github.com/openshift/installer

Added

• The installer binary now includes all required Terraform plugins, so there is no longer a need to separately download and install them. This will be most noticeable on libvirt, where users used to be required to install the libvirt plugin manually. This avoids issues with mismatched plugin versions (which we saw sometimes on libvirt) and network-connectivity issues during create cluster invocations (which we saw sometimes on all platforms).
• The configured base domain is now pushed into the cluster's config.openshift.io as a DNS custom resource.

Changed

• install-config.yml is now install-config.yaml to align with our usual YAML extension. install-config.yml is deprecated, and support for it will be removed completely in the next release.

• On AWS, we now use a select widget for the base-domain wizard prompt, making it easier to choose an existing public zone.

• On AWS, Route 53 rate limits during cluster destroy are now less disruptive, reducing the AWS request load in busy accounts.

• On OpenStack, the HAProxy configuration no longer hard-codes the cluster name and base domain.

• On OpenStack, the 0.7.0 fix for:

    FATAL Expected HTTP response code [202 204] when accessing [DELETE https://osp-xxxxx:13696/v2.0/routers/52093478-dcf1-4bcc-9a2c-dbb1e42da880], but got 409 instead
    {"NeutronError": {"message": "Router 52093478-dcf1-4bcc-9a2c-dbb1e42da880 still has ports", "type": "RouterInUse", "detail": ""}}
  

  was incorrect and has been reverted. We'll land a real fix for this issue in future work.

• On OpenStack, the service VM from 0.7.0 now has a floating IP address.

• All libvirt functionality is behind TAGS=libvirt now. Previously installer builds with TAGS=libvirt_destroy included all libvirt functionality, while builds without that tag would include create cluster but not destroy cluster functionality. With the change, all users using the installer with libvirt clusters will need to set the new build tag.

• Lots of doc and internal cleanup and minor fixes.

Removed

• On AWS and OpenStack, the tectonicClusterID tag which was deprecated in 0.7.0 has been removed.

Added

• We now validate install-config when loading it during staged installs. Previously we only validated that input when it was entered into the wizard or via environment variables. This also leads to some changes in which values are considered valid:

  • Base domains may no longer contain uppercase letters.
  • Cluster names may now be longer than 63 characters, although as specified in [RFC 1123][rfc-1123-s2.1] host software may not support names longer than 63 characters.
  • Pull secrets require more content (e.g. it must contain an auths property). Previously we only required pull secrets to be valid JSON.
  • SSH public keys must be parsable with [ParseAuthorizedKey][ssh.ParseAuthorizedKey]. Previously we had our own logic that was not as well developed.

• We've added images/installer/Dockerfile.ci.rhel7 for building installer images on a RHEL base.

• On AWS, we now create [an S3 endpoint][aws-s3-endpoint] for the VPC.

• We've added OpenStack documentation.

Changed

• The pull-secret prompt now masks the input to avoid leaking it to shoulder surfers and similar.
• The pull-secret prompt's help now points to cloud.openshift.com instead of try.openshift.com. This should make it easier to find the pull secret without digging through a page of introductory content.
• The initial kubeconfig inserted on master nodes used to have admin privileges, but only for 30 minutes. Now it has role bindings that allow it to create and receive automatic approval for certificate signing requests, but it does not have additional privileges beyond that.
• On AWS and OpenStack, master ports 10251 (scheduler) and 10252 (controller manager) have been opened to access from all machines. This allows Prometheus (which runs on the worker nodes) to scrape all machines for metrics.
• On AWS and OpenStack, the installer and subsequent cluster will now tag resources it creates with openshiftClusterID. tectonicClusterID is deprecated.
• On OpenStack, only the OpenStack clouds entry is marshalled into the openstack-creds secret. Previously we had injected the host's entire cloud configuration.
• On OpenStack, there is now a service VM to provide DNS and load balancing for the OpenShift cluster. The service VM will eventually be removed, but for now its a convenient hack to get usable clusters on OpenStack.
• On libvirt, we now document host DNS configuration as required, because too many users were skipping that step and then reporting errors with Kubernetes API detection when the install-host failed to resolve the cluster domain name while waiting for the bootstrap-complete event.
• Lots of doc and internal cleanup and minor fixes.

Fixed

• Fixed OpenShift manifest loading during staged installs. The installer had been ignoring changes to those files since 0.4.0.

• Fixed you must pass a pointer as the target of a Write operation errors introduced in 0.6.0 for the AWS access key ID prompt.

• When create cluster times out waiting for the Kubernetes API, we now exit immediately. Previously we'd wait through another 30 minutes of failed event-listener connections before failing this case. We've also fixed similar timeout detection for the code that waits for the OpenShift console route.

• On OpenStack, we've fixed a bug in router deletion:

    FATAL Expected HTTP response code [202 204] when accessing [DELETE https://osp-xxxxx:13696/v2.0/routers/52093478-dcf1-4bcc-9a2c-dbb1e42da880], but got 409 instead
    {"NeutronError": {"message": "Router 52093478-dcf1-4bcc-9a2c-dbb1e42da880 still has ports", "type": "RouterInUse", "detail": ""}}
  

• On libvirt, we've fixed a bug introduced in 0.6.0 and are now back to removing the bootstrap node from round-robin DNS when we destroy the bootstrap resources.

Removed

• The user-facing OPENSHIFT_INSTALL_* environment variables are gone. Instead, users who want to skip the wizard are encouraged to provide their own install-config.
• On AWS, the option to install a cluster into an existing VPC is gone. Users who would have previously done this can use [VPC peering][aws-vpc-peering].

Added

• We now push a kubeadmin user (with an internally-generated password) into the cluster for the new [bootstrap identity provider][bootstrap-identity-provider]. This gives users a way to access the web console, Prometheus, etc. without needing to configure a full-fledged identity provider or install oc. The create cluster subcommand now blocks until the web-console route is available and then exits after printing instructions for using the new credentials.
• The installer binary now includes Terraform so there is no longer a need to separately download and install it.

Changed

• The SSH public key configuration has moved a level up in the install config, now that the admin structure has been removed.
• build.sh now checks to make sure you have a new enough go, instead of erroring out partway through the build.
• We now resolve the update payload to a digest on the bootstrap node, so [the cluster-version-operator][cluster-version-operator] can figure out exactly which image we used.
• Creation logging has been overhauled to increase it's signal-to-noise while waiting for the Kubernetes API to come up.
• On AWS, the installer will now prompt you for an access key and secret if it cannot find your AWS credentials in the usual places.
• On AWS, the installer will look at AWS_DEFAULT_REGION and in other usual places when picking a default for the region prompt. You still have to set OPENSHIFT_INSTALL_AWS_REGION if you want to skip the prompt entirely.
• On libvirt, we've bumped masters from 3 GiB of memory to 4 GiB to address out-of-memory issues we had been seeing at 3 GiB.
• Lots of doc and internal cleanup and minor fixes.

Removed

• The old admin username and password inputs have been removed. They weren't being used anyway, and their intended role has been replaced by the newly-added kubeadmin user and bootstrap identity provider.
• The old openshift-web-console namespace is gone. The new console is in the openshift-console namespace.

Added

• We now push the ingress custom resource definition and initial configuration, allowing the ingress operator to configure itself without referencing the deprecated cluster-config-v1 resource.

Changed

• Pull secret documentation now points to try.openshift.com for pull-secret acquisition, instead of pointing at account.coreos.com. Users will need to update their pull secrets.
• If the automatic bootstrap teardown (which landed in 0.4.0) times out waiting for the bootstrap-complete event, the installer exits with a non-zero exit code. We had ignored watcher timeouts in 0.4.0 due to concerns about watcher robustness, but the current watcher code has been reliable in our continuous integration testing.
• The hard-coded quay.io/coreos/bootkube dependency has been replaced by the new [cluster-bootstrap][] image, which is referenced from the release image.
• The etcd service now uses [selectors][kube-selector] to determine the pods it exposes, and the explict etcd endpoints object is gone (replaced by the one Kubernetes maintains based on the selector).
• On AWS, both masters and worker have moved from t2.medium nodes m4.large nodes (more on AWS instance types [here][aws-instance-types]) to address CPU and memory constraints.
• On AWS, master volume size has been bumped from 30 GiB to 120 GiB to increase our baseline performance from on [gp2's sliding IOPS scale][aws-ebs-gp2-iops] from the 100 IOPS floor up to 360 IOPS. Volume information is not currently supported by [the cluster-API AWS provider's AWSMachineProviderConfig][cluster-api-provider-aws-012575c1-AWSMachineProviderConfig], so this change is currently limited to masters created by the installer.
• On Openstack, we now validate cloud, region, and image-name user input instead of blindly accepting entries.
• On libvirt, we pass Ignition information for masters and workers via secrets instead of passing a libvirt volume path. This makes the libvirt approach consistent with how we already handle AWS and OpenStack.
• Lots of internal cleanup, especially around trimming dead code.

Fixed

• The .openshift_install.log addition from 0.4.0 removed Terraform output from --log-level=debug. We've fixed that in 0.5.0; now .openshift_install.log will always contain the full Terraform output, while standard error returns to containing the Terraform output if and only if --log-level=debug or higher.

• On AWS teardown, errors retrieving tags for S3 buckets and Route 53 zones are no longer fatal. This allows the teardown code to continue it's exponential backoff and try to remove the bucket or zone later. It avoids some resource leaks we were seeing due to AWS rate limiting on those tag lookups as many simultaneous CI jobs searched for Route 53 zones with their cluster's tags. We'll still hit those rate limits, but they no longer cause us to give up on reaping resources.

• On AWS, we've removed some unused data blocks, fixing occasional errors like:

    data.aws_route_table.worker.1: Your query returned no results.
  

• On OpenStack, similar retry-during-teardown changes were made for removing ports and for removing subnets from routers.

• On libvirt, Terraform no longer errors out when launching clusters configured for more than one master, fixing a bug from 0.4.0.

Added

• The creation targets have been moved below a new create subcommand (e.g. openshift-install create cluster instead of the old openshift-install cluster). This makes them easier to distinguish from other openshift-install subcommands and also mirrors the approach taken by destroy in 0.3.0.
• A new manifest-templates target has been added to create, allowing users to edit templates and have descendant assets generated from their altered templates during a staged install.
• [The ingress operator][ingress-operator] is no longer masked. The old Tectonic ingress operator has been removed.
• The [the registry operator][registry-operator] has been added, and the kube-addon operator which used to provide a registry (among other things) has been removed.
• The [checkpointer operator][checkpointer-operator] is no longer masked. It runs on the production cluster, but not on the bootstrap node.
• Cloud credentials are now pushed into a secret where they can be consumed by cluster-API operators and other tools.
• OpenStack now has destroy support.
• We log verbosely to ${INSTALL_DIR}/.openshift_install.log for most operations, giving access to the logs for troubleshooting even if you neglected to run with --log-level=debug.
• We've grown troubleshooting documentation.

Changed

• The create cluster subcommand now waits for the bootstrap-complete event and automatically removes the bootstrap assets after receiving it. This means that after create cluster returns successfully, the cluster has its production control plane and topology (although there may still be operators working through their initialization). The bootstrap-complete event was new in 0.3.0, and it is now pushed at the appropriate time (it was too early in 0.3.0). The destroy bootstrap subcommand is still available, to allow users to manually trigger bootstrap deletion if the automatic removal fails for whatever reason.

• On AWS, bootstrap deletion now also removes the S3 bucket used for the bootstrap node's Ignition configuration.

• Asset state is preserved even while moving backwards through a staged install. For example:

  openshift-install --dir=example create ignition-configs
  openshift-install --dir=example create install-config
  

  now preserves the full state including the generated Ignition configuration. In 0.3.0, the install-config call would have removed the Ignition configuration and other downstream assets from the stored state.

• Some asset state is removed by successful destroy cluster runs. This reduces the change of contaminating future cluster creation with assets left over from a previous cluster, but users are still encouraged to remove state between clusters to avoid accidentally contaminating the subsequent cluster's state.

• etcd discovery now happens via SRV records. On libvirt, this requires a new Terraform provider, so users with older providers should install a newer version. This also allows all masters to use a single Ignition file.

• On AWS, the API and service load balancers have been changed from [classic load balancers][aws-elb] to [network load balancers][aws-nlb]. This should avoid [some latency issues we were seeing with classic load balancers][aws-elb-latency], and network load balancers are cheaper.

• On AWS, master Machine entries now include load balancer references, ensuring that new masters created by [the AWS cluster-API provider][cluster-api-provider-aws] will be attached to the load balancers.

• On AWS and OpenStack, the default network CIDRs have changed to 172.30.0.0/16 for services and 10.128.0.0/14 for the cluster, to be consistent with previous versions of OpenStack.

• The bootstrap kubelet is no longer part of the production cluster. This reduces complexity and keeps production pods off of the temporary bootstrap node.

• [The cluster-version operator][cluster-version-operator] now runs in a static pod on the bootstrap node until the production control plane comes up. This breaks a cyclic dependency between the production API server and operators.

• The bootstrap control plane now waits for some core pods to come up before exiting.

• [The machine-API operator][machine-api-operator] now reads the install-config from the cluster-config-v1 config-map, instead of from an operator-specific configuration.

• AWS AMIs and libvirt images are now pulled from the new [RHCOS pipeline][rhcos-pipeline].

• Updated the security contact information for CoreOS -> Red Hat.

• We push a ClusterVersion custom resource. The old CVOConfig is still being pushed, but it is deprecated.

• OpenStack credentials are loaded from standard system paths.

• On AWS and OpenStack, ports 9000-9999 are now open for host network services.

• Lots of doc and internal cleanup and minor fixes.

Fixed

• On AWS, destroy cluster is now more robust, removing resources with either the tectonicClusterID or kubernetes.io/cluster/<name>: owned tags. It also removes pending instances as well (it used to only remove running instances).
• On libvirt, destroy cluster is now more precise, only removing resources which are prefixed by the cluster name.
• Bootstrap Ignition edits (via create ignition-configs) no longer suffer from a worker.ign dependency cycle, which had been clobbering manual bootstrap.ign changes.
• The state-purging implementation respects --dir, avoiding remove ...: no such file or directory errors during staged installs.
• Cross-filesystem Terraform state recovery during destroy bootstrap no longer raises invalid cross-device link.
• Bootstrap binaries are now located under /usr/local/bin, avoiding SELinux violations on RHEL 8.

Removed

• All the old Tectonic operators and the tectonic-system namespace have been removed.
• On libvirt, the image URI prompt has been removed. You can still control this via the OPENSHIFT_INSTALL_LIBVIRT_IMAGE environment variable, but too many users were breaking their cluster by pointing the installer at an outdated RHCOS, so we removed the prompt to make that knob less obvious.
• On libvirt, we've removed .gz suffix handling for images. The new RHCOS pipeline supports Content-Encoding: gzip, so the suffix-based hack is no longer necessary.
• The destroy-cluster command, which was deprecated in favor of destroy cluster in 0.3.0, has been removed.
• The creation target subcommands of openshift-install have been removed. Use the target subcommands of create instead (e.g. openshift-install create cluster instead of openshift-install cluster).

Added

• Asset state is loaded from the install directory, allowing for a staged install.

• A new openshift-install destroy bootstrap command destroys the bootstrap resources. Ideally, this would be safe to run after the new bootstrap-complete event is pushed to the kube-system namespace, but there is currently a bug causing that event to be pushed too early. For now, you're on your own figuring out when to call this command.

  For consistency, the old destroy-cluster has been deprecated in favor of openshift-install destroy cluster.

• The installer creates worker MachineSets, instead of leaving that to [the machine-API operator][machine-api-operator].

• The installer creates master Machines and tags masters to be picked up by the [AWS cluster-API provider][cluster-api-provider-aws].

Changed

• The installer now respects the AWS_PROFILE environment variable when launching AWS clusters.
• Worker subnets are now created in the appropriate availability zone for AWS clusters.
• Use the released hyperkube and hypershift instead of hard-coded images.
• Lots of changes to keep up with the advancing release image, as OpenShift operators are added to control various cluster components.
• Lots of internal cleanup and minor fixes.

Removed

• The Tectonic kube-core operator, which has been replaced by OpenShift operators.

Added

• Asset state is preserved between invocations, allowing for a staged install like:

  $ openshift-install --dir=example install-config
  $ openshift-install --dir=example cluster
  

  which creates a cluster using the same data given in the install-config (including the same random cluster ID, etc.).

• [The kube-apiserver][kube-apiserver-operator] and [kube-controller-manager][kube-controller-manager-operator] operators are called to render additional cluster manifests.

• etcd is now available as a service in the kube-system namespace, and the new service is labeled so [Prometheus][] will scrape it.

• The service-serving-cert-signer-signing-key secret is now available in the openshift-service-cert-signer namespace, which gives [the service-serving cert signer][service-serving-cert-signer] the keys it needs to mint and manage certificates for Kubernetes services.

• The etcd-serving certificate is now passed through to [the kube-controller-manager operator][kube-controller-manager-operator].

• We disable some components which [the cluster-version operator][cluster-version-operator] would otherwise install but which conflict with the legacy tectonic-operators.

• The new openshift-install graph outputs the asset graph in [the DOT language][dot].

• openshift-install version now outputs the Terraform version as well as the installer version.

Changed

• The [cluster-version operator][cluster-version-operator] is no longer run as a static pod. Instead, we just wait until the control plane comes up and run it them.
• Terraform errors are logged to standard error even when --log-level is less than debug.
• Terraform is now invoked with -no-color and -input=false.
• The cluster target now includes both launching the cluster and populating metadata.json, regardless of whether the terraform invocation succeeds. This allows destroy-cluster to cleanup cluster resources even when the terraform invocation fails.
• Reported errors now include more context, making them less enigmatic.
• Libvirt image caching is more efficient, caching unzipped images with a cache that grows by one unzipped image per RHCOS release in $XDG_CACHE_HOME/openshift-install/libvirt/image. The previous implementation unzipped, when necessary, for every launched cluster, which was slow. And the previous implementation added one unzipped image to /tmp per cluster launch, which consumed more disk space.
• Work continues on the OpenStack platform.
• Lots of internal cleanup, especially around asset generation.

Removed

• The operatorstatus CRD. Now [the cluster-version operator][cluster-version-operator] creates this on its own.
• The machine-config-operator-images config-map. Now [the cluster-version operator][cluster-version-operator] pulls these from [the machine-config images][machine-config-operator].
• The machine-api app-version from the tectonic-system namespace.

Added

The openshift-install command. This moves us to the new install-config approach with asset generation in Go instead of in Terraform. Terraform is still used to push the assets out to resources on the backing platform (AWS, libvirt, or OpenStack), but that push happens in a single Terraform invocation instead of in multiple steps. This makes installation faster, because more resources can be created in parallel. openshift-install also dispenses with the distribution tarball; all required assets except for a terraform binary are distributed in the openshift-install binary.

The configuration and command-line interface are quite different, so previous tectonic users are encouraged to start from scratch when getting acquainted with openshift-install. AWS users should look here. Libvirt users should look here. The new openshift-install also includes an interactive configuration generator, so you can launch the installer and follow along as it guides you through the process.

Removed

The tectonic command and tarball distribution are gone. Please use the new openshift-install command instead.