Refactors setting status into its own top-level function to make it
easier to ensure we _always_ set the status if it's changed on every
reconcile. Previously, it was possible to have stale status if some
earlier part of the provision logic failed.
Updates #16327
Change-Id: Idab0cfc15ae426cf6914a82f0d37a5cc7845236b
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Previously, the operator checked the ProxyGroup status fields for
information on how many of the proxies had successfully authed. Use
their state Secrets instead as a more reliable source of truth.
containerboot has written device_fqdn and device_ips keys to the
state Secret since inception, and pod_uid since 1.78.0, so there's
no need to use the API for that data. Read it from the state Secret
for consistency. However, to ensure we don't read data from a
previous run of containerboot, make sure we reset containerboot's
state keys on startup.
One other knock-on effect of that is ProxyGroups can briefly be
marked not Ready while a Pod is restarting. Introduce a new
ProxyGroupAvailable condition to more accurately reflect
when downstream controllers can implement flows that rely on a
ProxyGroup having at least 1 proxy Pod running.
Fixes#16327
Change-Id: I026c18e9d23e87109a471a87b8e4fb6271716a66
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Adds Recorder fields to configure the name and annotations of the ServiceAccount
created for and used by its associated StatefulSet. This allows the created Pod
to authenticate with AWS without requiring a Secret with static credentials,
using AWS' IAM Roles for Service Accounts feature, documented here:
https://docs.aws.amazon.com/eks/latest/userguide/iam-roles-for-service-accounts.htmlFixes#15875
Change-Id: Ib0e15c0dbc357efa4be260e9ae5077bacdcb264f
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
The defaultEnv and defaultBool functions are copied over temporarily
to minimise diff. This lays the ground work for having both the operator
and the new k8s-proxy binary implement the API proxy
Updates #13358
Change-Id: Ieacc79af64df2f13b27a18135517bb31c80a5a02
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Adds a new diagram for ProxyGroups running in Ingress mode.
Documentation is currently not publicly available, but a link needs
adding once it is.
Updates tailscale/corp#24795
Change-Id: I0d5dd6bf6f0e1b8b0becae848dc97d8b4bfb9ccb
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Ensure no services are advertised as part of shutting down tailscaled.
Prefs are only edited if services are currently advertised, and they're
edited we wait for control's ~15s (+ buffer) delay to failover.
Note that editing prefs will trigger a synchronous write to the state
Secret, so it may fail to persist state if the ProxyGroup is getting
scaled down and therefore has its RBAC deleted at the same time, but that
failure doesn't stop prefs being updated within the local backend,
doesn't affect connectivity to control, and the state Secret is
about to get deleted anyway, so the only negative side effect is a harmless
error log during shutdown. Control still learns that the node is no
longer advertising the service and triggers the failover.
Note that the first version of this used a PreStop lifecycle hook, but
that only supports GET methods and we need the shutdown to trigger side
effects (updating prefs) so it didn't seem appropriate to expose that
functionality on a GET endpoint that's accessible on the k8s network.
Updates tailscale/corp#24795
Change-Id: I0a9a4fe7a5395ca76135ceead05cbc3ee32b3d3c
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
The test suite had grown to about 20s on my machine, but it doesn't
do much taxing work so was a good candidate to parallelise. Now runs
in under 2s on my machine.
Updates #cleanup
Change-Id: I2fcc6be9ca226c74c0cb6c906778846e959492e4
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Switch from using the Comment field to a ts-scoped annotation for
tracking which operators are cooperating over ownership of a
VIPService.
Updates tailscale/corp#24795
Change-Id: I72d4a48685f85c0329aa068dc01a1a3c749017bf
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
There was a flaky failure case where renaming a TLS hostname for an
ingress might leave the old hostname dangling in tailscaled config. This
happened when the proxygroup reconciler loop had an outdated resource
version of the config Secret in its cache after the
ingress-pg-reconciler loop had very recently written it to delete the
old hostname. As the proxygroup reconciler then did a patch, there was
no conflict and it reinstated the old hostname.
This commit updates the patch to an update operation so that if the
resource version is out of date it will fail with an optimistic lock
error. It also checks for equality to reduce the likelihood that we make
the update API call in the first place, because most of the time the
proxygroup reconciler is not even making an update to the Secret in the
case that the hostname has changed.
Updates tailscale/corp#24795
Change-Id: Ie23a97440063976c9a8475d24ab18253e1f89050
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Update the HA Ingress controller to wait until it sees AdvertisedServices
config propagated into at least 1 Pod's prefs before it updates the status
on the Ingress, to ensure the ProxyGroup Pods are ready to serve traffic
before indicating that the Ingress is ready
Updates tailscale/corp#24795
Change-Id: I1b8ce23c9e312d08f9d02e48d70bdebd9e1a4757
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
When the Ingress is updated to a new hostname, the controller does not
currently clean up the old VIPService from control. Fix this up to parse
the ownership comment correctly and write a test to enforce the improved
behaviour
Updates tailscale/corp#24795
Change-Id: I792ae7684807d254bf2d3cc7aa54aa04a582d1f5
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
PR #14771 added support for getting certs from alternate ACME servers, but the
certStore caching mechanism breaks unless you install the CA in system roots,
because we check the validity of the cert before allowing a cache hit, which
includes checking for a valid chain back to a trusted CA. For ease of testing,
allow cert cache hits when the chain is unknown to avoid re-issuing the cert
on every TLS request served. We will still get a cache miss when the cert has
expired, as enforced by a test, and this makes it much easier to test against
non-prod ACME servers compared to having to manage the installation of non-prod
CAs on clients.
Updates #14771
Change-Id: I74fe6593fe399bd135cc822195155e99985ec08a
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Now that packets flow for VIPServices, the last piece needed to start
serving them from a ProxyGroup is config to tell the proxy Pods which
services they should advertise.
Updates tailscale/corp#24795
Change-Id: Ic7bbeac8e93c9503558107bc5f6123be02a84c77
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
tailscaled's ipn package writes a collection of keys to state after
authenticating to control, but one at a time. If containerboot happens
to send a SIGTERM signal to tailscaled in the middle of writing those
keys, it may shut down with an inconsistent state Secret and never
recover. While we can't durably fix this with our current single-use
auth keys (no atomic operation to auth + write state), we can reduce
the window for this race condition by checking for partial state
before sending SIGTERM to tailscaled. Best effort only.
Updates #14080
Change-Id: I0532d51b6f0b7d391e538468bd6a0a80dbe1d9f7
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Adds an envknob setting for changing the client's ACME directory URL.
This allows testing cert issuing against LE's staging environment, as
well as enabling local-only test environments, which is useful for
avoiding the production rate limits in test and development scenarios.
Fixes#14761
Change-Id: I191c840c0ca143a20e4fa54ea3b2f9b7cbfc889f
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
The CN field is technically deprecated; set the requested name in a DNS SAN
extension in addition to maximise compatibility with RFC 8555.
Fixes#14762
Change-Id: If5d27f1e7abc519ec86489bf034ac98b2e613043
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
The new ProxyGroup-based Ingress reconciler is causing a fatal log at
startup because it has the same name as the existing Ingress reconciler.
Explicitly name both to ensure they have unique names that are consistent
with other explicitly named reconcilers.
Updates #14583
Change-Id: Ie76e3eaf3a96b1cec3d3615ea254a847447372ea
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Most users should not run into this because it's set in the helm chart
and the deploy manifest, but if namespace is not set we get confusing
authz errors because the kube client tries to fetch some namespaced resources
as though they're cluster-scoped and reports permission denied. Try to
detect namespace from the default projected volume, and otherwise fatal.
Fixes #cleanup
Change-Id: I64b34191e440b61204b9ad30bbfa117abbbe09c3
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
This is an experiment to see how useful we will find it to have some
text-based diagrams to document how various components of the operator
work. There are no plans to link to this from elsewhere yet, but
hopefully it will be a useful reference internally.
Updates #cleanup
Change-Id: If5911ed39b09378fec0492e87738ec0cc3d8731e
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
This is the start of an integration/e2e test suite for the tailscale operator.
It currently only tests two major features, ingress proxy and API server proxy,
but we intend to expand it to cover more features over time. It also only
supports manual runs for now. We intend to integrate it into CI checks in a
separate update when we have planned how to securely provide CI with the secrets
required for connecting to a test tailnet.
Updates #12622
Change-Id: I31e464bb49719348b62a563790f2bc2ba165a11b
Co-authored-by: Irbe Krumina <irbe@tailscale.com>
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Previously this unit test failed if it was run in a container. Update the assert
to focus on exactly the condition we are trying to assert: the package type
should only be 'container' if we use the build tag.
Updates #14317
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Every so often, the ProxyGroup and other controllers lose an optimistic locking race
with other controllers that update the objects they create. Stop treating
this as an error event, and instead just log an info level log line for it.
Fixes#14072
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
When the operator enables metrics on a proxy, it uses the port 9001,
and in the near future it will start using 9002 for the debug endpoint
as well. Make sure we don't choose ports from a range that includes
9001 so that we never clash. Setting TS_SOCKS5_SERVER, TS_HEALTHCHECK_ADDR_PORT,
TS_OUTBOUND_HTTP_PROXY_LISTEN, and PORT could also open arbitrary ports,
so we will need to document that users should not choose ports from the
10000-11000 range for those settings.
Updates #13406
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
* cmd/containerboot: serve health on local endpoint
We introduced stable (user) metrics in #14035, and `TS_LOCAL_ADDR_PORT`
with it. Rather than requiring users to specify a new addr/port
combination for each new local endpoint they want the container to
serve, this combines the health check endpoint onto the local addr/port
used by metrics if `TS_ENABLE_HEALTH_CHECK` is used instead of
`TS_HEALTHCHECK_ADDR_PORT`.
`TS_LOCAL_ADDR_PORT` now defaults to binding to all interfaces on 9002
so that it works more seamlessly and with less configuration in
environments other than Kubernetes, where the operator always overrides
the default anyway. In particular, listening on localhost would not be
accessible from outside the container, and many scripted container
environments do not know the IP address of the container before it's
started. Listening on all interfaces allows users to just set one env
var (`TS_ENABLE_METRICS` or `TS_ENABLE_HEALTH_CHECK`) to get a fully
functioning local endpoint they can query from outside the container.
Updates #14035, #12898
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
containerboot:
Adds 3 new environment variables for containerboot, `TS_LOCAL_ADDR_PORT` (default
`"${POD_IP}:9002"`), `TS_METRICS_ENABLED` (default `false`), and `TS_DEBUG_ADDR_PORT`
(default `""`), to configure metrics and debug endpoints. In a follow-up PR, the
health check endpoint will be updated to use the `TS_LOCAL_ADDR_PORT` if
`TS_HEALTHCHECK_ADDR_PORT` hasn't been set.
Users previously only had access to internal debug metrics (which are unstable
and not recommended) via passing the `--debug` flag to tailscaled, but can now
set `TS_METRICS_ENABLED=true` to expose the stable metrics documented at
https://tailscale.com/kb/1482/client-metrics at `/metrics` on the addr/port
specified by `TS_LOCAL_ADDR_PORT`.
Users can also now configure a debug endpoint more directly via the
`TS_DEBUG_ADDR_PORT` environment variable. This is not recommended for production
use, but exposes an internal set of debug metrics and pprof endpoints.
operator:
The `ProxyClass` CRD's `.spec.metrics.enable` field now enables serving the
stable user metrics documented at https://tailscale.com/kb/1482/client-metrics
at `/metrics` on the same "metrics" container port that debug metrics were
previously served on. To smooth the transition for anyone relying on the way the
operator previously consumed this field, we also _temporarily_ serve tailscaled's
internal debug metrics on the same `/debug/metrics` path as before, until 1.82.0
when debug metrics will be turned off by default even if `.spec.metrics.enable`
is set. At that point, anyone who wishes to continue using the internal debug
metrics (not recommended) will need to set the new `ProxyClass` field
`.spec.statefulSet.pod.tailscaleContainer.debug.enable`.
Users who wish to opt out of the transitional behaviour, where enabling
`.spec.metrics.enable` also enables debug metrics, can set
`.spec.statefulSet.pod.tailscaleContainer.debug.enable` to false (recommended).
Separately but related, the operator will no longer specify a host port for the
"metrics" container port definition. This caused scheduling conflicts when k8s
needs to schedule more than one proxy per node, and was not necessary for allowing
the pod's port to be exposed to prometheus scrapers.
Updates #11292
---------
Co-authored-by: Kristoffer Dalby <kristoffer@tailscale.com>
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
We currently annotate pods with a hash of the tailscaled config so that
we can trigger pod restarts whenever it changes. However, the hash
updates more frequently than is necessary causing more restarts than is
necessary. This commit removes two causes; scaling up/down and removing
the auth key after pods have initially authed to control. However, note
that pods will still restart on scale-up/down because of the updated set
of volumes mounted into each pod. Hopefully we can fix that in a planned
follow-up PR.
Updates #13406
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Our existing container-detection tricks did not work on Kubernetes,
where Docker is no longer used as a container runtime. Extends the
existing go build tags for containers to the other container packages
and uses that to reliably detect builds that were created by Tailscale
for use in a container. Unfortunately this doesn't necessarily improve
detection for users' custom builds, but that's a separate issue.
Updates #13825
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
As discussed in #13684, base the ProxyGroup's proxy definitions on the same
scaffolding as the existing proxies, as defined in proxy.yaml
Updates #13406
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Instead of converting our PortMap struct to a string during marshalling
for use as a key, convert the whole collection of PortMaps to a list of
PortMap objects, which improves the readability of the JSON config while
still keeping the data structure we need in the code.
Updates #13406
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
The default ProxyClass can be set via helm chart or env var, and applies
to all proxies that do not otherwise have an explicit ProxyClass set.
This ensures proxies created by the new ProxyGroup CRD are consistent
with the behaviour of existing proxies
Nearby but unrelated changes:
* Fix up double error logs (controller runtime logs returned errors)
* Fix a couple of variable names
Updates #13406
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Rearrange conditionals to reduce indentation and make it a bit easier to read
the logic. Also makes some error message updates for better consistency
with the recent decision around capitalising resource names and the
upcoming addition of config secrets.
Updates #cleanup
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Implements the controller for the new ProxyGroup CRD, designed for
running proxies in a high availability configuration. Each proxy gets
its own config and state Secret, and its own tailscale node ID.
We are currently mounting all of the config secrets into the container,
but will stop mounting them and instead read them directly from the kube
API once #13578 is implemented.
Updates #13406
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
The ProxyGroup CRD specifies a set of N pods which will each be a
tailnet device, and will have M different ingress or egress services
mapped onto them. It is the mechanism for specifying how highly
available proxies need to be. This commit only adds the definition, no
controller loop, and so it is not currently functional.
This commit also splits out TailnetDevice and RecorderTailnetDevice
into separate structs because the URL field is specific to recorders,
but we want a more generic struct for use in the ProxyGroup status field.
Updates #13406
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
containerboot's main.go had grown to well over 1000 lines with
lots of disparate bits of functionality. This commit is pure copy-
paste to group related functionality outside of the main function
into its own set of files. Everything is still in the main package
to keep the diff incremental and reviewable.
Updates #cleanup
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
cmd/k8s-operator,k8s-operator,kube: Add TSRecorder CRD + controller
Deploys tsrecorder images to the operator's cluster. S3 storage is
configured via environment variables from a k8s Secret. Currently
only supports a single tsrecorder replica, but I've tried to take early
steps towards supporting multiple replicas by e.g. having a separate
secret for auth and state storage.
Example CR:
```yaml
apiVersion: tailscale.com/v1alpha1
kind: Recorder
metadata:
name: rec
spec:
enableUI: true
```
Updates #13298
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
cmd/containerboot,cmd/k8s-operator: enable IPv6 for fqdn egress proxies
Don't skip installing egress forwarding rules for IPv6 (as long as the host
supports IPv6), and set headless services `ipFamilyPolicy` to
`PreferDualStack` to optionally enable both IP families when possible. Note
that even with `PreferDualStack` set, testing a dual-stack GKE cluster with
the default DNS setup of kube-dns did not correctly set both A and
AAAA records for the headless service, and instead only did so when
switching the cluster DNS to Cloud DNS. For both IPv4 and IPv6 to work
simultaneously in a dual-stack cluster, we require headless services to
return both A and AAAA records.
If the host doesn't support IPv6 but the FQDN specified only has IPv6
addresses available, containerboot will exit with error code 1 and an
error message because there is no viable egress route.
Fixes#12215
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
go get github.com/tailscale/mkctr@main
Pulls in changes to support a local target that only pushes
a single-platform image to the machine's local image store.
Fixestailscale/mkctr#18
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Adds a new TailscaleProxyReady condition type for use in corev1.Service
conditions.
Also switch our CRDs to use metav1.Condition instead of
ConnectorCondition. The Go structs are seralized identically, but it
updates some descriptions and validation rules. Update k8s
controller-tools and controller-runtime deps to fix the documentation
generation for metav1.Condition so that it excludes comments and
TODOs.
Stop expecting the fake client to populate TypeMeta in tests. See
kubernetes-sigs/controller-runtime#2633 for details of the change.
Finally, make some minor improvements to validation for service hostnames.
Fixes#12216
Co-authored-by: Irbe Krumina <irbe@tailscale.com>
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Add new fields TailnetIPs and Hostname to Connector Status. These
contain the addresses of the Tailscale node that the operator created
for the Connector to aid debugging.
Fixes#12214
Signed-off-by: Tom Proctor <tomhjp@users.noreply.github.com>
Tom ProctorandGitHub
