K8s

SUMO Kubernetes Support Guide¶

Links¶

High level:

Tech details:

K8s commands¶

Most of the examples use sumo-prod as an example namespace. SUMO dev/stage/prod run in the sumo-dev/sumo-stage/sumo-prod namespaces respectively.

General¶

Most examples are using the kubectl get ... subcommand. If you'd prefer output that's more readable, you can substitute the get subcommand with describe:

kubectl -n sumo-prod describe pod sumo-prod-web-76b74db69-dvxbh

Listing resources is easier with the get subcommand.

To see all SUMO pods currently running:

kubectl -n sumo-prod get pods

To see all pods running and the K8s nodes they are assigned to:

kubectl -n sumo-prod get pods -o wide

To show yaml for a single pod:

kubectl -n sumo-prod get pod sumo-prod-web-76b74db69-dvxbh -o yaml

To show all deployments:

 kubectl -n sumo-prod get deployments

NAME               DESIRED   CURRENT   UP-TO-DATE   AVAILABLE   AGE
sumo-prod-celery   3         3         3            3           330d
sumo-prod-cron     0         0         0            0           330d
sumo-prod-web      50        50        50           50          331d

To show yaml for a single deployment:

 kubectl -n sumo-prod get deployment sumo-prod-web -o yaml

Run a bash shell on a SUMO pod:

kubectl -n sumo-prod exec -it sumo-prod-web-76b74db69-xbfgj bash

Scaling a deployment:

kubectl -n sumo-prod scale --replicas=60 deployment/sumo-prod-web

Check rolling update status:

kubectl -n sumo-prod rollout status deployment/sumo-prod-web

Working with K8s command output¶

Filtering pods based on a label:

kubectl -n sumo-prod -l type=web get pods

Getting a list of pods:

kubectl -n sumo-prod -l type=web get pods | tail -n +2 | cut -d" " -f 1

Structured output:

See the jsonpath guide here

kubectl -n sumo-prod get pods -o=jsonpath='{.items[0].metadata.name}'

Processing K8s command json output with jq:

jsonpath may be more portable

kubectl -n sumo-prod get pods -o json | jq -r .items[].metadata.name

K8s Services¶

List SUMO services:

kubectl -n sumo-prod get services
NAME            TYPE       CLUSTER-IP      EXTERNAL-IP   PORT(S)         AGE
sumo-nodeport   NodePort   100.71.222.28   <none>        443:30139/TCP   341d

Secrets¶

K8s secrets docs

Secret values are base64 encoded when viewed in K8s output. Once setup as an environment variable or mounted file in a pod, the values are base64 decoded automatically.

Kitsune uses secrets specified as environment variables in a deployment spec:

example

To list secrets:

kubectl -n sumo-prod get secrets

To view a secret w/ base64-encoded values:

kubectl -n sumo-prod get secret sumo-secrets-prod -o yaml

To view a secret with decoded values (aka "human readable"):

This example uses the ksv utility

kubectl -n sumo-prod get secret sumo-secrets-prod -o yaml | ksv

To encode a secret value:

echo -n "somevalue" | base64

The -n flag strips the newline before base64 encoding. Values must be specified without newlines, the base64 command on Linux can take a -w 0 parameter that outputs without newlines. The base64 command in Macos Sierra seems to output encoded values without newlines.

Updating secrets:

kubectl -n sumo-prod apply -f ./some-secret.yaml

Monitoring¶

New Relic¶

Primary region
- sumo-prod-oregon
Failover region
- sumo-prod-frankfurt

Papertrail¶

All pod output is logged to Papertrail.

elastic.co¶

Our hosted Elasticsearch cluster is in the us-west-2 region of AWS. Elastic.co hosting status can be found on this page.

Operations¶

Cronjobs¶

The sumo-prod-cron deployment is a self-contained Python cron system that runs in both Primary and Failover clusters.

 # Oregon
kubectl -n sumo-prod get deployments
NAME               DESIRED   CURRENT   UP-TO-DATE   AVAILABLE   AGE
sumo-prod-celery   3         3         3            3           330d
sumo-prod-cron     1         1         1            1           330d
sumo-prod-web      25        25        25           25          331d

Manually adding/removing K8s Oregon/Frankfurt cluster nodes¶

If you are modifying the Frankfurt cluster, replace instances of oregon-* below with frankfurt.

login to the AWS console
ensure you are in the Oregon region
search for and select the EC2 service in the AWS console
select Auto Scaling Groups from the navigation on the left side of the page
click on the nodes.k8s.us-west-2a.sumo.mozit.cloud or nodes.k8s.us-west-2b.sumo.mozit.cloud row to select it
from the Actions menu (close to the top of the page), click Edit
the Details tab for the ASG should appear, set the appropriate Min, Desired and Max values.
1. it's probably good to set Min and Desired to the same value in case the cluster autoscaler decides to scale down the cluster smaller than the Min.
click Save
if you click on Instances from the navigation on the left side of the page, you can see the new instances that are starting/stopping.
you can see when the nodes join the K8s cluster with the following command:

watch 'kubectl get nodes | tail -n +2 | grep -v main | wc -l'

The number that is displayed should eventually match your ASG Desired value. Note this value only includes K8s workers.

Manually Blocking an IP address¶

login to the AWS console
ensure you are in the Oregon region
search for and select the VPC service in the AWS console
select Network ACLs from the navigation on the left side of the page
select the row containing the Oregon VPC
click on the Inbound Rules tab
click Edit
click Add another rule
for Rule#, select a value < 100 and > 0
for Type, select All Traffic
for Source, enter the IP address in CIDR format. To block a single IP, append /32 to the IP address.
1. example: 196.52.2.54/32
for Allow / Deny, select DENY
click Save

There are limits that apply to using VPC ACLs documented here.

Manually Initiating Cluster failover¶

Note: Route 53 will provide automated cluster failover, these docs cover things to consider if there is a catastrophic failure in Oregon and Frankfurt must be promoted to primary rather than a read-only failover.

verify the Frankfurt read replica
- eu-central-1 (Frankfurt) has a read-replica of the SUMO production database
- the replica is currently a db.m4.xlarge, while the prod DB is db.m4.4xlarge
  - this may be ok in maintenance mode, but if you are going to enable write traffic, the instance type must be scaled up.
    - SRE's performed a manual instance type change on the Frankfurt read-replica, and it took ~10 minutes to change from a db.t2.medium to a db.m4.xlarge.
- although we have alerting in place to notify the SRE team in the event of a replication error, it's a good idea to check the replication status on the RDS details page for the sumo MySQL instance.
  - specifically, check the DB Instance Status, Read Replica Source, Replication State, and Replication Error values.
- decide if promoting the read-replica to a main is appropriate.
  - it's preferrable to have a multi-AZ RDS instance, as we can take snapshots against the failover instance (RDS does this by default in a multi-AZ setup).
  - if data is written to a promoted instance, and failover back to the us-west-2 clusters is desirable, a full DB backup and restore in us-west-2 is required.
  - the replica is automatically rebooted before being promoted to a full instance.
ensure image versions are up to date
Most MySQL changes should already be replicated to the read-replica, however, if you're reading this, chances are things are broken. Ensure that the DB schema is correct for the iamges you're deploying.
scale cluster and pods
- the prod deployments A yaml contain the correct number of replicas, but here are some safe values to use in an emergency:
DNS
- point the prod-tp.sumo.mozit.cloud traffic policy at the Frankfurt ELB