How We Analyze Kubernetes Occasions In Actual Time


Kubernetes at Rockset

At Rockset, we use Kubernetes (k8s) for cluster orchestration. It runs all our manufacturing microservices — from our ingest employees to our query-serving tier. Along with internet hosting all of the manufacturing infrastructure, every engineer has their very own Kubernetes namespace and devoted assets that we use to regionally deploy and check new variations of code and configuration. This sandboxed atmosphere for growth permits us to make software program releases confidently a number of occasions each week. On this weblog submit, we are going to discover a software we constructed internally that offers us visibility into Kubernetes occasions, a wonderful supply of details about the state of the system, which we discover helpful in troubleshooting the system and understanding its long-term well being.

Why We Care About Kubernetes Occasions

Kubernetes emits occasions at any time when some change happens in any of the assets that it’s managing. These occasions usually include essential metadata in regards to the entity that triggered it, the kind of occasion (Regular, Warning, Error, and many others.) and the trigger. This information is often saved in etcd and made obtainable if you run sure kubectl instructions.

$ kubectl describe pods jobworker-c5dc75db8-7m5ln
  Kind     Cause     Age                From                                                    Message
  ----     ------     ----               ----                                                    -------
  Regular   Scheduled  7m                 default-scheduler                                       Efficiently assigned grasp/jobworker-c5dc75db8-7m5ln to
  Regular   Pulling    6m                 kubelet,  pulling picture "..."
  Regular   Pulled     6m                 kubelet,  Efficiently pulled picture "..."
  Regular   Created    6m                 kubelet,  Created container
  Regular   Began    6m                 kubelet,  Began container
  Warning  Unhealthy  2m (x2 over 2m)    kubelet,  Readiness probe failed: Get http://XXX.XXX.XXX.XXX:YYY/healthz: dial tcp join: connection refused

These occasions assist perceive what occurred behind the scenes when a specific entity entered a particular state. One other place to see an aggregated checklist of all occasions is by accessing all occasions by way of kubectl get occasions.

$ kubectl get occasions
LAST SEEN   TYPE      REASON                 KIND                      MESSAGE
5m          Regular    Scheduled              Pod                       Efficiently assigned grasp/jobworker-c5dc75db8-7m5ln to
5m          Regular    Pulling                Pod                       pulling picture "..."
4m          Regular    Pulled                 Pod                       Efficiently pulled picture "..."

As could be seen above, this offers us particulars – the entity that emitted the occasion, the kind/severity of the occasion, in addition to what triggered it. This info could be very helpful when seeking to perceive adjustments which might be occurring within the system. One further use of those occasions is to grasp long-term system efficiency and reliability. For instance, sure node and networking errors that trigger pods to restart could not trigger service disruptions in a extremely obtainable setup however usually could be hiding underlying circumstances that place the system at elevated threat.

In a default Kubernetes setup, the occasions are continued into etcd, a key-value retailer. etcd is optimized for fast strongly constant lookups, however falls quick on its means to supply analytical skills over the info. As dimension grows, etcd additionally has hassle maintaining and subsequently, occasions get compacted and cleaned up periodically. By default, solely the previous hour of occasions is preserved by etcd.

The historic context can be utilized to grasp long-term cluster well being, incidents that occurred previously and the actions taken to mitigate them inside Kubernetes, and construct correct submit mortems. Although we checked out different monitoring instruments for occasions, we realized that we had a chance to make use of our personal product to research these occasions in a means that no different monitoring product might, and use it to assemble a visualization of the states of all of our Kubernetes assets.



To ingest the Kubernetes occasions, we use an open supply software by Heptio referred to as eventrouter. It reads occasions from the Kubernetes API server and forwards them to a specified sink. The sink could be something from Amazon S3 to an arbitrary HTTP endpoint. As a way to connect with a Rockset assortment, we determined to construct a Rockset connector for eventrouter to regulate the format of the info uploaded to our assortment. We contributed this Rockset sink into the upstream eventrouter undertaking. This connector is admittedly easy — it takes all acquired occasions and emits them into Rockset. The actually cool half is that for ingesting these occasions, that are JSON payloads that fluctuate throughout various kinds of entities, we don’t must construct any schema or do structural transformations. We will emit the JSON occasion as-is right into a Rockset assortment and question it as if it had been a full SQL desk. Rockset robotically converts JSON occasions into SQL tables by first indexing all of the json fields utilizing Converged Indexing after which robotically schematizing them by way of Good Schemas.

The front-end software is a skinny layer over the SQL layer that permits filtering occasions by namespace and entity sort (Pod, Deployment, and many others.), after which inside these entity varieties, cluster occasions by regular/errors. The aim is to have a histogram of those occasions to visually examine and perceive the state of the cluster over an prolonged time period. After all, what we show is solely a subset of what might be constructed – one can think about rather more advanced analyses – like analyzing community stability, deployment processes, canarying software program releases and even utilizing the occasion retailer as a key diagnostic software to find correlations between cluster-level alerts and Kubernetes-level adjustments.


Earlier than we are able to start receiving occasions from eventrouter into Rockset, we should create a set in Rockset. That is the gathering that every one eventrouter occasions are saved in. You are able to do this with a free account from

A set in Rockset can ingest information from a specified supply, or could be despatched occasions by way of the REST API. We’ll use the latter, so, we create a set that’s backed by this Write API. Within the Rockset console, we are able to create such a set by selecting “Write API” as the info supply.


When creating the gathering, we are able to decide a retention, say, 120 days or any cheap period of time to present us some sense of cluster well being. This retention is utilized based mostly on a particular area in Rockset, _event_time. We are going to map this area to a particular area inside the JSON occasion payload we are going to obtain from eventrouter referred to as occasion.lastTimestamp. The transformation operate seems to be like the next:


After creating the gathering, we are able to now arrange and use eventrouter to start receiving Kubernetes occasions.


Now, receiving occasions from eventrouter requires another factor – a Rockset API key. We will use API keys in Rockset to jot down JSON to a set, and to make queries. On this case, we create an API key referred to as eventrouter_write from Handle > API keys.


Copy the API key as we would require it in our subsequent step establishing eventrouter to ship occasions into the Rockset assortment we simply arrange. You possibly can arrange eventrouter by cloning the eventrouter repository and edit the YAML file yaml/deployment.yaml to seem like the next:

# eventrouter/yaml/deployment.yaml
config.json: |-
"sink": "rockset"
"rocksetServer": "",
"rocksetAPIKey": "<API_KEY>",
"rocksetCollectionName": "eventrouter_events",
"rocksetWorkspaceName": "commons",

You possibly can substitute the <API_KEY> with the Rockset API key we simply created within the earlier step. Now, we’re prepared! Run kubectl apply -f yaml/deployment.yaml, and eventrouter can begin watching and forwarding occasions immediately. Trying on the assortment inside Rockset, you need to begin seeing occasions flowing in and being made obtainable as a SQL desk. We will question it as proven under from the Rockset console and get a way of a number of the occasions flowing in. We will run full SQL over it – together with all varieties of filters, joins, and many others.


Querying Knowledge

We will now begin asking some fascinating questions from our cluster and get an understanding of cluster well being. One query that we needed to ask was – how usually are we deploying new photos into manufacturing. We operated on a strict launch schedule, however there are occasions after we rollout and rollback photos.

With replicasets as (
            e.occasion.motive as motive,
            e.occasion.lastTimestamp as ts,
            e.occasion.metadata.title as title,
              REGEXP_EXTRACT(e.occasion.message, 'Created pod: (.*)', 1) as pod
            commons.eventrouter_events e
        the place
            e.occasion.involvedObject.sort = 'ReplicaSet'
            and e.occasion.metadata.namespace="manufacturing"
              and e.occasion.motive = 'SuccessfulCreate'
    pods as (
            e.occasion.motive as motive,
            e.occasion.message as message,
            e.occasion.lastTimestamp as ts,
            e.occasion.involvedObject.title as title,
                'pulling picture "imagerepo/folder/(.*?)"',
            ) as picture
            commons.eventrouter_events e
        the place
            e.occasion.involvedObject.sort = 'Pod'
            and e.occasion.metadata.namespace="manufacturing"
            and e.occasion.message like '%pulling picture%'
              and e.occasion.involvedObject.title like 'aggregator%'

SELECT * from (
    MAX(p.ts) as ts, MAX(r.pod) as pod, MAX(p.picture) as picture, r.title
    pods p
    JOIN replicasets r on p.title = r.pod
GROUP BY r.title) sq
restrict 100;

The above question offers with our deployments, which in flip create replicasets and finds the final date on which we deployed a specific picture.

| picture                              | title                                   | pod                         | ts                   |
| leafagg: | aggregator-c478b597.15c8811219b0c944   | aggregator-c478b597-z8fln   | 2019-09-28T04:53:05Z |
| leafagg: | aggregator-c478b597.15c881077898d3e0   | aggregator-c478b597-wvbdb   | 2019-09-28T04:52:20Z |
| leafagg: | aggregator-c478b597.15c880742e034671   | aggregator-c478b597-j7jjt   | 2019-09-28T04:41:47Z |
| leafagg: | aggregator-587f77c45c.15c8162d63e918ec | aggregator-587f77c45c-qjkm7 | 2019-09-26T20:14:15Z |
| leafagg: | aggregator-587f77c45c.15c8160fefed6631 | aggregator-587f77c45c-9c47j | 2019-09-26T20:12:08Z |
| leafagg: | aggregator-587f77c45c.15c815f341a24725 | aggregator-587f77c45c-2pg6l | 2019-09-26T20:10:05Z |
| leafagg: | aggregator-58d76b8459.15c77b4c1c32c387 | aggregator-58d76b8459-4gkml | 2019-09-24T20:56:02Z |
| leafagg: | aggregator-58d76b8459.15c77b2ee78d6d43 | aggregator-58d76b8459-jb257 | 2019-09-24T20:53:57Z |
| leafagg: | aggregator-58d76b8459.15c77b131e353ed6 | aggregator-58d76b8459-rgcln | 2019-09-24T20:51:58Z |

This excerpt of photos and pods, with timestamp, tells us rather a lot about the previous few deploys and once they occurred. Plotting this on a chart would inform us about how constant we now have been with our deploys and the way wholesome our deployment practices are.

Now, shifting on to efficiency of the cluster itself, working our personal hand-rolled Kubernetes cluster means we get a variety of management over upgrades and the system setup however it’s value seeing when nodes could have been misplaced/community partitioned inflicting them to be marked as unready. The clustering of such occasions can inform us rather a lot in regards to the stability of the infrastructure.

With nodes as (
          e.occasion.lastTimestamp as ts,
        commons.eventrouter_events e
    the place
        e.occasion.involvedObject.sort = 'Node'
          AND e.occasion.sort="Regular"
          AND e.occasion.motive = 'NodeNotReady'
    ORDER by ts DESC
Restrict 100;

This question offers us the occasions the node standing went NotReady and we are able to attempt to cluster this information utilizing SQL time features to grasp how usually points are occurring over particular buckets of time.

| message                                                                      | title                                                         | motive       | ts                   |
| Node standing is now: NodeNotReady | | NodeNotReady | 2019-09-30T02:13:19Z |
| Node standing is now: NodeNotReady | | NodeNotReady | 2019-09-30T02:13:19Z |
| Node standing is now: NodeNotReady | | NodeNotReady | 2019-09-30T02:14:20Z |
| Node standing is now: NodeNotReady | | NodeNotReady | 2019-09-30T02:13:19Z |
| Node standing is now: NodeNotReady  |  | NodeNotReady | 2019-09-30T00:10:11Z |

We will moreover search for pod and container degree occasions like once they get OOMKilled and correlate that with different occasions occurring within the system. In comparison with a time collection database like prometheus, the ability of SQL lets us write and JOIN various kinds of occasions to attempt to piece collectively various things that occurred round a specific time interval, which can be causal.

For visualizing occasions, we constructed a easy software that makes use of React that we use internally to look by and do some fundamental clustering of Kubernetes occasions and errors occurring in them. We’re releasing this dashboard into open supply and would like to see what the group may use this for. There are two foremost elements to the visualization of Kubernetes occasions. First is a high-level overview of the cluster at a per-resource granularity. This permits us to see a realtime occasion stream from our deployments and pods, and to see at what state each single useful resource in our Kubernetes system is. There’s additionally an choice to filter by namespace – as a result of sure units of providers run in their very own namespace, this permits us to drill down into a particular namespace to have a look at occasions.


If we have an interest within the well being and state of any explicit useful resource, every per-resource abstract is clickable and opens a web page with an in depth overview of the occasion logs of that useful resource, with a graph that reveals the occasions and errors over time to supply a holistic image of how the useful resource is being managed.


The graph on this visualization has adjustable granularity, and the change in time vary permits for viewing the occasions for a given useful resource over any specified interval. Hovering over a particular bar on the stacked bar chart permits us to see the varieties of errors occurring throughout that point interval for useful over-time analytics of what’s occurring to a particular useful resource. The desk of occasions listed under the graph is sorted by occasion time and in addition tells accommodates the identical info because the graph – that’s, a chronological overview of all of the occasions that occurred to this particular k8s useful resource. The graph and desk are useful methods to grasp why a Kubernetes useful resource has been failing previously, and any traits over time which will accompany that failure (for instance, if it coincides with the discharge of a brand new microservice).


At present, we’re utilizing the real-time visualization of occasions to research our personal Kubernetes deployments in each growth and manufacturing. This software and information supply permits us to see our deployments as they’re ongoing with out having to wrangle the kubectl interface to see what’s damaged and why. Moreover, this software is useful to get a retrospective look on previous incidents. For instance – if we spot transient points, we now have the ability to return in time and take a retrospective have a look at transient manufacturing points, discovering patterns of why it might have occurred, and what we are able to do to forestall the incident from occurring once more sooner or later.

The flexibility to entry historic Kubernetes occasion logs at nice granularity is a robust abstraction that gives us at Rockset a greater understanding of the state of our Kubernetes system than kubectl alone would enable us. This distinctive information supply and visualization permits us to watch our deployments and assets, in addition to have a look at points from a historic perspective. We’d love so that you can do this, and contribute to it in case you discover it helpful in your individual environments!



Leave a Reply