コンテンツにスキップ
|
Kong Gateway
2.8.x LTS
report-issue問題を報告する

このページは、まだ日本語ではご利用いただけません。翻訳中です。

旧バージョンのドキュメントを参照しています。 最新のドキュメントはこちらをご参照ください。

Proxy Reference

In this document we will cover Kong’s proxying capabilities by explaining its routing capabilities and internal workings in details.

Kong exposes several interfaces which can be tweaked by the following configuration properties:

  • proxy_listen, which defines a list of addresses/ports on which Kong will accept public HTTP (gRPC, WebSocket, etc) traffic from clients and proxy it to your upstream services (8000 by default).

  • admin_listen, which also defines a list of addresses and ports, but those should be restricted to only be accessed by administrators, as they expose Kong’s configuration capabilities: the Admin API (8001 by default).

    Important: If you need to expose the admin_listen port to the internet in a production environment, secure it with authentication.

  • stream_listen, which is similar to proxy_listen but for Layer 4 (TCP, TLS) generic proxy. This is turned off by default.

Terminology

  • client: Refers to the downstream client making requests to Kong’s proxy port.
  • upstream service: Refers to your own API/service sitting behind Kong, to which client requests/connections are forwarded.
  • Service: Service entities, as the name implies, are abstractions of each of your own upstream services. Examples of Services would be a data transformation microservice, a billing API, etc.
  • Route: This refers to the Kong Routes entity. Routes are entrypoints into Kong, and defining rules for a request to be matched, and routed to a given Service.
  • Plugin: This refers to Kong “plugins”, which are pieces of business logic that run in the proxying lifecycle. Plugins can be configured through the Admin API - either globally (all incoming traffic) or on specific Routes and Services.

Overview

From a high-level perspective, Kong listens for HTTP traffic on its configured proxy port(s) (8000 and 8443 by default) and L4 traffic on explicitly configured stream_listen ports. Kong will evaluate any incoming HTTP request or L4 connection against the Routes you have configured and try to find a matching one. If a given request matches the rules of a specific Route, Kong will process proxying the request.

Because each Route may be linked to a Service, Kong will run the plugins you have configured on your Route and its associated Service, and then proxy the request upstream. You can manage Routes via Kong’s Admin API. Routes have special attributes that are used for routing - matching incoming HTTP requests. Routing attributes differ by subsystem (HTTP/HTTPS, gRPC/gRPCS, and TCP/TLS).

Subsystems and routing attributes:

  • http: methods, hosts, headers, paths (and snis, if https)
  • tcp: sources, destinations (and snis, if tls)
  • grpc: hosts, headers, paths (and snis, if grpcs)

If one attempts to configure a Route with a routing attribute it doesn’t support (e.g., an http route with sources or destinations fields), an error message will be reported:

HTTP/1.1 400 Bad Request
Content-Type: application/json
Server: kong/<x.x.x>

{
    "code": 2,
    "fields": {
        "sources": "cannot set 'sources' when 'protocols' is 'http' or 'https'"
    },
    "message": "schema violation (sources: cannot set 'sources' when 'protocols' is 'http' or 'https')",
    "name": "schema violation"
}

If Kong receives a request that it cannot match against any of the configured Routes (or if no Routes are configured), it will respond with:

HTTP/1.1 404 Not Found
Content-Type: application/json
Server: kong/<x.x.x>

{
    "message": "no route and no Service found with those values"
}

How to configure a Service

The Configuring a Service quickstart guide explains how Kong is configured via the Admin API.

Adding a Service to Kong is done by sending an HTTP request to the Admin API:

$ curl -i -X POST http://localhost:8001/services/ \
    -d 'name=foo-service' \
    -d 'url=http://foo-service.com'
HTTP/1.1 201 Created
...

{
    "connect_timeout": 60000,
    "created_at": 1515537771,
    "host": "foo-service.com",
    "id": "d54da06c-d69f-4910-8896-915c63c270cd",
    "name": "foo-service",
    "path": "/",
    "port": 80,
    "protocol": "http",
    "read_timeout": 60000,
    "retries": 5,
    "updated_at": 1515537771,
    "write_timeout": 60000
}

This request instructs Kong to register a Service named “foo-service”, which points to http://foo-service.com (your upstream).

Note: The url argument is a shorthand argument to populate the protocol, host, port, and path attributes at once.

Now, in order to send traffic to this Service through Kong, we need to specify a Route, which acts as an entry point to Kong:

$ curl -i -X POST http://localhost:8001/routes/ \
    -d 'hosts[]=example.com' \
    -d 'paths[]=/foo' \
    -d 'service.id=d54da06c-d69f-4910-8896-915c63c270cd'
HTTP/1.1 201 Created
...

{
    "created_at": 1515539858,
    "hosts": [
        "example.com"
    ],
    "id": "ee794195-6783-4056-a5cc-a7e0fde88c81",
    "methods": null,
    "paths": [
        "/foo"
    ],
    "preserve_host": false,
    "priority": 0,
    "protocols": [
        "http",
        "https"
    ],
    "service": {
        "id": "d54da06c-d69f-4910-8896-915c63c270cd"
    },
    "strip_path": true,
    "updated_at": 1515539858
}

We have now configured a Route to match incoming requests matching the given hosts and paths, and forward them to the foo-service we configured, thus proxying this traffic to http://foo-service.com.

Kong is a transparent proxy, and it will by default forward the request to your upstream service untouched, with the exception of various headers such as Connection, Date, and others as required by the HTTP specifications.

Routes and matching capabilities

Let’s now discuss how Kong matches a request against the configured routing attributes.

Kong supports native proxying of HTTP/HTTPS, TCP/TLS, and GRPC/GRPCS protocols; as mentioned earlier, each of these protocols accept a different set of routing attributes:

  • http: methods, hosts, headers, paths (and snis, if https)
  • tcp: sources, destinations (and snis, if tls)
  • grpc: hosts, headers, paths (and snis, if grpcs)

Note that all of these fields are optional, but at least one of them must be specified.

For a request to match a Route:

  • The request must include all of the configured fields
  • The values of the fields in the request must match at least one of the configured values (While the field configurations accepts one or more values, a request needs only one of the values to be considered a match)

Let’s go through a few examples. Consider a Route configured like this:

{
    "hosts": ["example.com", "foo-service.com"],
    "paths": ["/foo", "/bar"],
    "methods": ["GET"]
}

Some of the possible requests matching this Route would look like:

GET /foo HTTP/1.1
Host: example.com

GET /bar HTTP/1.1
Host: foo-service.com

GET /foo/hello/world HTTP/1.1
Host: example.com

All three of these requests satisfy all the conditions set in the Route definition.

However, the following requests would not match the configured conditions:

GET / HTTP/1.1
Host: example.com

POST /foo HTTP/1.1
Host: example.com

GET /foo HTTP/1.1
Host: foo.com

All three of these requests satisfy only two of configured conditions. The first request’s path is not a match for any of the configured paths, same for the second request’s HTTP method, and the third request’s Host header.

Now that we understand how the routing properties work together, let’s explore each property individually.

Request header

Kong supports routing by arbitrary HTTP headers. A special case of this feature is routing by the Host header.

Routing a request based on its Host header is the most straightforward way to proxy traffic through Kong, especially since this is the intended usage of the HTTP Host header. Kong makes it easy to do via the hosts field of the Route entity.

hosts accepts multiple values, which must be comma-separated when specifying them via the Admin API, and is represented in a JSON payload:

$ curl -i -X POST http://localhost:8001/routes/ \
    -H 'Content-Type: application/json' \
    -d '{"hosts":["example.com", "foo-service.com"]}'
HTTP/1.1 201 Created
...

But since the Admin API also supports form-urlencoded content types, you can specify an array via the [] notation:

$ curl -i -X POST http://localhost:8001/routes/ \
    -d 'hosts[]=example.com' \
    -d 'hosts[]=foo-service.com'
HTTP/1.1 201 Created
...

To satisfy the hosts condition of this Route, any incoming request from a client must now have its Host header set to one of:

Host: example.com

or:

Host: foo-service.com

Similarly, any other header can be used for routing:

$ curl -i -X POST http://localhost:8001/routes/ \
    -d 'headers.region=north'
HTTP/1.1 201 Created

Incoming requests containing a Region header set to North will be routed to said Route.

Region: North

Using wildcard hostnames

To provide flexibility, Kong allows you to specify hostnames with wildcards in the hosts field. Wildcard hostnames allow any matching Host header to satisfy the condition, and thus match a given Route.

Wildcard hostnames must contain only one asterisk at the leftmost or rightmost label of the domain. Examples:

  • *.example.com would allow Host values such as a.example.com and x.y.example.com to match.
  • example.* would allow Host values such as example.com and example.org to match.

A complete example would look like this:

{
    "hosts": ["*.example.com", "service.com"]
}

Which would allow the following requests to match this Route:

GET / HTTP/1.1
Host: an.example.com

GET / HTTP/1.1
Host: service.com

The preserve_host property

When proxying, Kong’s default behavior is to set the upstream request’s Host header to the hostname specified in the Service’s host. The preserve_host field accepts a boolean flag instructing Kong not to do so.

For example, when the preserve_host property is not changed and a Route is configured like so:

{
    "hosts": ["service.com"],
    "service": {
        "id": "..."
    }
}

A possible request from a client to Kong could be:

GET / HTTP/1.1
Host: service.com

Kong would extract the Host header value from the Service’s host property, , and would send the following upstream request:

GET / HTTP/1.1
Host: <my-service-host.com>

However, by explicitly configuring a Route with preserve_host=true:

{
    "hosts": ["service.com"],
    "preserve_host": true,
    "service": {
        "id": "..."
    }
}

And assuming the same request from the client:

GET / HTTP/1.1
Host: service.com

Kong would preserve the Host on the client request and would send the following upstream request instead:

GET / HTTP/1.1
Host: service.com

Additional request headers

It’s possible to route requests by other headers besides Host.

To do this, use the headers property in your Route:

{
    "headers": { "version": ["v1", "v2"] },
    "service": {
        "id": "..."
    }
}

Given a request with a header such as:

GET / HTTP/1.1
version: v1

This request will be routed through to the Service. The same will happen with this one:

GET / HTTP/1.1
version: v2

But this request will not be routed to the Service:

GET / HTTP/1.1
version: v3

Note: The headers keys are a logical AND and their values a logical OR.

Request path

Another way for a Route to be matched is via request paths. To satisfy this routing condition, a client request’s normalized path must be prefixed with one of the values of the paths attribute.

For example, with a Route configured like so:

{
    "paths": ["/service", "/hello/world"]
}

The following requests would be matched:

GET /service HTTP/1.1
Host: example.com

GET /service/resource?param=value HTTP/1.1
Host: example.com

GET /hello/world/resource HTTP/1.1
Host: anything.com

For each of these requests, Kong detects that their normalized URL path is prefixed with one of the Routes’s paths values. By default, Kong would then proxy the request upstream without changing the URL path.

When proxying with path prefixes, the longest paths get evaluated first. This allow you to define two Routes with two paths: /service and /service/resource, and ensure that the former does not “shadow” the latter.

Using Regex in paths

Kong supports regular expression pattern matching for an Route’s paths field via PCRE (Perl Compatible Regular Expression). You can assign paths as both prefixes and Regex to a Route at the same time.

For a path to be considered as Regex, it must fall outside of the following regex:

^[a-zA-Z0-9\.\-_~/%]*$

In other words, if a path contains any character that is not alphanumerical, dot (.), dash (-), underscore (_), tilde (~), forward-slash (/), or percent (%), then it will be considered a Regex path. This determination is done on a per-path basis and it is allowed to mix plain text and regex paths inside the same paths array of the same Route object.

For example, if we consider the following Route:

{
    "paths": ["/users/\d+/profile", "/following"]
}

The following requests would be matched by this Route:

GET /following HTTP/1.1
Host: ...

GET /users/123/profile HTTP/1.1
Host: ...

The provided Regex are evaluated with the a PCRE flag (PCRE_ANCHORED), meaning that they will be constrained to match at the first matching point in the path (the root / character).

Note: Regex matching is in general very fast, but it is possible to construct expressions that take a very long time to determine that they don’t actually match. This happens when the expression results in a “backtrace exponential explosion”, which means that there’s an astronomical number of tests to be performed before it’s conclusive that they’re all negative. If undetected, this could take hours to finish with a single regular expression. The PCRE engine automatically detects most types of this and replaces with more direct approaches, but there are more complex expressions that could result in this scenario.

To limit the worst-case scenario, Kong applies the OpenResty lua_regex_match_limit to ensure that any regex operation terminates in around two seconds in the worst case. Apart from that, a smaller limit is applied to some “critical” regex operations, like those for path selection, in order to terminate them within two milliseconds, at most.

Evaluation order

As previously mentioned, Kong evaluates prefix paths by length, the longest prefix paths are evaluated first. However, Kong will evaluate Regex paths based on the regex_priority attribute of Routes from highest priority to lowest. Regex paths are furthermore evaluated before prefix paths.

Consider the following Routes:

[
    {
        "paths": ["/status/\d+"],
        "regex_priority": 0
    },
    {
        "paths": ["/version/\d+/status/\d+"],
        "regex_priority": 6
    },
    {
        "paths": ["/version"],
    },
    {
        "paths": ["/version/any/"],
    }
]

In this scenario, Kong will evaluate incoming requests against the following defined URIs, in this order:

  1. /version/\d+/status/\d+
  2. /status/\d+
  3. /version/any/
  4. /version

Take care to avoid writing Regex rules that are overly broad and may consume traffic intended for a prefix rule. Adding a rule with the path /version/.* to the ruleset above would likely consume some traffic intended for the /version prefix path. If you see unexpected behavior, sending Kong-Debug: 1 in your request headers will indicate the matched Route ID in the response headers for troubleshooting purposes.

As usual, a request must still match a Route’s hosts and methods properties as well, and Kong will traverse your Routes until it finds one that matches the most rules.

Capturing groups

Capturing groups are also supported, and the matched group will be extracted from the path and available for plugins consumption. If we consider the following regex:

/version/(?<version>\d+)/users/(?<user>\S+)

And the following request path:

/version/1/users/john

Kong will consider the request path a match, and if the overall Route is matched (considering other routing attributes), the extracted capturing groups will be available from the plugins in the ngx.ctx variable:

local router_matches = ngx.ctx.router_matches

-- router_matches.uri_captures is:
-- { "1", "john", version = "1", user = "john" }
Escaping special characters

Next, it is worth noting that characters found in Regex are often reserved characters according to RFC 3986 and as such, should be percent-encoded. When configuring Routes with regex paths via the Admin API, be sure to URL encode your payload if necessary. For example, with curl and using an application/x-www-form-urlencoded MIME type:

$ curl -i -X POST http://localhost:8001/routes \
    --data-urlencode 'uris[]=/status/\d+'
HTTP/1.1 201 Created
...

Note that curl does not automatically URL encode your payload, and note the usage of --data-urlencode, which prevents the + character to be URL decoded and interpreted as a space ` ` by Kong’s Admin API.

The strip_path property

It may be desirable to specify a path prefix to match a Route, but not include it in the upstream request. To do so, use the strip_path boolean property by configuring a Route like so:

{
    "paths": ["/service"],
    "strip_path": true,
    "service": {
        "id": "..."
    }
}

Enabling this flag instructs Kong that when matching this Route, and proceeding with the proxying to a Service, it should not include the matched part of the URL path in the upstream request’s URL. For example, the following client’s request to the above Route:

GET /service/path/to/resource HTTP/1.1
Host: ...

Will cause Kong to send the following upstream request:

GET /path/to/resource HTTP/1.1
Host: ...

The same way, if a Regex path is defined on a Route that has strip_path enabled, the entirety of the request URL matching sequence will be stripped. Example:

{
    "paths": ["/version/\d+/service"],
    "strip_path": true,
    "service": {
        "id": "..."
    }
}

The following HTTP request matching the provided Regex path:

GET /version/1/service/path/to/resource HTTP/1.1
Host: ...

Will be proxied upstream by Kong as:

GET /path/to/resource HTTP/1.1
Host: ...

Normalization behavior

To prevent trivial Route match bypass, the incoming request URI from client is always normalized according to RFC 3986 before router matching occurs. Specifically, the following normalization techniques are used for incoming request URIs, which are selected because they generally do not change semantics of the request URI:

  1. Percent-encoded triplets are converted to uppercase. For example: /foo%3a becomes /foo%3A.
  2. Percent-encoded triplets of unreserved characters are decoded. For example: /fo%6F becomes /foo.
  3. Dot segments are removed as necessary. For example: /foo/./bar/../baz becomes /foo/baz.
  4. Duplicate slashes are merged. For example: /foo//bar becomes /foo/bar.

The paths attribute of the Route object are also normalized. It is achieved by first determining if the path is a plain text or regex path. Based on the result, different normalization techniques are used.

For plain text Route path:

Same normalization technique as above is used, that is, methods 1 through 4.

For regex Route path:

Only methods 1 and 2 are used. In addition, if the decoded character becomes a regex meta character, it will be escaped with backslash.

Kong normalizes any incoming request URI before performing router matches. As a result, any request URI sent over to the upstream services will also be in normalized form that preserves the original URI semantics.

Request HTTP method

The methods field allows matching the requests depending on their HTTP method. It accepts multiple values. Its default value is empty (the HTTP method is not used for routing).

The following Route allows routing via GET and HEAD:

{
    "methods": ["GET", "HEAD"],
    "service": {
        "id": "..."
    }
}

Such a Route would be matched with the following requests:

GET / HTTP/1.1
Host: ...

HEAD /resource HTTP/1.1
Host: ...

But it would not match a POST or DELETE request. This allows for much more granularity when configuring plugins on Routes. For example, one could imagine two Routes pointing to the same service: one with unlimited unauthenticated GET requests, and a second one allowing only authenticated and rate-limited POST requests (by applying the authentication and rate limiting plugins to such requests).

Request source

Note: This section only applies to TCP and TLS routes.

The sources routing attribute allows matching a route by a list of incoming connection IP and/or port sources.

The following Route allows routing via a list of source IP/ports:

{
    "protocols": ["tcp", "tls"],
    "sources": [{"ip":"10.1.0.0/16", "port":1234}, {"ip":"10.2.2.2"}, {"port":9123}],
    "id": "...",
}

TCP or TLS connections originating from IPs in CIDR range “10.1.0.0/16” or IP address “10.2.2.2” or Port “9123” would match such Route.

Request destination

Note: This section only applies to TCP and TLS routes.

The destinations attribute, similarly to sources, allows matching a route by a list of incoming connection IP and/or port, but uses the destination of the TCP/TLS connection as routing attribute.

Request SNI

When using secure protocols (https, grpcs, or tls), a Server Name Indication can be used as a routing attribute. The following Route allows routing via SNIs:

{
  "snis": ["foo.test", "example.com"],
  "id": "..."
}

Incoming requests with a matching hostname set in the TLS connection’s SNI extension would be routed to this Route. As mentioned, SNI routing applies not only to TLS, but also to other protocols carried over TLS - such as HTTPS and If multiple SNIs are specified in the Route, any of them can match with the incoming request’s SNI. with the incoming request (OR relationship between the names).

The SNI is indicated at TLS handshake time and cannot be modified after the TLS connection has been established. This means, for example, that multiple requests reusing the same keepalive connection will have the same SNI hostname while performing router match, regardless of the Host header. has been established. This means keepalive connections that send multiple requests will have the same SNI hostnames while performing router match (regardless of the Host header).

Please note that creating a route with mismatched SNI and Host header matcher is possible, but generally discouraged.

Matching priorities

A Route may define matching rules based on its headers, hosts, paths, and methods (plus snis for secure routes - "https", "grpcs", "tls") fields. For Kong to match an incoming request to a Route, all existing fields must be satisfied. However, Kong allows for quite some flexibility by allowing two or more Routes to be configured with fields containing the same values - when this occurs, Kong applies a priority rule.

The rule is: when evaluating a request, Kong will first try to match the Routes with the most rules.

For example, if two Routes are configured like so:

{
    "hosts": ["example.com"],
    "service": {
        "id": "..."
    }
},
{
    "hosts": ["example.com"],
    "methods": ["POST"],
    "service": {
        "id": "..."
    }
}

The second Route has a hosts field and a methods field, so it will be evaluated first by Kong. By doing so, we avoid the first Route “shadowing” calls intended for the second one.

Thus, this request will match the first Route

GET / HTTP/1.1
Host: example.com

And this request will match the second one:

POST / HTTP/1.1
Host: example.com

Following this logic, if a third Route was to be configured with a hosts field, a methods field, and a uris field, it would be evaluated first by Kong.

If the rule count for the given request is the same in two Routes A and B, then the following tiebreaker rules will be applied in the order they are listed. Route A will be selected over B if:

  • A has only “plain” Host headers and B has one or more “wildcard” host headers
  • A has more non-Host headers than B.
  • A has at least one “regex” paths and B has only “plain” paths.
  • A’s longest path is longer than B’s longest path.
  • A.created_at < B.created_at

Proxying behavior

The proxying rules above detail how Kong forwards incoming requests to your upstream services. Below, we detail what happens internally between the time Kong matches an HTTP request with a registered Route, and the actual forwarding of the request.

1. Load balancing

Kong implements load balancing capabilities to distribute proxied requests across a pool of instances of an upstream service.

You can find more information about configuring load balancing by consulting the Load Balancing Reference.

2. Plugins execution

Kong is extensible via “plugins” that hook themselves in the request/response lifecycle of the proxied requests. Plugins can perform a variety of operations in your environment and/or transformations on the proxied request.

Plugins can be configured to run globally (for all proxied traffic) or on specific Routes and Services. In both cases, you must create a plugin configuration via the Admin API.

Once a Route has been matched (and its associated Service entity), Kong will run plugins associated to either of those entities. Plugins configured on a Route run before plugins configured on a Service, but otherwise, the usual rules of plugins association apply.

These configured plugins will run their access phase, which you can find more information about in the Plugin development guide.

3. Proxying and upstream timeouts

Once Kong has executed all the necessary logic (including plugins), it is ready to forward the request to your upstream service. This is done via Nginx’s ngx_http_proxy_module. You can configure the desired timeouts for the connection between Kong and a given upstream, via the following properties of a Service:

  • connect_timeout: defines in milliseconds the timeout for establishing a connection to your upstream service. Defaults to 60000.
  • write_timeout: defines in milliseconds a timeout between two successive write operations for transmitting a request to your upstream service. Defaults to 60000.
  • read_timeout: defines in milliseconds a timeout between two successive read operations for receiving a request from your upstream service. Defaults to 60000.

Kong will send the request over HTTP/1.1, and set the following headers:

  • Host: <your_upstream_host>, as previously described in this document.
  • Connection: keep-alive, to allow for reusing the upstream connections.
  • X-Real-IP: <remote_addr>, where $remote_addr is the variable bearing the same name provided by ngx_http_core_module. Please note that the $remote_addr is likely overridden by ngx_http_realip_module.
  • X-Forwarded-For: <address>, where <address> is the content of $realip_remote_addr provided by ngx_http_realip_module appended to the request header with the same name.
  • X-Forwarded-Proto: <protocol>, where <protocol> is the protocol used by the client. In the case where $realip_remote_addr is one of the trusted addresses, the request header with the same name gets forwarded if provided. Otherwise, the value of the $scheme variable provided by ngx_http_core_module will be used.
  • X-Forwarded-Host: <host>, where <host> is the host name sent by the client. In the case where $realip_remote_addr is one of the trusted addresses, the request header with the same name gets forwarded if provided. Otherwise, the value of the $host variable provided by ngx_http_core_module will be used.
  • X-Forwarded-Port: <port>, where <port> is the port of the server which accepted a request. In the case where $realip_remote_addr is one of the trusted addresses, the request header with the same name gets forwarded if provided. Otherwise, the value of the $server_port variable provided by ngx_http_core_module will be used.

  • X-Forwarded-Prefix: <path>, where <path> is the path of the request which was accepted by Kong. In the case where $realip_remote_addr is one of the trusted addresses, the request header with the same name gets forwarded if provided. Otherwise, the value of the $request_uri variable (with the query string stripped) provided by ngx_http_core_module will be used.

    Note: Kong returns "/" for an empty path, but it doesn’t do any other normalization on the request path.

All the other request headers are forwarded as-is by Kong.

One exception to this is made when using the WebSocket protocol. If so, Kong will set the following headers to allow for upgrading the protocol between the client and your upstream services:

  • Connection: Upgrade
  • Upgrade: websocket

More information on this topic is covered in the Proxy WebSocket traffic section.

4. Errors and retries

Whenever an error occurs during proxying, Kong will use the underlying Nginx retries mechanism to pass the request on to the next upstream.

There are two configurable elements here:

  1. The number of retries: this can be configured per Service using the retries property. See the Admin API for more details on this.

  2. What exactly constitutes an error: here Kong uses the Nginx defaults, which means an error or timeout occurring while establishing a connection with the server, passing a request to it, or reading the response headers.

The second option is based on Nginx’s [proxy_next_upstream][proxy_next_upstream] directive. This option is not directly configurable through Kong, but can be added using a custom Nginx configuration. See the configuration reference for more details.

5. Response

Kong receives the response from the upstream service and sends it back to the downstream client in a streaming fashion. At this point Kong will execute subsequent plugins added to the Route and/or Service that implement a hook in the header_filter phase.

Once the header_filter phase of all registered plugins has been executed, the following headers will be added by Kong and the full set of headers be sent to the client:

  • Via: kong/x.x.x, where x.x.x is the Kong version in use
  • X-Kong-Proxy-Latency: <latency>, where latency is the time in milliseconds between Kong receiving the request from the client and sending the request to your upstream service.
  • X-Kong-Upstream-Latency: <latency>, where latency is the time in milliseconds that Kong was waiting for the first byte of the upstream service response.

Once the headers are sent to the client, Kong will start executing registered plugins for the Route and/or Service that implement the body_filter hook. This hook may be called multiple times, due to the streaming nature of Nginx. Each chunk of the upstream response that is successfully processed by such body_filter hooks is sent back to the client. You can find more information about the body_filter hook in the Plugin development guide.

Configuring a fallback Route

As a practical use-case and example of the flexibility offered by Kong’s proxying capabilities, let’s try to implement a “fallback Route”, so that in order to avoid Kong responding with an HTTP 404, “no route found”, we can catch such requests and proxy them to a special upstream service, or apply a plugin to it (such a plugin could, for example, terminate the request with a different status code or response without proxying the request).

Here is an example of such a fallback Route:

{
    "paths": ["/"],
    "service": {
        "id": "..."
    }
}

As you can guess, any HTTP request made to Kong would actually match this Route, since all URIs are prefixed by the root character /. As we know from the Request path section, the longest URL paths are evaluated first by Kong, so the / path will eventually be evaluated last by Kong, and effectively provide a “fallback” Route, only matched as a last resort. You can then send traffic to a special Service or apply any plugin you wish on this Route.

Configuring TLS for a Route

Kong provides a way to dynamically serve TLS certificates on a per-connection basis. TLS certificates are directly handled by the core, and configurable via the Admin API. Clients connecting to Kong over TLS must support the Server Name Indication extension to make use of this feature.

TLS certificates are handled by two resources in the Kong Admin API:

  • /certificates, which stores your keys and certificates.
  • /snis, which associates a registered certificate with a Server Name Indication.

You can find the documentation for those two resources in the Admin API Reference.

Here is how to configure a TLS certificate on a given Route: first, upload your TLS certificate and key via the Admin API:

$ curl -i -X POST http://localhost:8001/certificates \
    -F "cert=@/path/to/cert.pem" \
    -F "key=@/path/to/cert.key" \
    -F "snis=*.tls-example.com,other-tls-example.com"
HTTP/1.1 201 Created
...

The snis form parameter is a convenience parameter. Under the hood, it directly inserts an SNI and associates the uploaded certificate to it.

Note that one of the SNI names defined in snis above contains a wildcard (*.tls-example.com). An SNI may contain a single wildcard in the leftmost (prefix) or rightmost (suffix) position. This can be useful when maintaining multiple subdomains. A single sni configured with a wildcard name can be used to match multiple subdomains, instead of creating an SNI for each.

Valid wildcard positions are mydomain.*, *.mydomain.com, and *.www.mydomain.com.

A default certificate can be added using the following parameters in Kong configuration:

  1. ssl_cert
  2. ssl_cert_key

Or, by dynamically configuring the default certificate with an SNI of *:

$ curl -i -X POST http://localhost:8001/certificates \
    -F "cert=@/path/to/default-cert.pem" \
    -F "key=@/path/to/default-cert.key" \
    -F "snis=*"
HTTP/1.1 201 Created
...

Matching of snis respects the following priority:

  1. Exact SNI matching certificate
  2. Search for a certificate by prefix wildcard
  3. Search for a certificate by suffix wildcard
  4. Search for a certificate associated with the SNI *
  5. The default certificate on the file system

You must now register the following Route within Kong. We will match requests to this Route using only the Host header for convenience:

$ curl -i -X POST http://localhost:8001/routes \
    -d 'hosts=prefix.tls-example.com,other-tls-example.com' \
    -d 'service.id=d54da06c-d69f-4910-8896-915c63c270cd'
HTTP/1.1 201 Created
...

You can now expect the Route to be served over HTTPS by Kong:

$ curl -i https://localhost:8443/ \
  -H "Host: prefix.tls-example.com"
HTTP/1.1 200 OK
...

When establishing the connection and negotiating the TLS handshake, if your client sends prefix.tls-example.com as part of the SNI extension, Kong will serve the cert.pem certificate previously configured. This is the same for both HTTPS and TLS connections.

Restricting the client protocol

Routes have a protocols property to restrict the client protocol they should listen for. This attribute accepts a set of values, which can be "http", "https", "grpc", "grpcs", "tcp", or "tls".

A Route with http and https will accept traffic in both protocols.

{
    "hosts": ["..."],
    "paths": ["..."],
    "methods": ["..."],
    "protocols": ["http", "https"],
    "service": {
        "id": "..."
    }
}

Not specifying any protocol has the same effect, since routes default to ["http", "https"].

However, a Route with only https would only accept traffic over HTTPS. It would also accept unencrypted traffic if TLS termination previously occurred from a trusted IP. TLS termination is considered valid when the request comes from one of the configured IPs in trusted_ips and if the X-Forwarded-Proto: https header is set:

{
    "hosts": ["..."],
    "paths": ["..."],
    "methods": ["..."],
    "protocols": ["https"],
    "service": {
        "id": "..."
    }
}

If a Route such as the above matches a request, but that request is in plain-text without valid prior TLS termination, Kong responds with:

HTTP/1.1 426 Upgrade Required
Content-Type: application/json; charset=utf-8
Transfer-Encoding: chunked
Connection: Upgrade
Upgrade: TLS/1.2, HTTP/1.1
Server: kong/x.y.z

{"message":"Please use HTTPS protocol"}

It’s possible to create routes for raw TCP (not necessarily HTTP) connections by using "tcp" in the protocols attribute:

{
    "hosts": ["..."],
    "paths": ["..."],
    "methods": ["..."],
    "protocols": ["tcp"],
    "service": {
        "id": "..."
    }
}

Similarly, we can create routes which accept raw TLS traffic (not necessarily HTTPS) with the "tls" value:

{
    "hosts": ["..."],
    "paths": ["..."],
    "methods": ["..."],
    "protocols": ["tls"],
    "service": {
        "id": "..."
    }
}

A Route with only TLS would only accept traffic over TLS.

It is also possible to accept both TCP and TLS simultaneously:

{
    "hosts": ["..."],
    "paths": ["..."],
    "methods": ["..."],
    "protocols": ["tcp", "tls"],
    "service": {
        "id": "..."
    }
}

For L4 TLS proxy to work, it is necessary to create route that accepts the tls protocol, as well as having the appropriate TLS certificate uploaded and their sni attribute properly set to match incoming connection’s SNI. Please refer to the Configuring TLS for a Route section above for instructions on setting this up.

Proxy WebSocket traffic

Kong supports WebSocket traffic thanks to the underlying Nginx implementation. When you wish to establish a WebSocket connection between a client and your upstream services through Kong, you must establish a WebSocket handshake. This is done via the HTTP Upgrade mechanism. This is what your client request made to Kong would look like:

GET / HTTP/1.1
Connection: Upgrade
Host: my-websocket-api.com
Upgrade: WebSocket

This will make Kong forward the Connection and Upgrade headers to your upstream service, instead of dismissing them due to the hop-by-hop nature of a standard HTTP proxy.

WebSocket and TLS

Kong will accept ws and wss connections on its respective http and https ports. To enforce TLS connections from clients, set the protocols property of the Route to https only.

When setting up the Service to point to your upstream WebSocket service, you should carefully pick the protocol you want to use between Kong and the upstream. If you want to use TLS (wss), then the upstream WebSocket service must be defined using the https protocol in the Service protocol property, and the proper port (usually 443). To connect without TLS (ws), then the http protocol and port (usually 80) should be used in protocol instead.

If you want Kong to terminate TLS, you can accept wss only from the client, but proxy to the upstream service over plain text, or ws.

Proxy gRPC traffic

gRPC proxying is natively supported in Kong. In order to manage gRPC services and proxy gRPC requests with Kong, create Services and Routes for your gRPC Services (check out the Configuring a gRPC Service guide).

Only observability and logging plugins are supported with gRPC - plugins known to be supported with gRPC have “grpc” and “grpcs” listed under the supported protocols field in their Kong Hub page - for example, check out the File Log plugin’s page.

Proxy TCP/TLS traffic

TCP and TLS proxying is natively supported in Kong.

In this mode, data of incoming connections reaching the stream_listen endpoints will be passed through to the upstream. It is possible to terminate TLS connections from clients using this mode as well.

To use this mode, aside from defining stream_listen, appropriate Route/Service object with protocol types of tcp or tls should be created.

If TLS termination by Kong is desired, the following conditions must be met:

  1. The Kong port where TLS connection connects to must have the ssl flag enabled
  2. A certificate/key that can be used for TLS termination must be present inside Kong, as shown in Configuring TLS for a Route

Kong will use the connecting client’s TLS SNI server name extension to find the appropriate TLS certificate to use.

On the Service side, depends on whether connection between Kong and the upstream service need to be encrypted, tcp or tls protocol types can be set accordingly. This means all of the below setup are supported in this mode:

  1. Client <- TLS -> Kong <- TLS -> Upstream
  2. Client <- TLS -> Kong <- Cleartext -> Upstream
  3. Client <- Cleartext -> Kong <- TLS -> Upstream

Note: In L4 proxy mode, only plugins that has tcp or tls in the supported protocol list are supported. This list can be found in their respective documentation on the Kong Plugin Hub.

Back to top

Proxy TLS passthrough traffic

Kong Gateway supports proxying a TLS request without terminating or known as SNI proxy.

Kong Gateway uses the connecting client’s TLS SNI extension to find the matching Route and Service and forward the complete TLS request upstream, without decrypting the incoming TLS traffic.

In this mode, you need to:

  • Create a Route object with the protocol tls_passthrough, and the snis field set to one or more SNIs.
  • Set the corresponding Service object’s protocol to tcp.
  • Send requests to a port that has the ssl flag in its stream_listen directive.

Separate SNI and Certificate entities aren’t required and won’t be used.

Routes are not allowed to match tls and tls_passthrough protocols at same time. However, the same SNI is allowed to match tls and tls_passthrough in different Routes.

It’s also possible to set Route to tls_passthrough and Service to tls. In this mode, the connection to the upstream will be TLS-encrypted twice.

Note: To run any plugins in this mode, the plugin’s protocols field must include tls_passthrough.

Back to top

Conclusion

Through this guide, we hope you gained knowledge of the underlying proxying mechanism of Kong, from how does a request match a Route to be routed to its associated Service, on to how to allow for using the WebSocket protocol or setup dynamic TLS certificates.

This website is open-source and can be found at github.com/Kong/docs.konghq.com. Feel free to provide feedback to this document there, or propose improvements!

If you haven’t already, we suggest that you also read the Load balancing Reference, as it closely relates to the topic we just covered.