, these metrics should be monitored on all servers. Missing (or 0) metrics from non-leaders may be safely ignored.
Nomad clients have separate metrics for the host they are running on as well as for each allocation being run. Both of these metrics must be explicitly enabled.
By default, the Nomad agent collects telemetry data at a 1 second interval. Note that Nomad supports gauges, counters, and timers.
There are three ways to obtain metrics from Nomad:
Query the /v1/metrics API endpoint to return metrics for the current Nomad process. 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 such as DataDog, Prometheus, statsd, and Circonus.
Alerting
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
Nomad servers' memory, CPU, disk, and network usage all scales linearly with cluster size and scheduling throughput. The most important aspect of ensuring Nomad operates normally is monitoring these system resources to ensure the servers are not encountering resource constraints.
The sections below cover a number of other 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.
The 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
latency. nomad.raft.leader.lastContact should not get too close to
the leader lease timeout of 500ms.
The 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 sysstat
package contains a number of useful tools. Here are examples to
consider.
CPU -
vmstat 1, cloud provider metrics for "CPU %"IO -
iostat,sar -d, cloud provider metrics for "volume write/read ops" and "burst balance"Network -
sar -n,netstat -s, cloud provider metrics for interface "allowance"
The 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. Try shrinking the size
of the job either by putting distinct task groups into separate jobs,
downloading templates instead of embedding them, or reducing the count on
task groups.
Scheduling
The Scheduling documentation describes the workflow of how evaluations become scheduled plans and placed allocations.
Progress
There is a class of bug possible in Nomad where the two parts of the scheduling pipeline, the workers and the leader's plan applier, disagree about the validity of a plan. In the pathological case this can cause a job to never finish scheduling, as workers produce the same plan and the plan applier repeatedly rejects it.
While this class of bug is very rare, it can be detected by repeated log lines
on the Nomad servers containing plan for node rejected:
While it is possible for these log lines to occur infrequently due to normal cluster conditions, they should not appear repeatedly and prevent the job from eventually running (look up the evaluation ID logged to find the job).
Plan rejection tracker
Nomad provides a mechanism to track the history of plan rejections per client
and mark them as ineligible if the number goes above a given threshold within a
time window. This functionality can be enabled using the
plan_rejection_tracker server configuration.
When a node is marked as ineligible due to excessive plan rejections, the following node event is registered:
Along with the log line:
If a client is marked as ineligible due to repeated plan rejections, try draining the node and shutting it down. Misconfigurations not caught by validation can cause nodes to enter this state: #11830.
If the plan for node rejected log does appear repeatedly with the same
node_id referenced but the client is not being set as ineligible you can try
adjusting the plan_rejection_tracker configuration of servers.
Performance
The following metrics allow observing 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.blocked_evals.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.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 examplenomad.broker.batch_unackedshows the number of unacknowledged evaluations for the batch scheduler.nomad.broker.total_pending - The number of pending evaluations in the eval broker. Nomad processes only one evaluation for a given job concurrently. When an unacked evaluation is acknowledged, Nomad will discard all but the latest evaluation for a job. An increase in this metric may mean that the cluster state is changing more rapidly than the schedulers can keep up.
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. If this metric approaches 5 seconds, scheduling operations may fail and be retried. If possible reduce scheduling load until metrics improve.
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.evaluateandnomad.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.evaluateandnomad.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.
Capacity
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.
- nomad.client.allocated.cpu
- nomad.client.unallocated.cpu
- nomad.client.allocated.disk
- nomad.client.unallocated.disk
- nomad.client.allocated.iops
- nomad.client.unallocated.iops
- nomad.client.allocated.memory
- nomad.client.unallocated.memory
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
See Job Summary Metrics for monitoring the health and status of workloads running on Nomad.
Runtime Metrics
Runtime metrics apply to all clients and servers. The following metrics are general indicators of load and memory pressure.
- nomad.runtime.num_goroutines
- nomad.runtime.heap_objects
- nomad.runtime.alloc_bytes
It is recommended to alert on upticks in any of the above, server memory usage in particular.
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 member.flap and/or 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.