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>
cmd/{containerboot,k8s-operator},kube: add preshutdown hook for egress PG proxies
This change is part of work towards minimizing downtime during update
rollouts of egress ProxyGroup replicas.
This change:
- updates the containerboot health check logic to return Pod IP in headers,
if set
- always runs the health check for egress PG proxies
- updates ClusterIP Services created for PG egress endpoints to include
the health check endpoint
- implements preshutdown endpoint in proxies. The preshutdown endpoint
logic waits till, for all currently configured egress services, the ClusterIP
Service health check endpoint is no longer returned by the shutting-down Pod
(by looking at the new Pod IP header).
- ensures that kubelet is configured to call the preshutdown endpoint
This reduces the possibility that, as replicas are terminated during an update,
a replica gets terminated to which cluster traffic is still being routed via
the ClusterIP Service because kube proxy has not yet updated routig rules.
This is not a perfect check as in practice, it only checks that the kube
proxy on the node on which the proxy runs has updated rules. However, overall
this might be good enough.
The preshutdown logic is disabled if users have configured a custom health check
port via TS_LOCAL_ADDR_PORT env var. This change throws a warnign if so and in
future setting of that env var for operator proxies might be disallowed (as users
shouldn't need to configure this for a Pod directly).
This is backwards compatible with earlier proxy versions.
Updates tailscale/tailscale#14326
Signed-off-by: Irbe Krumina <irbe@tailscale.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>
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>
Tom Proctor
Irbe Krumina