This project contains Kubernetes controllers for managing external ingress with AWS or IPVS load balancers.
There are two controllers provided,
feed-ingress which runs an NGINX instance, and
feed-dns which manages Amazon Route 53 entries.
They can be run independently as needed, or together to provide a full ingress solution.
feed-ingress can be arbitrarily scaled up to support any traffic load.
Feed is actively used in production and should be stable enough for general usage. We can scale to many thousands of requests per second with only a handful of replicas.
Comparison to official NGINX ingress controller
Feed was started before the official NGINX ingress controller became production ready. The main differences that exist now are:
- Feed has fewer features, as we only built it for our needs.
- Feed pods attach directly to AWS load balancers or IPVS nodes. The official controller relies on the
LoadBalancerservice type, which generally forwards traffic to every node in your cluster (
service.spec.externalTrafficPolicycan be set in some providers to mitigate this). We found this problematic:
- It increases the amount of traffic flowing through your cluster, as traffic is routed through every node unnecessarily.
- ELB health checks don't work - the ELBs will disable arbitrary nodes, rather than a broken ingress pod.
- Feed uses services, while the official controller uses endpoints:
- Primarily to reduce the number of NGINX reloads that occur, which are problematic in busy environments. It may be possible to mitigate this though with a dynamic update of NGINX (via plugin), and is something we've discussed doing for service updates.
- It's debatable whether using endpoints directly is a good idea conceptually, as it bypasses kube-proxy and any service mesh in place.
Docker images for
feed-dns are released using semantic versioning.
See the examples for deployment YAML files that can be applied to a cluster.
The following RBAC permissions are required by the service account under which feed runs:
rules: - apiGroups: - "" - extensions resources: - ingresses - namespaces - services verbs: - get - list - watch - apiGroups: - extensions resources: - ingresses/status verbs: - update
feed-ingress with AWS load balancers, the following are required:
- An internal and internet-facing load balancer which can reach your Kubernetes cluster.
The load balancers should be tagged with
sky.uk/KubernetesClusterFrontend=<name>which is used by feed to discover them. If you are using v2 of
feed-ingressthey should also be tagged with
sky.uk/KubernetesClusterIngressClass=<name>. See upgrading from v1 to v2 for more information.
feed-dns, a Route 53 hosted zone to match your ingress resources.
feed-ingress manages an NGINX instance, updating its configuration dynamically for ingress resources. It attaches to
ELBs which are intended to be the frontend for all traffic.
See the command line options with:
docker run --rm skycirrus/feed-ingress:v2.0.0 -h
- NGINX reloads can be disruptive. On reload, NGINX will finish in-flight requests, then abruptly
close all server connections. This is a limitation of NGINX, and affects all NGINX solutions. We mitigate this by:
- Rate limiting reloads. This is user configurable.
- Using service IPs, which are stable. Reloads will only happen if an ingress or service changes, which is rare compared to pod changes.
Upgrading from v1 to v2
This is a breaking change to support multiple ingress controllers per cluster. The feed-ingress command-line structure has also changed. There are subcommands for the various load balancer types and arguments use double dashes to be POSIX-compliant.
To upgrade, follow these steps:
- Tag the ELBs with
sky.uk/KubernetesClusterIngressClass=<name>to indicate which feed-ingress controllers should attach to them
- Annotate all ingresses with
- Replace deprecated ingress resource annotations with their replacements:
- Use double dashes for all arguments
- Provide the mandatory argument
--ingress-class=<name>to feed-ingress with a value matching the ELB tag. For migrating existing deployments, you may provide the new (but deprecated) flag
--include-classless-ingresseswhich instructs feed-ingress to additionally consider ingress resources that have no
- Instead of using the argument
-registration-frontend-type, use instead a subcommand of
feed-ingress -registration-frontend-type=elb <args...>becomes
feed-ingress elb <args>)
- Rename the argument
- Rename the argument
SSL termination on load balancer
SSL termination could be done on ELBs, and we believe that this is the safest and best performing
approach for production usage. Unfortunately, ELBs don't support SNI at this time, so this limits SSL usage to
a single domain. One workaround is to use a wildcard certificate for the entire zone that
Another is to place an SSL termination EC2 instance in front of the ELBs.
SSL termination on feed-ingress
SSL termination can be done on feed-ingress. This approach still requires a layer 4 load balancer, eg. ELB or IPVS, in front.
For the moment you can setup a default wildcard SSL:
# Set default SSL path + name file without extension. Feed expects two files: one ending in .crt (the CA) # and the other in .key (the private key), for example: --ssl-path=/etc/ssl/default-ssl/default-ssl # will cause feed to look for /etc/ssl/default-ssl.crt and /etc/ssl/default.ssl.key
You can mount the
.crt though a Kubernetes Secret see feed-ingress-deployment-ssl.
Merlin is a distributed load balancer based on IPVS, with a gRPC based API. Feed supports attaching to merlin as a frontend for ingress.
See the example for details.
GORB / IPVS Support
GORB support is deprecated, and will be removed at some point. Use merlin instead.
Feed has support for configuring IPVS via GORB. GORB exposes a REST api to interrogate and modify the IPVS configuration such as virtual services and backends. The configuration can be stored in a distributed key/value store.
Although IPVS supports multiple packet-forwarding methods, feed currently only supports 'DR' aka Direct Server Return.
It provides the ability to manage the loopback interface so the ingress instance can pretend to be IPVS at the IP level.
feed-ingress pod will need to define the
NET_ADMIN Linux capability to be able to manage the loopback interface.
securityContext: capabilities: add: - NET_ADMIN
See the example deployment for GORB
AWS load balancer support
Feed supports the three types of load balancer offered by AWS.
Classic Load Balancers (ELBs)
See the example deployment for ELB
Network Load Balancers (NLBs)
See the example deployment for NLB
Please note that when a new Feed pod is added as a target to the target group of an NLB, it takes some time for the new target to be registered and health-checked by the NLB before the NLB will send it any requests. Please see the "Target Health Status" section of the AWS documentation on NLB target group health checks for more details.
Application Load Balancers (ALBs)
Feed has support for ALBs. Unfortunately, ALBs have a bug that prevents non-disruptive deployments of feed (specifically, they don't respect the deregistration delay). As a result, we don't recommend using ALBs at this time.
The build now includes support for OpenTracing, and the default Docker image includes the Jaeger tracing vendor implementation.
To enable OpenTracing, you will need to provide the following options:
# Define the path to the OpenTracing vendor plugin --nginx-opentracing-plugin-path=/usr/local/lib/libjaegertracing_plugin.linux_amd64.so # Define the path to the config for the vendor plugin --nginx-opentracing-config-path=/etc/jaeger-nginx-config.json
Note that the status and metrics endpoints will not have OpenTracing applied.
Handling large client requests
feed-ingress now supports handling of large client requests (header and body). The following are the default values for the same.
client_header_buffer_size 16k; client_body_buffer_size 16k; large_client_header_buffers 4 16k
They can be overridden by passing the following arguments during startup.
--nginx-client-header-buffer-size-in-kb=16 --nginx-client-body-buffer-size-in-kb=16 # Set the maximum number and size of buffers used for reading large client request header. If this value is set, # --nginx-client-header-buffer-size-in-kb should be passed in as well. Otherwise, this value will be ignored. --nginx-large-client-header-buffer-blocks=4
When using the ELB, NLB or Merlin updaters, the ingress status will be updated with relevant
load balancer information. This can then be used with other controllers such as
external-dns which can set DNS for any
given ingress using the ingress status.
AWS load balancers
Feed will automatically discover all of your load balancers and then use the
sky.uk/frontend-scheme annotation to
match a load balancer label to an ingress. The updater will then set the ingress status to the load balancer's DNS name.
The Merlin updater is currently unable to auto-discover all hosted load balancers on a Merlin server; instead the status
updater supports two different types:
internet-facing. These two load balancers are set using the
merlin-internet-facing-hostname flags respectively.
An ingress can select which load balancer it wants to be associated with by setting the
annotation to either
Running feed-ingress on privileged ports
feed-ingress can be run on privileged ports by defining the
NET_BIND_SERVICE Linux capability.
securityContext: capabilities: add: - NET_BIND_SERVICE
Multiple ingress controllers per cluster
Multiple feed-ingress controllers can be created per cluster. Load balancers should be tagged with
sky.uk/KubernetesClusterIngressClass=<name> and feed instances started with
Feed instances will attach to load balancers with matching ingress class names.
A feed ingress controller will adopt ingress resources with a matching
Ingress resources with no annotation will normally not be adopted and will have no traffic sent to their associated services.
However, see the deprecated flag
--include-classless-ingresses which instructs feed-ingress to additionally consider
ingress resources with no
Use the script
classless-ingresses.sh to find ingresses without this annotation.
This feature is supported by
feed-ingress and the
nlb load balancer. It is currently not supported by
or any other load balancer type. PRs are welcome.
feed-dns manages a Route 53 hosted zone, updating entries to point to ELBs or arbitrary hostnames. It is designed to
be run as a single instance per zone in your cluster.
See the command line options with:
docker run skycirrus/feed-dns:v2.0.0 -h
The feed-dns controller assumes that it can overwrite any entry in the supplied DNS zone and manages ALIAS and CNAME records per ingress.
On startup, all ingress entries are queried and compared to all the Record Sets in the configured hosted zone.
Any records pointing to one of the endpoints associated with this controller that do not have an ingress entry are deleted. For any new ingress entry, a record is created to point to the correct endpoint. Existing records which do not meet these conditions remain untouched.
Each ingress must have the following be annotated with
sky.uk/frontend-scheme set to
so the record can be set to the correct endpoint.
If you're using ELBs then ALIAS (A) records will be created. If you've explicitly provided CNAMEs of your load balancers then CNAMEs will be created.
feed-dnsonly supports a single hosted zone at this time, but this should be straightforward to add support for. PRs are welcome.
Install the required tools and setup dependencies:
Build and test with:
Tag the commit in master and push it to release it. Only maintainers can do this.
Dependencies are managed with Go Modules.