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

13 KiB

Architecture: Evaluations into Allocations

The Scheduling Concepts docs provide an overview of how the scheduling process works. This document is intended to go a bit deeper than those docs and walk thru the lifecycle of an Evaluation from job registration to running Allocations on the clients. The process can be broken into 4 parts:

Note that in all the diagrams below, writing to the State Store is considered atomic. This means the Nomad leader has replicated to all the followers, and all the servers have applied the raft log entry to their local FSM. So long as raftApply returns without an error, we have a guarantee that all servers will be able to retrieve the entry from their state store at some point in the future.

Job Registration

Creating or updating a Job is a synchronous API operation. By the time the response has returned to the API consumer, the Job and an Evaluation for that Job have been written to the state store of the server nodes.

  • Note: parameterized or periodic batch jobs don't create an Evaluation at registration, only once dispatched.
  • Note: The scheduler is very fast! This means that once the CLI gets a response it can immediately start making queries to get information about the next steps.
  • Note: This workflow is different for Multi-Region Deployments (found in Nomad Enterprise). That will be documented separately.

The diagram below shows this initial synchronous phase. Note here that there's no scheduling happening yet, so no Allocation (or Deployment) has been created.

sequenceDiagram

    participant user as User
    participant cli as CLI
    participant httpAPI as HTTP API
    participant leaderRpc as Leader RPC
    participant stateStore as State Store

    user ->> cli: nomad job run
    activate cli
    cli ->> httpAPI: Create Job API
    httpAPI ->> leaderRpc: Job.Register

    activate leaderRpc

    leaderRpc ->> stateStore: write Job
    activate stateStore
    stateStore -->> leaderRpc: ok
    deactivate stateStore

    leaderRpc ->> stateStore: write Evaluation

    activate stateStore
    stateStore -->> leaderRpc: ok

    Note right of stateStore: EvalBroker.Enqueue
    Note right of stateStore: (see Scheduling below)

    deactivate stateStore

    leaderRpc -->> httpAPI: Job.Register response
    deactivate leaderRpc

    httpAPI -->> cli: Create Job response
    deactivate cli

Scheduling

A long-lived goroutine on the Nomad leader called the Eval Broker maintains a queue of Evaluations previously written to the state store and enqueued via the EvalBroker.Enqueue method. (When a leader transition occurs, the leader queries all the Evaluations in the state store and enqueues them in its new Eval Broker.)

Scheduler workers are long-lived goroutines running on all server nodes. Typically this will be one per core on followers and 1/4 that number on the leader. The workers poll for Evaluations from the Eval Broker with the Eval.Dequeue RPC. Once a worker has an evaluation, it instantiates a scheduler for that evaluation (of type service, system, sysbatch, batch, or core).

Because a worker is running one scheduler at a time, Nomad's documentation often refers to "workers" and "schedulers" interchangeably, but the worker is the long-lived goroutine and the scheduler is the struct that contains the code and state around processing a single evaluation. The scheduler mutates itself and is thrown away once the evaluation is processed.

The scheduler takes a snapshot of that server node's state store so that it has a constant current view of the cluster state. The scheduler executes 3 main steps:

  • Reconcile: compare the cluster state and job specification to determine what changes need to be made -- starting or stopping Allocations. The scheduler creates new Allocations in this step. But note that these Allocations are not yet persisted to the state store.
  • Feasibility Check: For each Allocation it needs, the scheduler iterates over Nodes until it finds up to 2 Nodes that match the Allocation's resource requirements and constraints.
    • Note: for system jobs or jobs with with spread blocks, the scheduler has to check all Nodes.
  • Scoring: For each feasible node, the scheduler ranks them and picks the one with the highest score.

If the job is a service job, the scheduler will also create (or update) a Deployment. When the scheduler determines the number of Allocations to create, it examine the job's update block. Only the number of Allocations needed to complete the next phase of the update will be created in a given allocation. The Deployment is used by the deployment watcher (see below) to monitor the health of allocations and create new Evaluations to continue the update.

If the scheduler cannot place all Allocations, it will create a new Evaluation in the blocked state and submit it to the leader. The Eval Broker will re-enque that Evaluation once cluster state has changed. (This process is the first green box in the sequence diagram below.)

Once the scheduler has completed processing the Evaluation, if there are Allocations (and possibly a Deployment) to update, it will submit this work as a Plan to the leader. The leader needs to validate this plan and serialize it:

  • The scheduler took a snapshot of cluster state at the start of its work, so that state may have changed in the meantime.
  • Schedulers run concurrently across the cluster, so they may generate conflicting plans (particularly on heavily-packed clusters).

The leader processes the plan in the plan applier. If the plan is valid, the plan applier will write the Allocations (and Deployment) update to the state store. If not, it will reject the plan and the scheduler will try to create a new plan with a refreshed state. If the scheduler fails to submit a valid plan too many times it submits a blocked Evaluation that is triggered by max-plan-attempts type. (The plan submit process is the second green box in the sequence diagram below.)

Once the scheduler has a response from the leader, it will tell the Eval Broker to Ack the Evaluation (if it successfully submitted the plan) or Nack the Evaluation (if it failed to do so) so that another scheduler can try processing it.

The diagram below shows the scheduling phase, including submitting plans to the planner. Note that at the end of this phase, Allocations (and Deployment) have been persisted to the state store.

sequenceDiagram

    participant leaderRpc as Leader RPC
    participant stateStore as State Store
    participant planner as Planner
    participant broker as Eval Broker
    participant sched as Scheduler

    leaderRpc ->> stateStore: write Evaluation (see above)
    stateStore ->> broker: EvalBroker.Enqueue
    activate broker
    broker -->> broker: enqueue eval
    broker -->> stateStore: ok
    deactivate broker

    sched ->> broker: Eval.Dequeue (blocks until work available)
    activate sched
    broker -->> sched: Evaluation

    sched ->> stateStore: query to get job and cluster state
    stateStore -->> sched: results

    sched -->> sched: Reconcile (how many allocs are needed?)
    deactivate sched

    alt
        Note right of sched: Needs new allocs
        activate sched
        sched -->> sched: create Allocations
        sched -->> sched: create Deployment (service jobs)
        Note left of sched: for Deployments, only 1 "batch" of Allocs will get created for each Eval
        sched -->> sched: Feasibility Check
        Note left of sched: iterate over Nodes to find a placement for each Alloc
        sched -->> sched: Ranking
        Note left of sched: pick best option for each Alloc

        %% start rect highlight for blocked eval
        rect rgb(213, 246, 234)
            Note left of sched: Not enough room! (But we can submit a partial plan)
            sched ->> leaderRpc: Eval.Upsert (blocked)
            activate leaderRpc
            leaderRpc ->> stateStore: write Evaluation
            activate stateStore
            stateStore -->> leaderRpc: ok
            deactivate stateStore
            leaderRpc --> sched: ok
            deactivate leaderRpc
        end
        %% end rect highlight for blocked eval


        %% start rect highlight for planner
        rect rgb(213, 246, 234)
            Note over sched, planner: Note: scheduler snapshot may be stale state relative to leader so it's serialized by plan applier

            sched ->> planner: Plan.Submit (Allocations + Deployment)
            planner -->> planner: is plan still valid?

            Note right of planner: Plan is valid
            activate planner

            planner ->> leaderRpc: Allocations.Upsert + Deployment.Upsert
            activate leaderRpc
            leaderRpc ->> stateStore: write Allocations and Deployment
            activate stateStore
            stateStore -->> leaderRpc: ok
            deactivate stateStore
            leaderRpc -->> planner: ok
            deactivate leaderRpc
            planner -->> sched: ok
            deactivate planner

        end
        %% end rect highlight for planner

        sched -->> sched: retry on failure, if exceed max attempts will Eval.Nack

    else

    end

    sched ->> broker: Eval.Ack (Eval.Nack if failed)
    activate broker
    broker ->> stateStore: complete Evaluation
    activate stateStore
    stateStore -->> broker: ok
    deactivate stateStore
    broker -->> sched: ok
    deactivate broker
    deactivate sched

Deployment Watcher

As noted under Scheduling above, a Deployment is created for service jobs. A deployment watcher runs on the leader. Its job is to watch the state of Allocations being placed for a given job version and to emit new Evaluations so that more Allocations for that job can be created.

The "deployments watcher" (plural) makes a blocking query for Deployments and spins up a new "deployment watcher" (singular) for each one. That goroutine will live until its Deployment is complete or failed.

The deployment watcher makes blocking queries on Allocation health and its own Deployment (which can be canceled or paused by a user). When there's a change in any of those states, it compares the current state against the update block and the timers it maintains for min_healthy_time, healthy_deadline, and progress_deadline. It then updates the Deployment state and creates a new Evaluation if the current step of the update is complete.

The diagram below shows deployments from a high level. Note that Deployments do not themselves create Allocations -- they create Evaluations and then the schedulers process those as they do normally.

sequenceDiagram

    participant leaderRpc as Leader RPC
    participant stateStore as State Store
    participant dw as Deployment Watcher

    dw ->> stateStore: blocking query for new Deployments
    activate dw
    stateStore -->> dw: new Deployment
    dw -->> dw: start watcher

    dw ->> stateStore: blocking query for Allocation health
    stateStore -->> dw: Allocation health updates
    dw ->> dw: next step?

    Note right of dw: Update state and create evaluations for next batch...
    Note right of dw: Or fail the Deployment and update state

    dw ->> leaderRpc: Deployment.Upsert + Evaluation.Upsert
    activate leaderRpc
    leaderRpc ->> stateStore: write Deployment and Evaluations
    activate stateStore
    stateStore -->> leaderRpc: ok
    deactivate stateStore
    leaderRpc -->> dw: ok
    deactivate leaderRpc
    deactivate dw

Client Allocs

Once the plan applier has persisted Allocations to the state store (with an associated Node ID), they become available to get placed on the client. Clients pull new allocations (and changes to allocations), so a new allocation will be in the pending state until it's been pulled down by a Client and the allocation has been instantiated.

Once the Allocation is running and healthy, the Client will send a Node.UpdateAlloc RPC back to the server so that info can be persisted in the state store. This is the allocation health data the Deployment Watcher is querying for above.

sequenceDiagram

    participant followerRpc as Follower RPC
    participant stateStore as State Store

    participant client as Client
    participant allocrunner as Allocation Runner

    client ->> followerRpc: Alloc.GetAllocs RPC
    activate client

    Note right of client: this query can be stale

    followerRpc ->> stateStore: query for Allocations
    activate followerRpc
    activate stateStore
    stateStore -->> followerRpc: Allocations
    deactivate stateStore
    followerRpc -->> client: Allocations
    deactivate followerRpc

    client ->> allocrunner: Create or update allocation runners

    client ->> followerRpc: Node.UpdateAlloc
    Note right of followerRpc: will be forwarded to leader

    deactivate client