metadata.md

Metadata Collection

How To:

• Metadata Collection
  • How it Works
  • Targeted Init Containers
  • DaemonSet
  • Differences between targeted and daemonset
• Enable Collection
  • Additional Options
    • Customizing backpack image
    • Privileged Pods
    • Additional k8s Information
    • Stockpile tags
    • Redis integration
  • Running Additional Collections

How it Works

The metadata collection is done through the use of Stockpile, Backpack and Scribe. The data is then uploaded to a defined Elasticsearch instance.

Metadata collection is enabled by default, it can be disabled setting collection to false in the metadata section:

metadata:
  collection: false

There are two ways to launch metadata collection from a workload.

Targeted Init Containers

The first, and default behavior, is through init containers. These are defined in the workload template with an init container section that looks like:

{% if metadata.collection is sameas true and metadata.targeted is sameas true %}
      initContainers:
      - name: backpack-{{ trunc_uuid }}
        image: {{ metadata.image }}
        command: ["/bin/sh", "-c"]
        args:
          - >
            python3
            stockpile-wrapper.py
            -s={{ elasticsearch.url }}
            -u={{ uuid }}
            -n=${my_node_name}
            -N=${my_pod_name}
            --redisip={{ bo.resources[0].status.podIP }}
            --redisport=6379
{% if metadata.force is sameas true %}
            --force
{% endif %}
{% if metadata.stockpileTags|length > 0 %}
            --tags={{ metadata.stockpileTags|join(",") }}
{% endif %}
{% if metadata.stockpileSkipTags|length > 0 %}
            --skip-tags={{ metadata.stockpileSkipTags|join(",") }}
{% endif %}
{% if metadata.ssl is sameas true %}
            --sslskipverify
{% endif %}
        imagePullPolicy: Always
        securityContext:
          privileged: {{ metadata.privileged }}
        env:
          - name: my_node_name
            valueFrom:
              fieldRef:
                fieldPath: spec.nodeName
          - name: my_pod_name
            valueFrom:
              fieldRef:
                fieldPath: metadata.name
{% endif %}

This allows the targeted collection of metadata on nodes that only have workloads on them. This avoids trying to collect data from nodes we are not touching. The backpack container will run stockpile to gather data and then push that information into Elasticsearch. Once complete the init container will terminate and the workload will launch and continue as normal.

DaemonSet

If it is desired to run the metadata collection in the "classic" way (i.e. as a DaemonSet), then you will need to set metadata.targeted: false in your cr file.

When configured in this way, backpack will launch as a DaemonSet on all nodes in the cluster when the workload is applied. Backpack will then run stockpile to gather data and then push that information up to Elasticsearch exactly like the targeted/init container option.

It runs the data collection/upload immediately upon launch and will not enter "Ready" state until the initial data collection/upload has completed. Once this is done, your benchmark will then launch and continue as normal.

The DaemonSet option is able to take another optional parameters that the targeted variant cannot. You are able to supply a list of label/value pairs in label. This will have the DaemonSet only run on nodes that match any of the labels provided.

In the below example the metadata DaemonSet will only run on nodes labeled with foo=bar OR awesome=sauce. It can match either provided label.

apiVersion: ripsaw.cloudbulldozer.io/v1alpha1
kind: Benchmark
metadata:
  name: byowl-benchmark
  namespace: benchmark-operator
spec:
  elasticsearch:
    url: "http://my.elastic.server.foo:9200"
  metadata:
    collection: true
    targeted: false
    label:
      - [ 'foo', 'bar' ]
      - [ 'awesome', 'sauce' ]
  workload:
    name: byowl
    args:
      image: "quay.io/jtaleric/uperf:testing"
      clients: 1
      commands: "echo Test"

Differences between targeted and daemonset

There are a few notable differences between the DaemonSet and targeted options:

• First, and most important, the DaemonSet will run on ALL nodes of the cluster. This means that if you have 200 nodes it will collect data from all the nodes even if your relevant pod(s) are only running on a subset of nodes. You will end up with more data than you need and slow down the start of your workload.

• If running targeted and with a service account (more on that below) the entire workload will run as that service account. That is because service accounts are done at a level above the container definition and can only be applied once.

• When run as a daemonset the backpack pods will not complete/terminate until you delete your benchmark. This is done to allow additional collections to be done as an ad-hoc basis as well.

• When running as a targeted init container the benchmark metadata status will be set to "Collecting" however it will never be marked as completed as that would require additional logic in the workloads which is out of scope.

• Finally, running as a daemonset requires no additions to the workload template. This may be preferable in certain situations.

Enable Collection

By default metadata collection is turned off. To enable collection:

• Open the benchmark yaml that you will be running (for example byowl below)

apiVersion: ripsaw.cloudbulldozer.io/v1alpha1
kind: Benchmark
metadata:
  name: byowl-benchmark
  namespace: benchmark-operator
spec:
  workload:
    name: byowl
    args:
      image: "quay.io/jtaleric/uperf:testing"
      clients: 1
      commands: "echo Test"

• Add metadata.collection: true to the spec section of the yaml

apiVersion: ripsaw.cloudbulldozer.io/v1alpha1
kind: Benchmark
metadata:
  name: byowl-benchmark
  namespace: benchmark-operator
spec:
  metadata:
    collection: true
  workload:
    name: byowl
    args:
      image: "quay.io/jtaleric/uperf:testing"
      clients: 1
      commands: "echo Test"

The metadata collection will now be enabled however as there is no Elasticsearch information defined it will fail.

• Add the Elasticsearch server information

apiVersion: ripsaw.cloudbulldozer.io/v1alpha1
kind: Benchmark
metadata:
  name: byowl-benchmark
  namespace: benchmark-operator
spec:
  elasticsearch:
    url: "http://my.elastic.server.foo:9200"
  metadata:
    collection: true
  workload:
    name: byowl
    args:
      image: "quay.io/jtaleric/uperf:testing"
      clients: 1
      commands: "echo Test"

The metadata collection will now run as defined.

Additional Options

There are a few additional options that can be enabled to enhance the amount of data collected as well as reduce redundancy.

Customizing backpack image

By default backpack uses quay.io/cloud-bulldozer/backpack:latest, however it's possible to customize this image setting the image parameter from the metadata section:

metadata:
  image: quay.io/myorg/custom-backpack:latest

Privileged Pods

By default, pods are run in an unprivledged state. While this makes permissions a lesser issue, it does limit the amount of data collected. For example, dmidecode requires privileges to read the memory information and generate the data.

To enable privileged pods set:

metadata:
  privileged: true

In the spec section of the benchmark yaml as outlined for the other variables.

Elasticsearch SSL Verification

By default elasticsearch SSL verification is disabled. To enable it, set ssl to true in the metadata section.

To enable ssl verification:

metadata:
  ssl: true

Stockpile tags

Backpack leverages stockpile to collect metadata. Stockpile is basically a set of Ansible roles, these roles have different tags. It's possible to pass custom tags to the inner ansible-playbook command through the parameters stockpileSkipTags and stockpileTags from the metadata section. These parameters are translated to the Ansible's flags --tags and --skip-tags respectively. By default stockpileTags has the value ["common", "k8s", "openshift"].

An example to only collect metadata from the roles tagged with memory and cpu would be:

metadata:
  stockpileTags: ["memory", "cpu"]

An example to skip these tags would be:

metadata:
  stockpileSkipTags: ["memory", "cpu"]

Redis integration

When the stockpile-wrapper.py script is passed --redisip [ip of redis] and --redisport [redis port] it will attempt to check Redis to see if the host has had its metadata collected for the current uuid. If it has already been collected we will not run the collection again. This will reduce some redundant data that was being collected when pods were launched on the same node.

If it is desired to run the metadata collection regardless of if Redis claims it was already run then passing the script the --force option will force the metadata collection to occur.

NOTE As Redis integration is already configured for the existing workloads nothing needs to be done to enable it. However, if you wish to use the --force option you will need to add force: true to the metadata section of your workload cr file.

metadata:
  force: true

Running Additional Collections

NOTE This is only applicable to the DaemonSet collection method

While upon initial creation metadata is collected, it may be useful to collect additional runs of data at other times. To do this you will need to loop through each backpack Pod and exec the python command below:

python3 stockpile-wrapper.py -s [es_url] -u [uuid]

Where es_url points to the Elasticsearch server to index to. The UUID can be any uuid string you would like (if you do not supply one it will create one for you and you will see it defined in the output). On the initial run at boot this is the same UUID as the benchmark UUID.