...

Text file src/sigs.k8s.io/gateway-api/geps/gep-1742.md

Documentation: sigs.k8s.io/gateway-api/geps

     1# GEP-1742: HTTPRoute Timeouts
     2
     3* Issue: [#1742](https://github.com/kubernetes-sigs/gateway-api/issues/1742)
     4* Status: Experimental
     5
     6(See status definitions [here](overview.md#status).)
     7
     8## TLDR
     9
    10Create some sort of design so that Gateway API objects can be used to configure
    11timeouts for different types of connection.
    12
    13## Goals
    14
    15- Create some method to configure some timeouts.
    16- Timeout config must be applicable to most if not all Gateway API implementations.
    17
    18## Non-Goals
    19
    20- A standard API for every possible timeout that implementations may support.
    21
    22## Introduction
    23
    24In talking about Gateway API objects, particularly HTTPRoute, we've mentioned
    25timeout configuration many times in the past as "too hard" to find the common
    26ground necessary to make more generic configuration. This GEP intends firstly
    27to make this process less difficult, then to find common timeouts that we can
    28build into Gateway API.
    29
    30For this initial round, we'll focus on Layer 7 HTTP traffic, while acknowledging
    31that Layer 4 connections have their own interesting timeouts as well.
    32
    33The following sections will review all the implementations, then document what
    34timeouts are _available_ for the various data planes.
    35
    36### Background on implementations
    37
    38Most implementations that handle HTTPRoute objects use a proxy as the data plane
    39implementation, that actually forwards flows as directed by Gateway API configuration.
    40
    41The following table is a review of all the listed implementations of Gateway API
    42at the time of writing, with the data plane they use for Layer 7, based on what information
    43could be found online. If there are errors here, or if the implementation doesn't
    44support layer 7, please feel free to correct them.
    45
    46| Implementation | Data Plane       |
    47|----------------|------------|
    48| Acnodal EPIC   | Envoy      |
    49| Apache APISIX  | Nginx      |
    50| BIG-IP Kubernetes Gateway| F5 BIG-IP  |
    51| Cilium         | Envoy      |
    52| Contour        | Envoy      |
    53| Emissary Ingress| Envoy     |
    54| Envoy Gateway  | Envoy      |
    55| Flomesh Service Mesh | Pipy |
    56| Gloo Edge      | Envoy      |
    57| Google Kubernetes Engine (GKE) | Similar to Envoy Timeouts |
    58| HAProxy Ingress | HAProxy   |
    59| Hashicorp Consul | Envoy    |
    60| Istio          | Envoy      |
    61| Kong           | Nginx      |
    62| Kuma           | Envoy      |
    63| Litespeed      | Litespeed WebADC |
    64| NGINX Gateway Fabric | Nginx |
    65| Traefik        | Traefik    |
    66
    67
    68### Flow diagrams with available timeouts
    69
    70The following flow diagrams are based off the basic diagram below, with all the
    71timeouts I could find included.
    72
    73In general, timeouts are recorded with the setting name or similar that the data
    74plane uses for them, and are correct as far as I've parsed the documentation
    75correctly.
    76
    77Idle timeouts are marked as such.
    78
    79```mermaid
    80sequenceDiagram
    81    participant C as Client
    82    participant P as Proxy
    83    participant U as Upstream
    84    C->>P: Connection Started
    85    C->>P: Starts sending Request
    86    C->>P: Finishes Headers
    87    C->>P: Finishes request
    88    P->>U: Connection Started
    89    P->>U: Starts sending Request
    90    P->>U: Finishes request
    91    P->>U: Finishes Headers
    92    U->>P: Starts Response
    93    U->>P: Finishes Headers
    94    U->>P: Finishes Response
    95    P->>C: Starts Response
    96    P->>C: Finishes Headers
    97    P->>C: Finishes Response
    98    Note right of P: Repeat if connection sharing
    99    U->>C: Connection ended
   100```
   101
   102#### Envoy Timeouts
   103
   104For Envoy, some timeouts are configurable at either the HTTP Connection Manager
   105(very, very roughly equivalent to a Listener), the Route (equivalent to a HTTPRoute)
   106level, or the Cluster (usually close to the Service) or some combination. These
   107are noted in the below diagram with a `CM`, `R`, or `Cluster` prefix respectively.
   108
   109```mermaid
   110sequenceDiagram
   111    participant C as Client
   112    participant P as Envoy
   113    participant U as Upstream
   114    C->>P: Connection Started
   115    activate P
   116    Note left of P: transport_socket_connect_timeout for TLS
   117    deactivate P
   118    C->>P: Starts sending Request
   119    activate C
   120    activate P
   121    activate P
   122    C->>P: Finishes Headers
   123    note left of P: CM request_headers_timeout
   124		C->>P: Finishes request
   125    deactivate P
   126    activate U
   127    note left of U: Cluster connect_timeout
   128    deactivate U
   129    P->>U: Connection Started
   130		activate U
   131    note right of U: CM idle_timeout<br />CM max_connection_duration
   132    P->>U: Starts sending Request
   133    P->>U: Finishes Headers
   134    note left of P: CM request_timeout
   135		P->>U: Finishes request
   136    deactivate P
   137    activate U
   138    U->>P: Starts Response
   139    U->>P: Finishes Headers
   140		note right of U: R timeout<br/>R per_try_timeout<br/>R per_try_idle_timeout
   141    U->>P: Finishes Response
   142    deactivate U
   143    P->>C: Starts Response
   144    P->>C: Finishes Headers
   145    P->>C: Finishes Response
   146    Note left of C: CM stream_idle_timeout<br />R idle_timeout<br />CM,R max_stream_duration<br/>TCP proxy idle_timeout<br />TCP protocol idle_timeout
   147    deactivate C
   148    Note right of P: Repeat if connection sharing
   149    U->>C: Connection ended
   150    deactivate U
   151```
   152
   153#### Nginx timeouts
   154
   155Nginx allows setting of GRPC and general HTTP timeouts separately, although the
   156purposes seem to be roughly equivalent.
   157
   158```mermaid
   159sequenceDiagram
   160    participant C as Client
   161    participant P as Nginx
   162    participant U as Upstream
   163    C->>P: Connection Started
   164    activate P
   165    C->>P: Starts sending Request
   166    C->>P: Finishes Headers
   167    Note right of P: client_headers_timeout
   168    deactivate P
   169    activate P
   170    C->>P: Finishes request
   171    deactivate P
   172    Note right of P: client_body_timeout
   173    activate U
   174    note left of U: proxy_connect_timeout<br/>grpc_connect_timeout
   175    deactivate U
   176    P->>U: Connection Started
   177    Activate U
   178	  Activate U
   179    P->>U: Starts sending Request
   180    P->>U: Finishes Headers
   181		P->>U: Finishes request
   182    Note right of U: (between write operations)<br/>proxy_send_timeout<br/>grpc_send_timeout
   183    deactivate U
   184		activate U
   185    U->>P: Starts Response
   186    U->>P: Finishes Headers
   187        Note right of U: (between read operations)<br/>proxy_read_timeout<br/>grpc_read_timeout
   188    U->>P: Finishes Response
   189    deactivate U
   190    activate P
   191    P->>C: Starts Response
   192    P->>C: Finishes Headers
   193    P->>C: Finishes Response
   194    deactivate P
   195    Note left of P: send_timeout (only between two successive write operations)
   196    Note left of C: Repeat if connection is shared until server's keepalive_timeout is hit
   197    Note Right of U: upstream's keepalive_timeout (if keepalive enabled)
   198    U->>C: Connection ended
   199		deactivate U
   200```
   201
   202#### HAProxy timeouts
   203
   204```mermaid
   205sequenceDiagram
   206    participant C as Client
   207    participant P as Proxy
   208    participant U as Upstream
   209
   210    C->>P: Connection Started
   211    activate U
   212    activate C
   213    activate P
   214    note left of P: timeout client (idle)
   215    C->>P: Starts sending Request
   216    C->>P: Finishes Headers
   217    C->>P: Finishes request
   218    note left of C: timeout http-request
   219    deactivate C
   220			activate C
   221    note left of C: timeout client-fin
   222    deactivate C
   223		deactivate P
   224    activate U
   225    note left of U: timeout queue<br/>(wait for available server)
   226    deactivate U
   227
   228    P->>U: Connection Started
   229    activate U
   230    P->>U: Starts sending Request
   231    activate U
   232    P->>U: Finishes Headers
   233    P->>U: Finishes request
   234
   235    note right of U: timeout connect
   236    deactivate U
   237    note left of U: timeout server<br/>(idle timeout)
   238    deactivate U
   239    activate U
   240    note left of U: timeout server-fin
   241    deactivate U
   242    U->>P: Starts Response
   243    U->>P: Finishes Headers
   244    U->>P: Finishes Response
   245    P->>C: Starts Response
   246    P->>C: Finishes Headers
   247    P->>C: Finishes Response
   248    activate C
   249    note left of C: timeout http-keep-alive
   250    deactivate C
   251    Note right of P: Repeat if connection sharing
   252    Note right of U: timeout tunnel<br/>(for upgraded connections)
   253    deactivate U
   254    U->>C: Connection ended
   255
   256```
   257
   258#### Traefik timeouts
   259
   260```mermaid
   261sequenceDiagram
   262    participant C as Client
   263    participant P as Proxy
   264    participant U as Upstream
   265    C->>P: Connection Started
   266    activate U
   267    C->>P: Starts sending Request
   268    activate P
   269    C->>P: Finishes Headers
   270    Note right of P: respondingTimeouts<br/>readTimeout
   271    C->>P: Finishes request
   272    deactivate P
   273    P->>U: Connection Started
   274    activate U
   275    Note right of U: forwardingTimeouts<br/>dialTimeout
   276    deactivate U
   277    P->>U: Starts sending Request
   278    P->>U: Finishes request
   279    P->>U: Finishes Headers
   280    U->>P: Starts Response
   281    activate U
   282    note right of U: forwardingTimeouts<br/>responseHeaderTimeout
   283    U->>P: Finishes Headers
   284    deactivate U
   285    U->>P: Finishes Response
   286    P->>C: Starts Response
   287    activate P
   288    P->>C: Finishes Headers
   289    Note right of P: respondingTimeouts<br/>writeTimeout
   290    P->>C: Finishes Response
   291    deactivate P
   292    Note right of P: Repeat if connection sharing
   293    Note right of U: respondingTimeouts<br/>idleTimeout<br/>Keepalive connections only
   294    deactivate U
   295    U->>C: Connection ended
   296
   297```
   298#### F5 BIG-IP Timeouts
   299
   300Could not find any HTTP specific timeouts. PRs welcomed. 😊
   301
   302#### Pipy Timeouts
   303
   304Could not find any HTTP specific timeouts. PRs welcomed. 😊
   305
   306#### Litespeed WebADC Timeouts
   307
   308Could not find any HTTP specific timeouts. PRs welcomed. 😊
   309
   310## API
   311
   312The above diagrams show that there are many different kinds of configurable timeouts
   313supported by Gateway implementations: connect, idle, request, upstream, downstream.
   314Although there may be opportunity for the specification of a common API for more of
   315them in the future, this GEP will focus on the L7 timeouts in HTTPRoutes that are
   316most valuable to clients.
   317
   318From the above analysis, it appears that most implementations are capable of
   319supporting the configuration of simple client downstream request timeouts on HTTPRoute
   320rules. This is a relatively small addition that would benefit many users.
   321
   322Some implementations support configuring a timeout for individual backend requests,
   323separate from the overall client request timeout. This is particularly useful if a
   324client HTTP request to a gateway can result in more than one call from the gateway
   325to the destination backend service, for example, if automatic retries are supported.
   326Adding support for this would also benefit many users.
   327
   328### Timeout values
   329
   330There are 2 kinds of timeouts that can be configured in an `HTTPRouteRule`:
   331
   3321. `timeouts.request` is the timeout for the Gateway API implementation to send a
   333    response to a client HTTP request. Whether the gateway starts the timeout before
   334    or after the entire client request stream has been received, is implementation dependent.
   335    This field is optional `Extended` support.
   336
   3371. `timeouts.backendRequest` is a timeout for a single request from the gateway to a backend.
   338    This field is optional `Extended` support. Typically used in conjunction with retry configuration,
   339    if supported by an implementation.
   340    Note that retry configuration will be the subject of a separate GEP (GEP-1731).
   341
   342```mermaid
   343sequenceDiagram
   344    participant C as Client
   345    participant P as Proxy
   346    participant U as Upstream
   347    C->>P: Connection Started
   348    note left of P: timeouts.request start time (min)
   349    C->>P: Starts sending Request
   350    C->>P: Finishes Headers
   351    C->>P: Finishes request
   352    note left of P: timeouts.request start time (max)
   353    P->>U: Connection Started
   354    note right of P: timeouts.backendRequest start time
   355    P->>U: Starts sending Request
   356    P->>U: Finishes request
   357    P->>U: Finishes Headers
   358    U->>P: Starts Response
   359    U->>P: Finishes Headers
   360    note right of P: timeouts.backendRequest end time
   361    note left of P: timeouts.request end time
   362    U->>P: Finishes Response
   363    note right of P: Repeat if retry
   364    P->>C: Starts Response
   365    P->>C: Finishes Headers
   366    P->>C: Finishes Response
   367    Note right of P: Repeat if connection sharing
   368    U->>C: Connection ended
   369```
   370
   371Both timeout fields are [GEP-2257 Duration] values. A zero-valued timeout
   372("0s") MUST be interpreted as disabling the timeout; a non-zero-valued timeout
   373MUST be >= 1ms.
   374
   375[GEP-2257 Duration]:/geps/gep-2257/
   376
   377### GO
   378
   379```go
   380type HTTPRouteRule struct {
   381	// Timeouts defines the timeouts that can be configured for an HTTP request.
   382	//
   383	// Support: Extended
   384	//
   385	// +optional
   386	// <gateway:experimental>
   387	Timeouts *HTTPRouteTimeouts `json:"timeouts,omitempty"`
   388
   389	// ...
   390}
   391
   392// HTTPRouteTimeouts defines timeouts that can be configured for an HTTPRoute.
   393// Timeout values are represented with Gateway API Duration formatting.
   394// Specifying a zero value such as "0s" is interpreted as no timeout.
   395//
   396// +kubebuilder:validation:XValidation:message="backendRequest timeout cannot be longer than request timeout",rule="!(has(self.request) && has(self.backendRequest) && duration(self.request) != duration('0s') && duration(self.backendRequest) > duration(self.request))"
   397type HTTPRouteTimeouts struct {
   398	// Request specifies the maximum duration for a gateway to respond to an HTTP request.
   399	// If the gateway has not been able to respond before this deadline is met, the gateway
   400	// MUST return a timeout error.
   401	//
   402	// For example, setting the `rules.timeouts.request` field to the value `10s` in an
   403	// `HTTPRoute` will cause a timeout if a client request is taking longer than 10 seconds
   404	// to complete.
   405	//
   406	// This timeout is intended to cover as close to the whole request-response transaction
   407	// as possible although an implementation MAY choose to start the timeout after the entire
   408	// request stream has been received instead of immediately after the transaction is
   409	// initiated by the client.
   410	//
   411	// When this field is unspecified, request timeout behavior is implementation-specific.
   412	//
   413	// Support: Extended
   414	//
   415	// +optional
   416	Request *Duration `json:"request,omitempty"`
   417
   418	// BackendRequest specifies a timeout for an individual request from the gateway
   419	// to a backend. This covers the time from when the request first starts being
   420	// sent from the gateway to when the full response has been received from the backend.
   421	//
   422	// An entire client HTTP transaction with a gateway, covered by the Request timeout,
   423	// may result in more than one call from the gateway to the destination backend,
   424	// for example, if automatic retries are supported.
   425	//
   426	// Because the Request timeout encompasses the BackendRequest timeout, the value of
   427	// BackendRequest must be <= the value of Request timeout.
   428	//
   429	// Support: Extended
   430	//
   431	// +optional
   432	BackendRequest *Duration `json:"backendRequest,omitempty"`
   433}
   434
   435// Duration is a string value representing a duration in time. The foramat is as specified
   436// in GEP-2257, a strict subset of the syntax parsed by Golang time.ParseDuration.
   437//
   438// +kubebuilder:validation:Pattern=`^([0-9]{1,5}(h|m|s|ms)){1,4}$`
   439type Duration string
   440```
   441
   442### YAML
   443
   444```yaml
   445apiVersion: gateway.networking.k8s.io/v1beta1
   446kind: HTTPRoute
   447metadata:
   448  name: timeout-example
   449spec:
   450  ...
   451  rules:
   452  - backendRefs:
   453    - name: some-service
   454      port: 8080
   455    timeouts:
   456      request: 10s
   457      backendRequest: 2s
   458```
   459
   460### Conformance Details
   461
   462Gateway implementations can indicate support for the optional behavior in this GEP using
   463the following feature names:
   464
   465- `HTTPRouteRequestTimeout`: supports `rules.timeouts.request` in an `HTTPRoute`.
   466- `HTTPRouteBackendTimeout`: supports `rules.timeouts.backendRequest` in an `HTTPRoute`.
   467
   468## Alternatives
   469
   470Timeouts could be configured using policy attachments or in objects other than `HTTPRouteRule`.
   471
   472### Policy Attachment
   473
   474Instead of configuring timeouts directly on an API object, they could be configured using policy
   475attachments. The advantage to this approach would be that timeout policies can be not only
   476configured for an `HTTPRouteRule`, but can also be added/overriden at a more fine
   477(e.g., `HTTPBackendRef`) or course (e.g. `HTTPRoute`) level of granularity.
   478
   479The downside, however, is complexity introduced for the most common use case, adding a simple
   480timeout for an HTTP request. Setting a single field in the route rule, instead of needing to
   481create a policy resource, for this simple case seems much better.
   482
   483In the future, we could consider using policy attachments to configure less common kinds of
   484timeouts that may be needed, but it would probably be better to instead extend the proposed API
   485to support those timeouts as well.
   486
   487The default values of the proposed timeout fields could also be overridden
   488using policy attachments in the future. For example, a policy attachment could be used to set the
   489default value of `rules.timeouts.request` for all routes under an `HTTPRoute` or `Gateway`.
   490
   491### Other API Objects
   492
   493The new timeouts field could be added to a different API struct, instead of `HTTPRouteRule`.
   494
   495Putting it on an `HTTPBackendRef`, for example, would allow users to set different timeouts for different
   496backends. This is a feature that we believe has not been requested by existing proxy or service mesh
   497clients and is also not implementable using available timeouts of most proxies.
   498
   499Another alternative is to move the timeouts configuration up a level in the API to `HTTPRoute`. This
   500would be convenient when a user wants the same timeout on all rules, but would be overly restrictive.
   501Using policy attachments to override the default timeout value for all rules, as described in the
   502previous section, is likely a better way to handle timeout configuration above the route rule level.
   503
   504## References
   505
   506(Add any additional document links. Again, we should try to avoid
   507too much content not in version control to avoid broken links)

View as plain text