konsumerator

Konsumerator

Konsumerator is a Kubernetes operator intended to automate management and resource allocations for kafka consumers.

This software is in active development phase, any logic and CRD structure could change between versions. Use on your own risk.

Operator creates and manages Consumer CRD.

apiVersion: konsumerator.lwolf.org/v1alpha1
kind: Consumer
metadata:
  name: consumer-sample
spec:
  numPartitions: 100
  name: "test-consumer"
  namespace: "default"
  autoscaler:
    # only one provider is supported yet - prometheus.
    # `prometheus` is configured using prometheus provider and user specific metrics
    mode: "prometheus"
    prometheus:
      # minimum allowed period to query prometheus for the lag
      # information to avoid DDoS of that service
      minSyncPeriod: "1m"
      # do not scale up if the lag is less than 5 minutes
      tolerableLag: "5m"
      # approximate consumption rate per CPU
      ratePerCore: 20000
      # approximate memory requirements per CPU
      # if ratePerCore is 10k ops, this value is amount of
      # RAM needed during the processing of this 10k ops
      ramPerCore: "100M"
      # criticalLag is some value close to the SLO.
      # if lag has reached this point, autoscaler will
      # give maximum allowed resource to that deployment
      criticalLag: "60m"
      # preferable recovery time. During lag, Consumer will try to allocate 
      # as much resources as possible to recover from lag during this period
      preferableCatchupPeriod: "30m"
      # prometheus addresses to query
      address:
        - "http://prometheus-operator-prometheus.monitoring.svc.cluster.local:9090"
      # Offset query should return number of messages that is not
      # processed yet a.k.a lag per partitionLabel
      offset:
        query: "sum(rate(kafka_messages_last_offset{topic=''}[5m])) by (partition)"
        partitionLabel: "partition"
      # Production query should return number of messages is being
      # produced per partitionLabel per unit of time (second)
      production:
        query: "sum(rate(kafka_messages_produced_total{topic=''}[5m])) by (partition)"
        partitionLabel: "partition"
      # Consumption query should return number of messages is being
      # consumed per partitionLabel per unit of time (second)
      consumption:
        query: "sum(rate(kafka_messages_consumed_total{topic=''}[5m])) by (partition)"
        partitionLabel: "partition"
  # partitionEnvKey - the name of the environment variable
  # containing partition number the deployment is responsible for
  partitionEnvKey: "PARTITION"
  # DeploymentSpec to run the consumer
  deploymentTemplate:
    replicas: 1
    strategy:
      type: Recreate
    selector:
      matchLabels:
        app: my-dep
    template:
      metadata:
        labels:
          app: my-dep
      spec:
        containers:
          - image: busybox
            name: busybox-sidecar
            command: ["/bin/sh", "-ec", "env && sleep 3000"]
          - image: busybox
            name: busybox
            command: ["/bin/sh", "-ec", "sleep 2000"]
  # resource boundaries, this policy protects
  # the consumer from scaling to 0 or infinity in case
  # of incidents
  resourcePolicy:
    containerPolicies:
    - containerName: busybox
      minAllowed:
        cpu: "100m"
        memory: "100M"
      maxAllowed:
        cpu: "1"
        memory: "1G"
    - containerName: busybox-sidecar
      minAllowed:
        cpu: "100m"
        memory: "100M"
      maxAllowed:
        cpu: "100m"
        memory: "100M"

When such Consumer are being created, operator will create .spec.numPartitions unique deployments.

$ kubectl get consumers

  NAME                 REPLICAS   AUTOSCALER    LAGGING AVAILABLE   AGE
  consumer-sample      100        Prometheus    6       100         7s

Deployment specifics:

• Deployment will be created based on the template provided in .spec.deploymentTemplate without any modifications. It does not make sense to set resource field, since it will be overridden by autoscaler.
• Deployment will be named {consumerName}-{index} where consumerName is .spec.name and index is in range from 0 to .spec.numPartitions.
• Resource requests/limits for each deployment will be estimated based on metrics and configuration
• Each container in the deployment will get few environment variables set: KONSUMERATOR_PARTITION GOMAXPROCS. First one contains kafka partition number assigned to this deployment. Name of this variable is configurable. Second is golang specific setting, always equals to the resources.limit.cpu.

Metrics Providers

At the moment only one metrics provider is implemented - Prometheus.

Here is a critical settings for the provider:

spec:
  ...
  autoscaler:
    prometheus:
      # minimum allowed period to query prometheus for the lag
      # information to avoid DDoS of that service
      minSyncPeriod: "1m"
      address:
        - "http://prometheus-operator-prometheus.monitoring.svc.cluster.local:9090"
      # Offset query should return number of messages that is not
      # processed yet a.k.a lag per partitionLabel
      offset:
        query: "sum(rate(kafka_messages_last_offset{topic=''}[5m])) by (partition)"
        partitionLabel: "partition"
      # Production query should return number of messages is being
      # produced per partitionLabel per unit of time (second)
      production:
        query: "sum(rate(kafka_messages_produced_total{topic=''}[5m])) by (partition)"
        partitionLabel: "partition"
      # Consumption query should return number of messages is being
      # consumed per partitionLabel per unit of time (second)
      consumption:
        query: "sum(rate(kafka_messages_consumed_total{topic=''}[5m])) by (partition)"
        partitionLabel: "partition"

Resource Predictors

At the moment only one resource predictor is implemented which is tightly coupled with Prometheus metrics provider. NaivePredictor operates using following settings provided by a user in .spec.autoscaler.prometheus

• ratePerCore- approximate number of message that could be processed by a single core during normal operations
• ramPerCore - approximate amount of RAM required for the amount of messages process by a core. Sometimes you don't have such dependency between cpu and memory and want to allocate fixed amount of memory, in this case, put here any value and set resourcePolicy for the container with minAllowed.memory = maxAllowed.memory.
• preferableCatchupPeriod - if there is a lag of the partition, predictor will try to give consumer as much resources as possible to catchup during this period. It also require production and offset metrics from the Prometheus.

Installation

coming soon...

Operations

How to stop/start consumer

At the moment, the only way to stop the consumer is to set .spec.deploymentTemplate.replicas to 0. This will trigger reconciliation, operator will notice that the deployment spec was changed and apply the change, in this case change the number of replicas to 0.

How to pause auto scaling

It is possible to temporary disable autoscaling of the managed deployments. To do so, add the following annotation to the consumer object with any value:

annotations:
  konsumerator.lwolf.org/disable-autoscaler: "true"

In case of disabled autoscaling, operator will allocate minimum resources from the resourcePolicy for this container. If there is no such policy set, no resource request/limit will be set.

Development

make build

make test

make install

make run