open-nomad/contributing/architecture-eval-triggers.md

10 KiB

Architecture: Evaluation Triggers

The Scheduling in Nomad internals documentation covers the path that an evaluation takes through the leader, worker, and plan applier. This document describes what events within the cluster cause Evaluations to be created.

Evaluations have a TriggeredBy field which is always one of the values defined in structs.go:

const (
	EvalTriggerJobRegister          = "job-register"
	EvalTriggerJobDeregister        = "job-deregister"
	EvalTriggerPeriodicJob          = "periodic-job"
	EvalTriggerNodeDrain            = "node-drain"
	EvalTriggerNodeUpdate           = "node-update"
	EvalTriggerAllocStop            = "alloc-stop"
	EvalTriggerScheduled            = "scheduled"
	EvalTriggerRollingUpdate        = "rolling-update"
	EvalTriggerDeploymentWatcher    = "deployment-watcher"
	EvalTriggerFailedFollowUp       = "failed-follow-up"
	EvalTriggerMaxPlans             = "max-plan-attempts"
	EvalTriggerRetryFailedAlloc     = "alloc-failure"
	EvalTriggerQueuedAllocs         = "queued-allocs"
	EvalTriggerPreemption           = "preemption"
	EvalTriggerScaling              = "job-scaling"
	EvalTriggerMaxDisconnectTimeout = "max-disconnect-timeout"
	EvalTriggerReconnect            = "reconnect"
)

The list below covers each trigger and what can trigger it.

  • job-register: Creating or updating a Job will result in 1 Evaluation created, plus any follow-up Evaluations associated with scheduling, planning, or deployments.
  • job-deregister: Stopping a Job will result in 1 Evaluation created, plus any follow-up Evaluations associated with scheduling, planning, or deployments.
  • periodic-job: A periodic job that hits its timer and dispatches a child job will result in 1 Evaluation created, plus any additional Evaluations associated with scheduling or planning.
  • node-drain: Draining a node will create 1 Evaluation for each Job on the node that's draining, plus any additional Evaluations associated with scheduling or planning.
  • node-update: When the fingerprint of a client node has changed or the node has changed state (from up to down), Nomad creates 1 Evaluation for each Job running on the Node, plus 1 Evaluation for each system job that has datacenters that include the datacenter for that Node.
  • alloc-stop: When the API that serves the nomad alloc stop command is hit, Nomad creates 1 Evaluation.
  • scheduled: Nomad's internal housekeeping will periodically create Evaluations for garbage collection.
  • rolling-update: When a system job is updated, the update block's stagger field controls how many Allocations will be scheduled at a time. The scheduler will create 1 Evaluation to follow-up for the next set.
  • deployment-watcher: When a service job is updated, the update block controls how many Allocations will be scheduled at a time. The deployment watcher runs on the leader and monitors Allocation healthy. It will create 1 Evaluation when the Deployment has reached the next step.
  • failed-follow-up: Evaluations that hit a delivery limit and will not be retried by the eval broker are marked as failed. The leader periodically reaps failed Evaluations and creates 1 new Evaluation for these, with a delay.
  • max-plan-attempts: The scheduler will retry Evaluations that are rejected by the plan applier with a new cluster state snapshot. If the scheduler exceeds the maximum number of retries, it will create 1 new Evaluation in the blocked state.
  • alloc-failure: If an Allocation fails and exceeds its maximum restart attempts, Nomad creates 1 new Evaluation.
  • queued-allocs: When a scheduler processes an Evaluation, it may not be able to place all Allocations. It will create 1 new Evaluation in the blocked state to be processed later when node updates arrive.
  • preemption: When Allocations are preempted, the plan applier creates 1 Evaluation for each Job that has been preempted.
  • job-scaling: Scaling a Job will result in 1 Evaluation created, plus any follow-up Evaluations associated with scheduling, planning, or deployments.
  • max-disconnect-timeout: When an Allocation is in the unknown state for longer than the max_client_disconnect window, the scheduler will create 1 Evaluation.
  • reconnect: When a Node in the disconnected state reconnects, Nomad will create 1 Evaluation per job with an allocation on the reconnected Node.

Follow-up Evaluations

Almost any Evaluation processed by the scheduler can result in additional Evaluations being created, whether because the scheduler needs to follow-up from failed scheduling or because the resulting plan changes the state of the cluster. This can result in a large number of Evaluations when the cluster is in an unstable state with frequent changes.

Consider the following example where a node running 1 system job and 2 service jobs misses its heartbeat and is marked lost. The Evaluation for the system job is successfully planned. One of the service jobs no longer meets constraints. The other service job is successfully scheduled but the resulting plan is rejected because the scheduler has fallen behind in raft replication. A total of 6 Evaluations are created.

flowchart TD

    event((Node\nmisses\nheartbeat))

    system([system\nnode-update])
    service1([service 1\nnode-update])
    service2([service 2\nnode-update])

    blocked([service 1\nblocked\nqueued-allocs])
    failed([service 2\nfailed\nmax-plan-attempts])
    followup([service 2\nfailed-follow-up])

    %% style classes
    classDef eval fill:#d5f6ea,stroke-width:4px,stroke:#1d9467
    classDef other fill:#d5f6ea,stroke:#1d9467
    class event other;
    class system,service1,service2,blocked,failed,followup eval;

    event --> system
    event --> service1
    event --> service2

    service1 --> blocked

    service2 --> failed
    failed --> followup

Next, consider this example where a service job has been updated. The task group has count = 3 and the following update block:

update {
  max_parallel = 1
  canary       = 1
}

After each Evaluation is processed, the Deployment Watcher will be waiting to receive information on updated Allocation health. Then it will emit a new Evaluation for the next step. A total of 4 Evaluations are created.

flowchart TD

    registerEvent((Job\nRegister))
    alloc1health((Canary\nHealthy))
    alloc2health((Alloc 2\nHealthy))
    alloc3health((Alloc 3\nHealthy))

    register([job-register])
    dwPostCanary([deployment-watcher])
    dwPostAlloc2([deployment-watcher])
    dwPostAlloc3([deployment-watcher])

    %% style classes
    classDef eval fill:#d5f6ea,stroke-width:4px,stroke:#1d9467
    classDef other fill:#d5f6ea,stroke:#1d9467
    class registerEvent,alloc1health,alloc2health,alloc3health other
    class register,dwPostCanary,dwPostAlloc2,dwPostAlloc3 eval

    registerEvent --> register
    register --> wait1
    alloc1health --> wait1
    wait1 --> dwPostCanary

    dwPostCanary --> wait2
    alloc2health --> wait2
    wait2 --> dwPostAlloc2

    dwPostAlloc2 --> wait3
    alloc3health --> wait3
    wait3 --> dwPostAlloc3

Lastly, consider this example where 2 nodes each running 5 Allocations that are all for system jobs are "flapping" by missing heartbeats and then re-registering, or frequently changing fingerprints. This diagram will show the results from each node going down once and then coming back up.

flowchart TD

    %% style classes
    classDef eval fill:#d5f6ea,stroke-width:4px,stroke:#1d9467
    classDef other fill:#d5f6ea,stroke:#1d9467

    eventAdown((Node A\nmisses\nheartbeat))
    eventAup((Node A\nheartbeats))
    eventBdown((Node B\nmisses\nheartbeat))
    eventBup((Node B\nheartbeats))

    eventAdown --> eventAup
    eventBdown --> eventBup

    A01down([job 1 node A\nnode-update])
    A02down([job 2 node A\nnode-update])
    A03down([job 3 node A\nnode-update])
    A04down([job 4 node A\nnode-update])
    A05down([job 5 node A\nnode-update])

    B01down([job 1 node B\nnode-update])
    B02down([job 2 node B\nnode-update])
    B03down([job 3 node B\nnode-update])
    B04down([job 4 node B\nnode-update])
    B05down([job 5 node B\nnode-update])

    A01up([job 1 node A\nnode-update])
    A02up([job 2 node A\nnode-update])
    A03up([job 3 node A\nnode-update])
    A04up([job 4 node A\nnode-update])
    A05up([job 5 node A\nnode-update])

    B01up([job 1 node B\nnode-update])
    B02up([job 2 node B\nnode-update])
    B03up([job 3 node B\nnode-update])
    B04up([job 4 node B\nnode-update])
    B05up([job 5 node B\nnode-update])

    eventAdown:::other --> A01down:::eval
    eventAdown:::other --> A02down:::eval
    eventAdown:::other --> A03down:::eval
    eventAdown:::other --> A04down:::eval
    eventAdown:::other --> A05down:::eval

    eventAup:::other --> A01up:::eval
    eventAup:::other --> A02up:::eval
    eventAup:::other --> A03up:::eval
    eventAup:::other --> A04up:::eval
    eventAup:::other --> A05up:::eval

    eventBdown:::other --> B01down:::eval
    eventBdown:::other --> B02down:::eval
    eventBdown:::other --> B03down:::eval
    eventBdown:::other --> B04down:::eval
    eventBdown:::other --> B05down:::eval

    eventBup:::other --> B01up:::eval
    eventBup:::other --> B02up:::eval
    eventBup:::other --> B03up:::eval
    eventBup:::other --> B04up:::eval
    eventBup:::other --> B05up:::eval

You can extrapolate this example to large clusters: 100 nodes each running 10 system jobs and 40 service jobs that go down once and come back up will result in 100 * 40 * 2 == 8000 Evaluations created for the service jobs, which will result in rescheduling of service allocations to new nodes. For the system jobs, 2000 Evaluations will be created and all of these will be no-op Evaluations that will still need to be replicated to all raft peers, canceled by the scheduler, and eventually need to be garbage collected.