title | authors | reviewers | approvers | creation-date | last-updated | |||
---|---|---|---|---|---|---|---|---|
cluster-fleet-evaluation |
|
|
|
2022-08-16 |
2022-08-16 |
OpenShift components sometimes need to introduce a change that could result in highly visible alerts, breakages or upgrade failures on user clusters. In scenarios like this it will be useful for the component owner to assess how many customers would be affected before rolling out the change and based on that feedback decide on the level of investment required to properly inform the customer base, or change implementation details.
Rather than introducing the change, add just the detection logic to the operator for the change and associate it with a ClusterStatusConditionTypes called EvaluationConditionDetected. If the ConditionStatus associated with this type is ConditionTrue and includes the Reason you are expecting, it will indicate that the change would have affected that cluster. However, if the ConditionStatus is false, or it's true but the Reason does not include the Reason you are expecting, the change doesn't affect the cluster. You can join multiple Reason to test multiple scenarios.
Now release the detection logic to master and backport it to a N-1 Z stream release where N is the current release under development. The Z stream backport is useful as it allows the team to gather the data without going through a full release cycle and waiting for customers to upgrade to newer minor versions. However, we still have to wait for them to upgrade to a patch level that includes the new info gathering logic.
How far back the backports should be done is subjective and the teams should aim to backport as far back as possible so as to collect a reasonable amount data that allows them to reach a conclusion. This decision will be largely based on what versions the bulk of the fleet of clusters are running and the nature of the problem such as whether or not the problem domain is likely influenced by cluster age or the version it was installed with. For instance clusters running 4.8 today, 12 months after GA, are disproportionately older and installed on older versions than those which were running 4.8 in the first 90 days after GA.
Once the Z streams have been released the component developers can interrogate the telemetry data to figure out the prevalence of the scenario that the change will introduce. Allow sufficient time for customers to upgrade to the Z stream release. The upgrade data is also available in telemetry.
At this point the team needs to validate that the telemetry data matches their expectations and make a decision if they still plan to move forward with the change. In particular the team needs to measure the level of impact the change has. If the percentage of impact is very low then it might be preferable to contact the affected customers. Alternatively if the percentage of impact is very high, it would be preferable that the team re-evaluate the change and investigate if there is a way to automate the fix process in a seamless manner.
If the component team decides to move forward, they will then be responsible for writing a Knowledge Centered Support (KCS) document providing sufficient information describing the scenario and how to resolve it. Refer to the KCS section for more information.
Once the KCS article/solution is ready and the list of customers are available, the component team can alert the affected customers using the methods described in the Alerting customers reference section. Teams should make every attempt to use all methods of alerting customers (this ensures contact with customers is made), but are free to drop options if they feel using that method may not reach the proper audience or may be too expensive.
Note: Before the team goes down this path it will be useful to discuss the outcome and expectations with their respective OpenShift staff engineer, product manager and PLM contact. PLM contacts are listed under column S in the OCP Team Member Tracking sheet. In particular the team should set their expectations for action they intend to take (abort the change, add automated migration, contact individual customers, etc) for a given impact percentage. This avoids post hoc rationalizing a decision if the data that comes back is unexpected.
To access the telemetry data please follow these instructions. Once you have access, you will be able to perform queries against the data collected from customer clusters.
You can query the telemetry data either using the Thanos or Grafana interfaces. For example, if you have introduced a new ConditionStatus/Reason called FooBar, it can be queried using:
count by (_id, ebs_account) (
cluster_operator_conditions{name="component-name",condition="ClusterFleetEvaluation",reason="FooBar"}
* on (_id) group_left (ebs_account)
topk by (_id) (1, id_version_ebs_account_internal:cluster_subscribed)
)
This will return a list of clusters and the associated customers. The above query is just an example and the team is free to create their own custom queries.
KCS docs are split into solutions and articles. Solutions explain breakages and how to fix them and are transactional in nature. The scenario presented here would typically result in a KCS solution. Articles are targeted towards highlighting features / explanations that are outside the scope of the OpenShift documentation. If you are new to writing a KCS doc, you can reach out to Albert Myles who can connect you with a KCS coach.
Once the component team has collected the telemetry data and the list of affected customers, they can alert them using any combination of the methods listed below. This should only be done if the impact of the change being introduced is very low. For help in facilitating use of any of these methods you can reach out to your PLM contact. The way to choose method(s) are subjective to the scope and impact of the change and should ideally be done in collaboration with your respective OpenShift staff engineer, product manager and PLM contact.
In order of preference, in what teams should choose:
- Insight rule notification -
These appear on cloud.redhat.com and on cluster when customers are logged in)
- Insights Rule Contribution wiki
- Contact Jan Holecek for questions.
- Technical Topic Torrent (T3) -
This is a communication to TAM/CSM’s (GCS) and GCS customers; delivered via
email.
- Read the T3 FAQ for more details.
- Contact gcs-tam-ocp@redhat.com if you have questions.
- Email Campaigns to affected customers -
This is direct email contact with customers
- While this should ideally be viewed as a last option, history has indicated that this is more effective 90% of time over customer portal banners.
- Contact Will Wang for questions.
- Customer Portal Banners -
This will reach customers visiting our access.redhat.com (Customer Portal)
web properties (IE: Docs, Product Pages, Downloads, etc)
- The linked document above is used to request the banner. The resulting banner can live on any page (negotiated with the Customer Portal Team) on the Red Hat Customer Portal. Here is an example.
- Note: This option can be targeted to ‘customers’, ‘web properties’ on the customer portal (IE: docs, or product pages, etc).
- Note: The SRE and Services teams also status pages available to them that could also be potentially used to inform customers based on the scenario.
- Contact Jacquelene Booker for questions.
In addition to the above methods, the team has an option to backport a ClusterOperator Upgradable=False guard in certain scenarios. This is not a preferable but in situations where the percentage of impacted clusters is small and the impact of upgrading without addressing the situation is high, it might be applicable.