Traffic Management Problems

Requests are rejected by Envoy

Requests may be rejected for various reasons. The best way to understand why requests are being rejected is by inspecting Envoy’s access logs. By default, access logs are output to the standard output of the container. Run the following command to see the log:

  1. $ kubectl logs PODNAME -c istio-proxy -n NAMESPACE

In the default access log format, Envoy response flags are located after the response code, if you are using a custom log format, make sure to include %RESPONSE_FLAGS%.

Refer to the Envoy response flags for details of response flags.

Common response flags are:

  • NR: No route configured, check your DestinationRule or VirtualService.
  • UO: Upstream overflow with circuit breaking, check your circuit breaker configuration in DestinationRule.
  • UF: Failed to connect to upstream, if you’re using Istio authentication, check for a mutual TLS configuration conflict.

Route rules don’t seem to affect traffic flow

With the current Envoy sidecar implementation, up to 100 requests may be required for weighted version distribution to be observed.

If route rules are working perfectly for the Bookinfo sample, but similar version routing rules have no effect on your own application, it may be that your Kubernetes services need to be changed slightly. Kubernetes services must adhere to certain restrictions in order to take advantage of Istio’s L7 routing features. Refer to the Requirements for Pods and Services for details.

Another potential issue is that the route rules may simply be slow to take effect. The Istio implementation on Kubernetes utilizes an eventually consistent algorithm to ensure all Envoy sidecars have the correct configuration including all route rules. A configuration change will take some time to propagate to all the sidecars. With large deployments the propagation will take longer and there may be a lag time on the order of seconds.

503 errors after setting destination rule

You should only see this error if you disabled automatic mutual TLS during install.

If requests to a service immediately start generating HTTP 503 errors after you applied a DestinationRule and the errors continue until you remove or revert the DestinationRule, then the DestinationRule is probably causing a TLS conflict for the service.

For example, if you configure mutual TLS in the cluster globally, the DestinationRule must include the following trafficPolicy:

  1. trafficPolicy:
  2.   tls:
  3.     mode: ISTIO_MUTUAL

Otherwise, the mode defaults to DISABLE causing client proxy sidecars to make plain HTTP requests instead of TLS encrypted requests. Thus, the requests conflict with the server proxy because the server proxy expects encrypted requests.

Whenever you apply a DestinationRule, ensure the trafficPolicy TLS mode matches the global server configuration.

Route rules have no effect on ingress gateway requests

Let’s assume you are using an ingress Gateway and corresponding VirtualService to access an internal service. For example, your VirtualService looks something like this:

  1. apiVersion: networking.istio.io/v1beta1
  2. kind: VirtualService
  3. metadata:
  4.   name: myapp
  5. spec:
  6.   hosts:
  7.   - "myapp.com" # or maybe "*" if you are testing without DNS using the ingress-gateway IP (e.g., http://1.2.3.4/hello)
  8.   gateways:
  9.   - myapp-gateway
  10.   http:
  11.   - match:
  12.     - uri:
  13.         prefix: /hello
  14.     route:
  15.     - destination:
  16.         host: helloworld.default.svc.cluster.local
  17.   - match:
  18.     ...

You also have a VirtualService which routes traffic for the helloworld service to a particular subset:

  1. apiVersion: networking.istio.io/v1beta1
  2. kind: VirtualService
  3. metadata:
  4.   name: helloworld
  5. spec:
  6.   hosts:
  7.   - helloworld.default.svc.cluster.local
  8.   http:
  9.   - route:
  10.     - destination:
  11.         host: helloworld.default.svc.cluster.local
  12.         subset: v1

In this situation you will notice that requests to the helloworld service via the ingress gateway will not be directed to subset v1 but instead will continue to use default round-robin routing.

The ingress requests are using the gateway host (e.g., myapp.com) which will activate the rules in the myapp VirtualService that routes to any endpoint of the helloworld service. Only internal requests with the host helloworld.default.svc.cluster.local will use the helloworld VirtualService which directs traffic exclusively to subset v1.

To control the traffic from the gateway, you need to also include the subset rule in the myapp VirtualService:

  1. apiVersion: networking.istio.io/v1beta1
  2. kind: VirtualService
  3. metadata:
  4.   name: myapp
  5. spec:
  6.   hosts:
  7.   - "myapp.com" # or maybe "*" if you are testing without DNS using the ingress-gateway IP (e.g., http://1.2.3.4/hello)
  8.   gateways:
  9.   - myapp-gateway
  10.   http:
  11.   - match:
  12.     - uri:
  13.         prefix: /hello
  14.     route:
  15.     - destination:
  16.         host: helloworld.default.svc.cluster.local
  17.         subset: v1
  18.   - match:
  19.     ...

Alternatively, you can combine both VirtualServices into one unit if possible:

  1. apiVersion: networking.istio.io/v1beta1
  2. kind: VirtualService
  3. metadata:
  4.   name: myapp
  5. spec:
  6.   hosts:
  7.   - myapp.com # cannot use "*" here since this is being combined with the mesh services
  8.   - helloworld.default.svc.cluster.local
  9.   gateways:
  10.   - mesh # applies internally as well as externally
  11.   - myapp-gateway
  12.   http:
  13.   - match:
  14.     - uri:
  15.         prefix: /hello
  16.       gateways:
  17.       - myapp-gateway #restricts this rule to apply only to ingress gateway
  18.     route:
  19.     - destination:
  20.         host: helloworld.default.svc.cluster.local
  21.         subset: v1
  22.   - match:
  23.     - gateways:
  24.       - mesh # applies to all services inside the mesh
  25.     route:
  26.     - destination:
  27.         host: helloworld.default.svc.cluster.local
  28.         subset: v1

Envoy is crashing under load

Check your ulimit -a. Many systems have a 1024 open file descriptor limit by default which will cause Envoy to assert and crash with:

  1. [2017-05-17 03:00:52.735][14236][critical][assert] assert failure: fd_ != -1: external/envoy/source/common/network/connection_impl.cc:58

Make sure to raise your ulimit. Example: ulimit -n 16384

Envoy won’t connect to my HTTP/1.0 service

Envoy requires HTTP/1.1 or HTTP/2 traffic for upstream services. For example, when using NGINX for serving traffic behind Envoy, you will need to set the proxy_http_version directive in your NGINX configuration to be “1.1”, since the NGINX default is 1.0.

Example configuration:

  1. upstream http_backend {
  2.     server 127.0.0.1:8080;
  4.     keepalive 16;
  5. }
  7. server {
  8.     ...
  10.     location /http/ {
  11.         proxy_pass http://http_backend;
  12.         proxy_http_version 1.1;
  13.         proxy_set_header Connection "";
  14.         ...
  15.     }
  16. }

503 error while accessing headless services

Assume Istio is installed with the following configuration:

  • mTLS mode set to STRICT within the mesh
  • meshConfig.outboundTrafficPolicy.mode set to ALLOW_ANY

Consider nginx is deployed as a StatefulSet in the default namespace and a corresponding Headless Service is defined as shown below:

  1. apiVersion: v1
  2. kind: Service
  3. metadata:
  4.   name: nginx
  5.   labels:
  6.     app: nginx
  7. spec:
  8.   ports:
  9.   - port: 80
  10.     name: http-web  # Explicitly defining an http port
  11.   clusterIP: None   # Creates a Headless Service
  12.   selector:
  13.     app: nginx
  14. ---
  15. apiVersion: apps/v1
  16. kind: StatefulSet
  17. metadata:
  18.   name: web
  19. spec:
  20.   selector:
  21.     matchLabels:
  22.       app: nginx
  23.   serviceName: "nginx"
  24.   replicas: 3
  25.   template:
  26.     metadata:
  27.       labels:
  28.         app: nginx
  29.     spec:
  30.       containers:
  31.       - name: nginx
  32.         image: registry.k8s.io/nginx-slim:0.8
  33.         ports:
  34.         - containerPort: 80
  35.           name: web

The port name http-web in the Service definition explicitly specifies the http protocol for that port.

Let us assume we have a sleep pod Deployment as well in the default namespace. When nginx is accessed from this sleep pod using its Pod IP (this is one of the common ways to access a headless service), the request goes via the PassthroughCluster to the server-side, but the sidecar proxy on the server-side fails to find the route entry to nginx and fails with HTTP 503 UC.

  1. $ export SOURCE_POD=$(kubectl get pod -l app=sleep -o jsonpath='{.items..metadata.name}')
  2. $ kubectl exec -it $SOURCE_POD -c sleep -- curl 10.1.1.171 -s -o /dev/null -w "%{http_code}"
  3.   503

10.1.1.171 is the Pod IP of one of the replicas of nginx and the service is accessed on containerPort 80.

Here are some of the ways to avoid this 503 error:

  1. Specify the correct Host header:

    The Host header in the curl request above will be the Pod IP by default. Specifying the Host header as nginx.default in our request to nginx successfully returns HTTP 200 OK.

    1. $ export SOURCE_POD=$(kubectl get pod -l app=sleep -o jsonpath='{.items..metadata.name}')
    2. $ kubectl exec -it $SOURCE_POD -c sleep -- curl -H "Host: nginx.default" 10.1.1.171 -s -o /dev/null -w "%{http_code}"
    3.   200

  2. Set port name to tcp or tcp-web or tcp-<custom_name>:

    Here the protocol is explicitly specified as tcp. In this case, only the TCP Proxy network filter on the sidecar proxy is used both on the client-side and server-side. HTTP Connection Manager is not used at all and therefore, any kind of header is not expected in the request.

    A request to nginx with or without explicitly setting the Host header successfully returns HTTP 200 OK.

    This is useful in certain scenarios where a client may not be able to include header information in the request.

    1. $ export SOURCE_POD=$(kubectl get pod -l app=sleep -o jsonpath='{.items..metadata.name}')
    2. $ kubectl exec -it $SOURCE_POD -c sleep -- curl 10.1.1.171 -s -o /dev/null -w "%{http_code}"
    3.   200
    1. $ kubectl exec -it $SOURCE_POD -c sleep -- curl -H "Host: nginx.default" 10.1.1.171 -s -o /dev/null -w "%{http_code}"
    2.   200

  3. Use domain name instead of Pod IP:

    A specific instance of a headless service can also be accessed using just the domain name.

    1. $ export SOURCE_POD=$(kubectl get pod -l app=sleep -o jsonpath='{.items..metadata.name}')
    2. $ kubectl exec -it $SOURCE_POD -c sleep -- curl web-0.nginx.default -s -o /dev/null -w "%{http_code}"
    3.   200

    Here web-0 is the pod name of one of the 3 replicas of nginx.

Refer to this traffic routing page for some additional information on headless services and traffic routing behavior for different protocols.

TLS configuration mistakes

Many traffic management problems are caused by incorrect TLS configuration. The following sections describe some of the most common misconfigurations.

Sending HTTPS to an HTTP port

If your application sends an HTTPS request to a service declared to be HTTP, the Envoy sidecar will attempt to parse the request as HTTP while forwarding the request, which will fail because the HTTP is unexpectedly encrypted.

  1. apiVersion: networking.istio.io/v1beta1
  2. kind: ServiceEntry
  3. metadata:
  4.   name: httpbin
  5. spec:
  6.   hosts:
  7.   - httpbin.org
  8.   ports:
  9.   - number: 443
  10.     name: http
  11.     protocol: HTTP
  12.   resolution: DNS

Although the above configuration may be correct if you are intentionally sending plaintext on port 443 (e.g., curl http://httpbin.org:443), generally port 443 is dedicated for HTTPS traffic.

Sending an HTTPS request like curl https://httpbin.org, which defaults to port 443, will result in an error like curl: (35) error:1408F10B:SSL routines:ssl3_get_record:wrong version number. The access logs may also show an error like 400 DPE.

To fix this, you should change the port protocol to HTTPS:

  1. spec:
  2.   ports:
  3.   - number: 443
  4.     name: https
  5.     protocol: HTTPS

Gateway to virtual service TLS mismatch

There are two common TLS mismatches that can occur when binding a virtual service to a gateway.

  1. The gateway terminates TLS while the virtual service configures TLS routing.
  2. The gateway does TLS passthrough while the virtual service configures HTTP routing.

Gateway with TLS termination

  1. apiVersion: networking.istio.io/v1beta1
  2. kind: Gateway
  3. metadata:
  4.   name: gateway
  5.   namespace: istio-system
  6. spec:
  7.   selector:
  8.     istio: ingressgateway
  9.   servers:
  10.   - port:
  11.       number: 443
  12.       name: https
  13.       protocol: HTTPS
  14.     hosts:
  15.       - "*"
  16.     tls:
  17.       mode: SIMPLE
  18.       credentialName: sds-credential
  19. ---
  20. apiVersion: networking.istio.io/v1beta1
  21. kind: VirtualService
  22. metadata:
  23.   name: httpbin
  24. spec:
  25.   hosts:
  26.   - "*.example.com"
  27.   gateways:
  28.   - istio-system/gateway
  29.   tls:
  30.   - match:
  31.     - sniHosts:
  32.       - "*.example.com"
  33.     route:
  34.     - destination:
  35.         host: httpbin.org

In this example, the gateway is terminating TLS while the virtual service is using TLS based routing. The TLS route rules will have no effect since the TLS is already terminated when the route rules are evaluated.

With this misconfiguration, you will end up getting 404 responses because the requests will be sent to HTTP routing but there are no HTTP routes configured. You can confirm this using the istioctl proxy-config routes command.

To fix this problem, you should switch the virtual service to specify http routing, instead of tls:

  1. spec:
  2.   ...
  3.   http:
  4.   - match: ...

Gateway with TLS passthrough

  1. apiVersion: networking.istio.io/v1beta1
  2. kind: Gateway
  3. metadata:
  4.   name: gateway
  5. spec:
  6.   selector:
  7.     istio: ingressgateway
  8.   servers:
  9.   - hosts:
  10.     - "*"
  11.     port:
  12.       name: https
  13.       number: 443
  14.       protocol: HTTPS
  15.     tls:
  16.       mode: PASSTHROUGH
  17. ---
  18. apiVersion: networking.istio.io/v1beta1
  19. kind: VirtualService
  20. metadata:
  21.   name: virtual-service
  22. spec:
  23.   gateways:
  24.   - gateway
  25.   hosts:
  26.   - httpbin.example.com
  27.   http:
  28.   - route:
  29.     - destination:
  30.         host: httpbin.org

In this configuration, the virtual service is attempting to match HTTP traffic against TLS traffic passed through the gateway. This will result in the virtual service configuration having no effect. You can observe that the HTTP route is not applied using the istioctl proxy-config listener and istioctl proxy-config route commands.

To fix this, you should switch the virtual service to configure tls routing:

  1. spec:
  2.   tls:
  3.   - match:
  4.     - sniHosts: ["httpbin.example.com"]
  5.     route:
  6.     - destination:
  7.         host: httpbin.org

Alternatively, you could terminate TLS, rather than passing it through, by switching the tls configuration in the gateway:

  1. spec:
  2.   ...
  3.     tls:
  4.       credentialName: sds-credential
  5.       mode: SIMPLE

Double TLS (TLS origination for a TLS request)

When configuring Istio to perform TLS origination, you need to make sure that the application sends plaintext requests to the sidecar, which will then originate the TLS.

The following DestinationRule originates TLS for requests to the httpbin.org service, but the corresponding ServiceEntry defines the protocol as HTTPS on port 443.

  1. apiVersion: networking.istio.io/v1beta1
  2. kind: ServiceEntry
  3. metadata:
  4.   name: httpbin
  5. spec:
  6.   hosts:
  7.   - httpbin.org
  8.   ports:
  9.   - number: 443
  10.     name: https
  11.     protocol: HTTPS
  12.   resolution: DNS
  13. ---
  14. apiVersion: networking.istio.io/v1beta1
  15. kind: DestinationRule
  16. metadata:
  17.   name: originate-tls
  18. spec:
  19.   host: httpbin.org
  20.   trafficPolicy:
  21.     tls:
  22.       mode: SIMPLE

With this configuration, the sidecar expects the application to send TLS traffic on port 443 (e.g., curl https://httpbin.org), but it will also perform TLS origination before forwarding requests. This will cause the requests to be double encrypted.

For example, sending a request like curl https://httpbin.org will result in an error: (35) error:1408F10B:SSL routines:ssl3_get_record:wrong version number.

You can fix this example by changing the port protocol in the ServiceEntry to HTTP:

  1. spec:
  2.   hosts:
  3.   - httpbin.org
  4.   ports:
  5.   - number: 443
  6.     name: http
  7.     protocol: HTTP

Note that with this configuration your application will need to send plaintext requests to port 443, like curl http://httpbin.org:443, because TLS origination does not change the port. However, starting in Istio 1.8, you can expose HTTP port 80 to the application (e.g., curl http://httpbin.org) and then redirect requests to targetPort 443 for the TLS origination:

  1. spec:
  2.   hosts:
  3.   - httpbin.org
  4.   ports:
  5.   - number: 80
  6.     name: http
  7.     protocol: HTTP
  8.     targetPort: 443

404 errors occur when multiple gateways configured with same TLS certificate

Configuring more than one gateway using the same TLS certificate will cause browsers that leverage HTTP/2 connection reuse (i.e., most browsers) to produce 404 errors when accessing a second host after a connection to another host has already been established.

For example, let’s say you have 2 hosts that share the same TLS certificate like this:

  • Wildcard certificate *.test.com installed in istio-ingressgateway
  • Gateway configuration gw1 with host service1.test.com, selector istio: ingressgateway, and TLS using gateway’s mounted (wildcard) certificate
  • Gateway configuration gw2 with host service2.test.com, selector istio: ingressgateway, and TLS using gateway’s mounted (wildcard) certificate
  • VirtualService configuration vs1 with host service1.test.com and gateway gw1
  • VirtualService configuration vs2 with host service2.test.com and gateway gw2

Since both gateways are served by the same workload (i.e., selector istio: ingressgateway) requests to both services (service1.test.com and service2.test.com) will resolve to the same IP. If service1.test.com is accessed first, it will return the wildcard certificate (*.test.com) indicating that connections to service2.test.com can use the same certificate. Browsers like Chrome and Firefox will consequently reuse the existing connection for requests to service2.test.com. Since the gateway (gw1) has no route for service2.test.com, it will then return a 404 (Not Found) response.

You can avoid this problem by configuring a single wildcard Gateway, instead of two (gw1 and gw2). Then, simply bind both VirtualServices to it like this:

  • Gateway configuration gw with host *.test.com, selector istio: ingressgateway, and TLS using gateway’s mounted (wildcard) certificate
  • VirtualService configuration vs1 with host service1.test.com and gateway gw
  • VirtualService configuration vs2 with host service2.test.com and gateway gw

Configuring SNI routing when not sending SNI

An HTTPS Gateway that specifies the hosts field will perform an SNI match on incoming requests. For example, the following configuration would only allow requests that match *.example.com in the SNI:

  1. servers:
  2. - port:
  3.     number: 443
  4.     name: https
  5.     protocol: HTTPS
  6.   hosts:
  7.   - "*.example.com"

This may cause certain requests to fail.

For example, if you do not have DNS set up and are instead directly setting the host header, such as curl 1.2.3.4 -H "Host: app.example.com", no SNI will be set, causing the request to fail. Instead, you can set up DNS or use the --resolve flag of curl. See the Secure Gateways task for more information.

Another common issue is load balancers in front of Istio. Most cloud load balancers will not forward the SNI, so if you are terminating TLS in your cloud load balancer you may need to do one of the following:

  • Configure the cloud load balancer to instead passthrough the TLS connection
  • Disable SNI matching in the Gateway by setting the hosts field to *

A common symptom of this is for the load balancer health checks to succeed while real traffic fails.

Unchanged Envoy filter configuration suddenly stops working

An EnvoyFilter configuration that specifies an insert position relative to another filter can be very fragile because, by default, the order of evaluation is based on the creation time of the filters. Consider a filter with the following specification:

  1. spec:
  2.   configPatches:
  3.   - applyTo: NETWORK_FILTER
  4.     match:
  5.       context: SIDECAR_OUTBOUND
  6.       listener:
  7.         portNumber: 443
  8.         filterChain:
  9.           filter:
  10.             name: istio.stats
  11.     patch:
  12.       operation: INSERT_BEFORE
  13.       value:
  14.         ...

To work properly, this filter configuration depends on the istio.stats filter having an older creation time than it. Otherwise, the INSERT_BEFORE operation will be silently ignored. There will be nothing in the error log to indicate that this filter has not been added to the chain.

This is particularly problematic when matching filters, like istio.stats, that are version specific (i.e., that include the proxyVersion field in their match criteria). Such filters may be removed or replaced by newer ones when upgrading Istio. As a result, an EnvoyFilter like the one above may initially be working perfectly but after upgrading Istio to a newer version it will no longer be included in the network filter chain of the sidecars.

To avoid this issue, you can either change the operation to one that does not depend on the presence of another filter (e.g., INSERT_FIRST), or set an explicit priority in the EnvoyFilter to override the default creation time-based ordering. For example, adding priority: 10 to the above filter will ensure that it is processed after the istio.stats filter which has a default priority of 0.

Virtual service with fault injection and retry/timeout policies not working as expected

Currently, Istio does not support configuring fault injections and retry or timeout policies on the same VirtualService. Consider the following configuration:

  1. apiVersion: networking.istio.io/v1alpha3
  2. kind: VirtualService
  3. metadata:
  4.   name: helloworld
  5. spec:
  6.   hosts:
  7.     - "*"
  8.   gateways:
  9.   - helloworld-gateway
  10.   http:
  11.   - match:
  12.     - uri:
  13.         exact: /hello
  14.     fault:
  15.       abort:
  16.         httpStatus: 500
  17.         percentage:
  18.           value: 50
  19.     retries:
  20.       attempts: 5
  21.       retryOn: 5xx
  22.     route:
  23.     - destination:
  24.         host: helloworld
  25.         port:
  26.           number: 5000

You would expect that given the configured five retry attempts, the user would almost never see any errors when calling the helloworld service. However since both fault and retries are configured on the same VirtualService, the retry configuration does not take effect, resulting in a 50% failure rate. To work around this issue, you may remove the fault config from your VirtualService and inject the fault to the upstream Envoy proxy using EnvoyFilter instead:

  1. apiVersion: networking.istio.io/v1alpha3
  2. kind: EnvoyFilter
  3. metadata:
  4.   name: hello-world-filter
  5. spec:
  6.   workloadSelector:
  7.     labels:
  8.       app: helloworld
  9.   configPatches:
  10.   - applyTo: HTTP_FILTER
  11.     match:
  12.       context: SIDECAR_INBOUND # will match outbound listeners in all sidecars
  13.       listener:
  14.         filterChain:
  15.           filter:
  16.             name: "envoy.filters.network.http_connection_manager"
  17.     patch:
  18.       operation: INSERT_BEFORE
  19.       value:
  20.         name: envoy.fault
  21.         typed_config:
  22.           "@type": "type.googleapis.com/envoy.extensions.filters.http.fault.v3.HTTPFault"
  23.           abort:
  24.             http_status: 500
  25.             percentage:
  26.               numerator: 50
  27.               denominator: HUNDRED

This works because this way the retry policy is configured for the client proxy while the fault injection is configured for the upstream proxy.