Followup to #2990, which refactored `linkerd endpoints` to use the
`Destination.Get` API instead of the `Discovery.Endpoints` API, leaving
the Discovery with no implented methods. This PR removes all the Discovery
code leftovers.
Fixes#3499
* Re-add the destination container to the controller spec
This fix is necessary to avoid data plane downtime during an upgrade to
stable-2.6. All existing older proxies will continue to send requests to
this destination container, until the data plane is restarted.
On restart, the new pods will start forwarding their requests to the new
linkerd-dst service.
* Use the 2.6 destination service fqdn
* Fixed unit tests
* Fix integration test failure
Signed-off-by: Ivan Sim <ivan@buoyant.io>
### Motivation
In order to expose arbitrary headers through tap, headers and trailers should be
read from the linkerd2-proxy-api `TapEvent`s and set in the public `TapEvent`s.
This change should have no user facing changes as it just prepares the events
for JSON output in linkerd/linkerd2#3390
### Solution
The public API has been updated with a headers field for
`TapEvent_Http_RequestInit_` and `TapEvent_Http_ResponseInit_`, and trailers
field for `TapEvent_Http_ResponseEnd_`.
These values are set by reading the corresponding fields off of the proxy's tap
events.
The proto changes are equivalent to the proto changes proposed in
linkerd/linkerd2-proxy-api#33
Closes#3262
Signed-off-by: Kevin Leimkuhler <kleimkuhler@icloud.com>
* Add the tracing environment variables to the proxy spec
* Add tracing event
* Remove unnecessary CLI change
* Update log message
* Handle single segment service name
* Use default service account if not provided
The injector doesn't read the defaults from the values.yaml
* Remove references to conf.workload.ownerRef in log messages
This nested field isn't always set.
Signed-off-by: Ivan Sim <ivan@buoyant.io>
This reverts commit edd3b1f6d4.
This is a temporary revert of #3461 while we sort out some details of how this should configured and how it should interact with configuring a trace collector on the Linkerd proxy. We will reintroduce this change once the config plan is straightened out.
Signed-off-by: Alex Leong <alex@buoyant.io>
* make identity use grpc server with prom metrics
Signed-off-by: Tarun Pothulapati <tarunpothulapati@outlook.com>
* linting fix
Signed-off-by: Tarun Pothulapati <tarunpothulapati@outlook.com>
* Fix injector timeout under high load
Fixes#3358
When retrieving a pod owner, we were hitting the k8s API directly because
at injection time the informer might not have been informed about the
existence of the parent object.
Under a large number of injection requests this ended up in the k8s API requests
being throttled, the proxy-injector getting blocked and the webhook requests
timing out.
Now we'll hit the shared informer first, and hit the k8s API only when
the informer doesn't return anything. After a few injection requests for
the same owner, the informer should have been updated.
Testing:
Scaling an emoji deployment to 1000 replicas, and after waiting for a
couple of minutes:
Before:
```bash
# a portion of the pods doesn't get injected
$ kubectl-n emojivoto get po | grep ./1 | wc -l
109
kubectl -n kube-system logs -f kube-apiserver-minikube | grep
failing.*timeout
.... (lots of errors)
```
After:
```bash
# all the pods get injected
$ kubectl -n emojivoto get po | grep ./1 | wc -l
0
kubectl -n kube-system logs -f kube-apiserver-minikube | grep
failing.*timeout
```
The repo depended on an old version of client-go. It also depended on
stern, which itself depended on an old version of client-go, making
client-go upgrade non-trivial.
Update the repo to client-go v12.0.0, and also replace stern with a
fork.
This fork of stern includes the following changes:
- updated to use Go Modules
- updated to use client-go v12.0.0
- fixed log line interleaving:
- https://github.com/wercker/stern/issues/96
- based on:
- 8723308e46Fixes#3382
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
The controller Docker image included 7 Go binaries (destination,
heartbeat, identity, proxy-injector, public-api, sp-validator, tap),
each roughly 35MB, with similar dependencies.
Change each controller binary into subcommands of a single `controller`
binary, decreasing the controller Docker image size from 315MB to 38MB.
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
Fixes#3356
1.16 removes some api groups that were already deprecated. From k8s blog
post (https://kubernetes.io/blog/2019/07/18/api-deprecations-in-1-16/):
```
- PodSecurityPolicy: will no longer be served from extensions/v1beta1 in
v1.16.
Migrate to the policy/v1beta1 API, available since v1.10. Existing
persisted data can be retrieved/updated via the policy/v1beta1 API.
- DaemonSet, Deployment, StatefulSet, and ReplicaSet: will no longer be
served from extensions/v1beta1, apps/v1beta1, or apps/v1beta2 in v1.16.
Migrate to the apps/v1 API, available since v1.9. Existing persisted
data can be retrieved/updated via the apps/v1 API.
```
Previous PRs had already made this change at the Helm templates level,
but we still needed to do it at the API calls and tests.
The integration tests ran fine for k8s 1.12 and 1.15. They fail on 1.16
because the upgrade integration test tries to install linkerd 2.5 which is not
compatible with 1.16.
Signed-off-by: Alejandro Pedraza <alejandro@buoyant.io>
The repo depended on a (recently broken) thrift package:
```
github.com/linkerd/linkerd2
-> contrib.go.opencensus.io/exporter/ocagent@v0.2.0
-> go.opencensus.io@v0.17.0
-> git.apache.org/thrift.git@v0.0.0-20180902110319-2566ecd5d999
```
... via this line in `controller/k8s`:
```go
_ "k8s.io/client-go/plugin/pkg/client/auth"
```
...which created a dependency on go.opencensus.io:
```bash
$ go mod why go.opencensus.io
...
github.com/linkerd/linkerd2/controller/k8s
k8s.io/client-go/plugin/pkg/client/auth
k8s.io/client-go/plugin/pkg/client/auth/azure
github.com/Azure/go-autorest/autorest
github.com/Azure/go-autorest/tracing
contrib.go.opencensus.io/exporter/ocagent
go.opencensus.io
```
Bump contrib.go.opencensus.io/exporter/ocagent from `v0.2.0` to
`v0.6.0`, creating this new dependency chain:
```
github.com/linkerd/linkerd2
-> contrib.go.opencensus.io/exporter/ocagent@v0.6.0
-> google.golang.org/api@v0.7.0
-> go.opencensus.io@v0.21.0
```
Bumping our go.opencensus.io dependency from `v0.17.0` to `v0.21.0`
pulls in this commit:
ed3a3f0bf0 (diff-37aff102a57d3d7b797f152915a6dc16)
...which removes our dependency on github.com/apache/thrift
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
* Avoid the dashboard requesting stats when not needed
Create an alternative to `urlsForResource` called
`urlsForResourceNoStats` that makes use of the `skip_stats` parameter in
the stats API (created in #1871) that doesn't query Prometheus when not needed.
When testing using the dashboard looking at the linkerd namespace,
queries per second went down from 2874 to 2756, a 4% decrease.
Signed-off-by: Alejandro Pedraza <alejandro@buoyant.io>
* Fix auto-injecting pods and integration tests reporting
When creating an Event when auto-injection occurs (#3316) we try to
fetch the parent object to associate the event to it. If the parent
doesn't exist (like in the case of stand-alone pods) the event isn't
created. I had missed dealing with one part where that parent was
expected.
This also adds a new integration test that I verified fails before this
fix.
Finally, I removed from `_test-run.sh` some `|| exit_code=$?` that was
preventing the whole suite to report failure whenever one of the tests
in `/tests` failed.
Signed-off-by: Alejandro Pedraza <alejandro@buoyant.io>
* Set custom cluster domain in GetServiceProfileFor
* Set custom cluster domain in tap server
Move fetching cluster domain for tap server to cmd main
* Handle fetchting cluster domain errors separately
* Use custom cluster domain for traffic split adaptor
Signed-off-by: Armin Buerkle <armin.buerkle@alfatraining.de>
* Have the proxy-injector emit events upon injection/skipping injection
Fixes#3253
Have the proxy-injector emit an event whenever a injection happens, or
when injection is skipped for some reason (also added that reason into
the proxy-injector logs). The level is associated to the parent workload
(it can't be associated to the pod because at this point the pod hasn't
been persisted).
The event recorder was setup at the `webhook/server.go` level and passed
to the proxy-injector's `Inject` function. The sp-validator thus also
has access to the event recorder, but for now it's not using it.
Related changes:
- Refactored `api.GetOwnerKindAndName()` to have it return a more
generic object.
- Refactored `report.Injectable()` to also have it return the reason why
a workload is not injectable.
Signed-off-by: Alejandro Pedraza <alejandro@buoyant.io>
Fixes#3052.
Adds a unit test for the edges API endpoint. To maintain a consistent order for
testing, the returned rows in api/public/edges.go are now sorted.
In preparation for #3242, the destination controller will need to
support a broader set of valid authorities including IP addresses.
This change modifies the destination controller's authority-parsing code
so that the is-this-a-kubernete-service-name decision is decoupled from
parsing of authorities into their consituent parts.
The `Get` API now explicitly handles IP address names, though it
currently fails all such resolutions.
* Rename template-values.go
* Define new constructor of charts.Values type
* Move all Helm values related code to the pkg/charts package
* Bump dependency
* Use '/' in filepath to remain compatible with VFS requirement
* Add unit test to verify Helm YAML output
* Alejandro's feedback
* Add unit test for Helm YAML validation (HA)
Signed-off-by: Ivan Sim <ivan@buoyant.io>
### Summary
As an initial attempt to secure the connection from clients to the gRPC tap
server on the tap Pod, the tap `addr` only listened on localhost.
As @adleong pointed out #3257, this was not actually secure because the inbound
proxy would establish a connection to localhost anyways.
This change removes the gRPC tap server listener and changes `TapByResource`
requests to interface with the server object directly.
From this, we know that all `TapByResourceRequests` have gone through the tap
APIServer and thus authorized by RBAC.
### Details
[NewAPIServer](ef90e0184f/controller/tap/apiserver.go (L25-L26)) now takes a [GRPCTapServer](f6362dfa80/controller/tap/server.go (L33-L34)) instead of a `pb.TapClient` so that
`TapByResource` requests can interact directly with the [TapByResource](f6362dfa80/controller/tap/server.go (L49-L50)) method.
`GRPCTapServer.TapByResource` now makes a private [grpcTapServer](ef90e0184f/controller/tap/handlers.go (L373-L374)) that satisfies
the [tap.TapServer](https://godoc.org/github.com/linkerd/linkerd2/controller/gen/controller/tap#TapServer) interface. Because this interface is satisfied, we can interact
with the tap server methods without spawning an additional listener.
Signed-off-by: Kevin Leimkuhler <kleimkuhler@icloud.com>
This PR adds a test for trafficsplits to stat_summary_test.go.
Because the test requires a consistent order for returned rows, trafficsplit
rows in stat_summary.go are now sorted by apex + leaf name before being
returned.
### Summary
After the addition of the tap APIServer, all the logic related to tap in the public API no longer needs to be there. The servers and clients that are created but not used, as well as all the old testing infrastrucure related to tap can be removed.
This deprecates TapByResource and therefore required an update to the protobuf files with `bin/protoc-go.sh`. While the change to deprecate this method was extremely small, a lot of protobuf fils were updated in the process. These changes to the code and protobuf files should probably remain coupled since `TapByResource` is officially deprecated in the public API, but a majority of the additions/deletions are related to those files.
This draft passes `go test` as well as a local run of the integration tests.
Signed-off-by: Kevin Leimkuhler <kleimkuhler@icloud.com>
PR #3217 re-introduced container metrics collection to
linkerd-prometheus. This enabled linkerd-heartbeat to collect mem and
cpu metrics at the container-level.
Add container cpu and mem metrics to heartbeat requests. For each of
(destination, prometheus, linkerd-proxy), collect maximum memory and p95
cpu.
Concretely, this introduces 7 new query params to heartbeat requests:
- p99-handle-us
- max-mem-linkerd-proxy
- max-mem-destination
- max-mem-prometheus
- p95-cpu-linkerd-proxy
- p95-cpu-destination
- p95-cpu-prometheus
Part of #2961
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
This PR adds `trafficsplit` as a supported resource for the `linkerd stat` command. Users can type `linkerd stat ts` to see the apex and leaf services of their trafficsplits, as well as metrics for those leaf services.
Go dependencies which are only used by generated code had not previously been checked into the repo. Because `go generate` does not respect the `-mod=readonly` flag, running `bin/linkerd` will add these dependencies and dirty the local repo. This can interfere with the way version tags are generated.
To avoid this, we simply check these deps in.
Note that running `go mod tidy` will remove these again. Thus, it is not recommended to run `go mod tidy`.
Signed-off-by: Alex Leong <alex@buoyant.io>
### Summary
Changes from `bin/protoc-go.sh`
An existing [draft PR](https://github.com/linkerd/linkerd2/pull/3240) has a majority of its changes related to protobuf file
updates. In order to separate these changes out into more related components,
this PR updates the generated protobuf files so that #3240 can be rebased off
this and have a more manageable diff.
Signed-off-by: Kevin Leimkuhler <kleimkuhler@icloud.com>
### Motivation
PR #3167 introduced the tap APIService and migrated `linkerd tap` to use it.
Subsequent PRs (#3186 and #3187) updated `linkerd top` and `linkerd profile
--tap` to use the tap APIService. This PR moves the web's Go server to now also
use the tap APIService instead of the public API. It also ensures an error
banner is shown to the user when unauthorized taps fail via `linkerd top`
command in *Overview* and *Top*, and `linkerd tap` command in *Tap*.
### Details
The majority of these changes are focused around piping through the HTTP error
that occurs and making sure the error banner generated displays the error
message explaining to view the tap RBAC docs.
`httpError` is now public (`HTTPError`) and the error message generated is short
enough to fit in a control frame (explained [here](https://github.com/linkerd/linkerd2/blob/kleimkuhler%2Fweb-tap-apiserver/web/srv/api_handlers.go#L173-L175)).
### Testing
The error we are testing for only occurs when the linkerd-web service account is
not authorzied to tap resources. Unforutnately that is not the case on Docker
For Mac (assuming that is what you use locally), so you'll need to test on a
different cluster. I chose a GKE cluster made through the GKE console--not made
through cluster-utils because it adds cluster-admin.
Checkout the branch locally and `bin/docker-build` or `ares-build` if you have
it setup. It should produce a linkerd with the version `git-04e61786`. I have
already pushed the dependent components, so you won't need to `bin/docker-push
git-04e61786`.
Install linkerd on this GKE cluster and try to run `tap` or `top` commands via
the web. You should see the following errors:
### Tap
### Top
Signed-off-by: Kevin Leimkuhler <kleimkuhler@icloud.com>
PR #3154 introduced an `l5d-require-id` header to Tap requests. That
header string was constructed based on the TapByResourceRequest, which
includes 3 notable fields (type, name, namespace). For namespace-level
requests (via commands like `linkerd tap ns linkerd`), type ==
`namespace`, name == `linkerd`, and namespace == "". This special casing
for namespace-level requests yielded invalid `l5d-require-id` headers,
for example: `pd-sa..serviceaccount.identity.linkerd.cluster.local`.
Fix `l5d-require-id` string generation to account for namespace-level
requests. The bulk of this change is tap unit test updates to validate
the fix.
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
* Refactor proxy injection to use Helm charts
Fixes#3128
A new chart `/charts/patch` was created, that generates the JSON patch
payload that is to be returned to the k8s API when doing the injection
through the proxy injector, and it's also leveraged by the `linkerd
inject --manual` CLI.
The VFS was used by `linkerd install` to access the old chart under
`/chart`. Now the proxy injection also uses the Helm charts to generate
the JSON patch (see above) so we've moved the VFS from `cli/static` to a
new common place under `/pkg/charts/static`, and the new root for the VFS is
now `/charts`.
`linkerd install` hasn't yet migrated to use the new charts (that'll
happen in #3127), so the only change in that regard was the creation of
`/charts/chart` which is a symlink pointing to `/chart` that
`install.go` now uses, so that the VFS contains both the old and new
charts, as a temporary measure.
You can see that `/bin/Dockerfile-bin`, `/controller/Dockerfile` and
`/bin/build-cli-bin` do now `go generate` pointing to the new location
(and the `go generate` annotation was moved from `/cli/main.go` to
`pkg/charts/static/templates.go`).
The symlink trick doesn't work when building the binaries through
Docker, so `/bin/Dockerfile-bin` replaces the symlink with an actual
copy of `/chart`.
Also note that in `/controller/Dockerfile` we now need to include the
`prod` tag in `go install` like we do in `/bin/Dockerfile-bin` so that
the proxy injector does use the VFS instead of the local file system.
- The common logic to parse a chart has been moved from `install.go` to
`/pkg/charts/util.go`.
- The special ENV var in the proxy for "outbound router capacity" that
only applies to the Prometheus pod is now handled directly in the proxy
partial and all the associated go code could be removed.
- The `patch.go` lib for generating the JSON patch in go along
with its tests `patch_test.go` are no longer needed.
- Lots of functions in `/pkg/inject/inject.go` got removed/simplified
with their logic being moved into the charts themselves. As a
consequence lots of things in `inject_test.go` became irrelevant.
- Moved `template-values.go` from `/pkg/inject` to `pkg/charts` as that
contains the go structs representation of the chart variables that
will be leveraged in #3127.
Don't forget to run `/bin/helm.sh` whenever you make changes to charts
;-)
Signed-off-by: Alejandro Pedraza <alejandro@buoyant.io>
The `/apis/tap.linkerd.io/v1alpha1` endpoint on the Tap APIServer
included tap resource/subresource pairs, such as `deployments/tap` and
`pods/tap`, but did not include the parent resources, such as
`deployments` and `pods`. This broke commands like `kubectl auth can-i`,
which expect the parent resource to exist.
Introduce parent resources for all tap-able subresources. Concretely, it
fixes this class of command:
```
$ kubectl auth can-i watch deployments.tap.linkerd.io/linkerd-grafana \
--subresource=tap -n linkerd --as siggy@buoyant.io
Warning: the server doesn't have a resource type 'deployments' in group 'tap.linkerd.io'
no - no RBAC policy matched
```
Fixed:
```
$ kubectl auth can-i watch deployments.tap.linkerd.io/linkerd-grafana \
--subresource=tap -n linkerd --as siggy@buoyant.io
yes
```
Additionally, when SubjectAccessReviews fail, return a 403 rather than
500.
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
The Tap Service enabled tapping of any meshed pod, regardless of user
privilege.
This change introduces a new Tap APIService. Kubernetes provides
authentication and authorization of Tap requests, and then forwards
requests to a new Tap APIServer, which implements a Kubernetes
aggregated APIServer. The Tap APIServer authenticates the client TLS
from Kubernetes, and authorizes the user via a SubjectAccessReview.
This change also modifies the `linkerd tap` command to make requests
against the new APIService.
The Tap APIService implements these Kubernetes-style endpoints:
POST /apis/tap.linkerd.io/v1alpha1/watch/namespaces/:ns/tap
POST /apis/tap.linkerd.io/v1alpha1/watch/namespaces/:ns/:res/:name/tap
GET /apis
GET /apis/tap.linkerd.io
GET /apis/tap.linkerd.io/v1alpha1
GET /healthz
GET /healthz/log
GET /healthz/ping
GET /metrics
GET /openapi/v2
GET /version
Users authorize to the new `tap.linkerd.io/v1alpha1` via RBAC. Only the
`watch` verb is supported. Access is also available via subresources
such as `deployments/tap` and `pods/tap`.
This change introduces the following resources into the default Linkerd
install:
- Global
- APIService/v1alpha1.tap.linkerd.io
- ClusterRoleBinding/linkerd-linkerd-tap-auth-delegator
- `linkerd` namespace:
- Secret/linkerd-tap-tls
- `kube-system` namespace:
- RoleBinding/linkerd-linkerd-tap-auth-reader
Tasks not covered by this PR:
- `linkerd top`
- `linkerd dashboard`
- `linkerd profile --tap`
- removal of the unauthenticated tap controller
Fixes#2725, #3162, #3172
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
Similar to `kubectl --as`, global flag across all linkerd subcommands
which sets a `ImpersonationConfig` in the Kubernetes API config.
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
### Summary
In order for Pods' tap servers to start authorizing tap clients, the tap
controller must open TLS connections so that it can identity itself to the
server.
This change introduces the use of `l5d-require-id` header on outbound tap
requests.
### Details
When tap requests are made by the tap controller, the `Authority` header is an
IP address. The proxy does not attempt to do service discovery on such requests
and therefore the connection is over plaintext. By introducing the
`l5d-require-id` header the proxy can require a server name on the connection.
This allows the tap controller to identity itself as the client making tap
requests. The name value for the header can be made from the Pod Spec and tap
request, so the change is rather minimal.
#### Proxy Changes
* Update h2 to v0.1.26
* Properly fall back in the dst_router (linkerd/linkerd2-proxy#291)
### Testing
Unit tests for the header have not been added mainly because [no test
infrastructure currently exists](065c221858/controller/tap/server_test.go (L241)) to mock proxy requests. After talking with
@siggy a little about this, it makes to do in a separate change at some point
when behavior like this cannot be reliably tested through integration tests
either.
Integration tests do test this well, and will continue to do once
linkerd/linkerd2-proxy#290 lands.
Signed-off-by: Kevin Leimkuhler <kleimkuhler@icloud.com>
Fixes https://github.com/linkerd/linkerd2/issues/2800#issuecomment-513740498
When the Linkerd proxy sends a query for a Kubernetes external name service to the destination service, the destination service returns `NoEndpoints: exists=false` because an external name service has no endpoints resource. Due to a change in the proxy's fallback logic, this no longer causes the proxy to fallback to either DNS or SO_ORIG_DST and instead fails the request. The net effect is that Linkerd fails all requests to external name services.
We change the destination service to instead return `InvalidArgument` for external name services. This causes the proxy to fallback to SO_ORIG_DST instead of failing the request.
Signed-off-by: Alex Leong <alex@buoyant.io>
The destination service's endpoints watcher assumed every `Endpoints`
object contained a `TargetRef`. This field is optional, and in cases
such as the default `ep/kubernetes` object, `TargetRef` is nil, causing
a nil pointer dereference.
Fix endpoints watcher to check for `TargetRef` prior to dereferencing.
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
The repo relied on `dep` for managing Go dependencies. Go 1.11 shipped
with Go modules support. Go 1.13 will be released in August 2019 with
module support enabled by default, deprecating GOPATH.
This change replaces `dep` with Go modules for dependency management.
All scripts, including Docker builds and ci, should work without any dev
environment changes.
To execute `go` commands directly during development, do one of the
following:
1. clone this repo outside of `GOPATH`; or
2. run `export GO111MODULE=on`
Summary of changes:
- Docker build scripts and ci set `-mod=readonly`, to ensure
dependencies defined in `go.mod` are exactly what is used for the
builds.
- Dependency updates to `go.mod` are accomplished by running
`go build` and `go test` directly.
- `bin/go-run`, `bin/build-cli-bin`, and `bin/test-run` set
`GO111MODULE=on`, permitting usage inside and outside of GOPATH.
- `gcr.io/linkerd-io/go-deps` tags hashed from `go.mod`.
- `bin/update-codegen.sh` still requires running from GOPATH,
instructions added to BUILD.md.
Fixes#1488
Signed-off-by: Andrew Seigner <siggy@buoyant.io>
Fixes#3136
When the destination service sends a destination profile with a traffic split to the proxy, the override destination authorities are absolute but do no contain a trailing dot. e.g. "bar.ns.svc.cluster.local:80". However, NameAddrs which have undergone canonicalization in the proxy will include the trailing dot. When a traffic split includes the apex service as one of the overrides, the original apex NameAddr will have the trailing dot and the override will not. Since these two NameAddrs are not identical, they will go into two distinct slots in the proxy's concrete dst router. This will cause two services to be created for the same destination which will cause the stats clobbering described in the linked issue.
We change the destination service to always return absolute dst overrides including the trailing dot.
Signed-off-by: Alex Leong <alex@buoyant.io>
Alejandro Pedraza
Ivan Sim
Kevin Leimkuhler
Alex Leong
Tarun Pothulapati
Andrew Seigner
Andrew Seigner
Bruno M. Custódio
陈谭军
arminbuerkle
Carol A. Scott
Oliver Gould
Guangming Wang
cpretzer
