Load balancing across backend servers

This page applies to Apigee and Apigee hybrid.

View Apigee Edge documentation.

Apigee enhances the availability of your APIs by providing built-in support for load balancing and failover across multiple backend server instances.

Target servers decouple concrete endpoint URLs from target endpoint configurations. Instead of defining a concrete URL in the configuration, you can configure one or more named target servers. Then, each target server is referenced by name in a target endpoint HTTPConnection.

Videos

Watch the following videos to learn more about API routing and load balancing using target servers

About the target server configuration

A target server configuration consists of a name, a host, a protocol, and a port, with an additional element to indicate whether the target server is enabled or disabled. A target server can also have an optional <sSLInfo> object.

The following code provides an example of a target server configuration:

{
  "name": "target1",
  "host": "1.mybackendservice.com",
  "protocol": "http",
  "port": "80",
  "isEnabled": "true"
}

For more information on the TargetServers API, see organizations.environments.targetservers.

See the schema for TargetServer and other entities on GitHub.

Creating target servers

Create target servers using the Apigee UI or API as described in the following sections.

curl "https://apigee.googleapis.com/v1/organizations/$ORG/environments/$ENV/targetservers" \
  -X POST \
  -H "Content-Type:application/json" \
  -H "Authorization: Bearer $TOKEN" \
  -d '
  {
  "name": "target1",
  "host": "1.mybackendservice.com",
  "protocol": "http",
  "port": "80",
  "isEnabled": "true",
  }'
 

{
  "host" : "1.mybackendservice.com",
  "protocol": "http",
  "isEnabled" : true,
  "name" : "target1",
  "port" : 80
}

curl "https://apigee.googleapis.com/v1/organizations/$ORG/environments/$ENV/targetservers" \
  -X POST \
  -H "Content-Type:application/json" \
  -H "Authorization: Bearer $TOKEN" \
  -d '
  {
  "name": "target2",
  "host": "2.mybackendservice.com",
  "protocol": "http",
  "port": "80",
  "isEnabled": "true",
  }'

{
  "host" : "2.mybackendservice.com",
  "protocol": "http",
  "isEnabled" : true,
  "name" : "target2",
  "port" : 80
}

curl https://apigee.googleapis.com/v1/organizations/$ORG/environments/$ENV/targetservers \
  -H "Authorization: Bearer $TOKEN"

[ "target2", "target1" ]

Configuring a target endpoint to load balance across named target servers

Now that you have two target servers available, you can modify the target endpoint's' <HTTPTargetConnection> element to reference those two target servers by name:

<TargetEndpoint name="default">
  <HTTPTargetConnection>
    <LoadBalancer>
      <Server name="target1" />
      <Server name="target2" />
    </LoadBalancer>
    <Path>/test</Path>
  </HTTPTargetConnection>
</TargetEndpoint>

The configuration above is the most basic load balancing configuration. The load balancer supports three load balancing algorithms, Round Robin, Weighted, and Least Connection. Round Robin is the default algorithm. Since no algorithm is specified in the configuration above, outbound requests from the API proxy to the backend servers will alternate, one for one, between target1 and target2.

The <Path> element forms the basepath of the target endpoint URI for all target servers. It is only used when <LoadBalancer> is used. Otherwise, it is ignored. In the above example, a request reaching to target1 will be http://target1/test and so for other target servers.

Configuring load balancer options

You can tune availability by configuring the options for load balancing and failover at the load balancer and target server level as described in the following sections.

Algorithm

Sets the algorithm used by <LoadBalancer>. The available algorithms are RoundRobin, Weighted, and LeastConnections, each of which is documented below.

Round robin

The default algorithm, round robin, forwards a request to each target server in the order in which the servers are listed in the target endpoint HTTP connection. For example:

<TargetEndpoint name="default">
  <HTTPTargetConnection>
      <LoadBalancer>
        <Algorithm>RoundRobin</Algorithm>
        <Server name="target1" />
        <Server name="target2" />
      </LoadBalancer>
      <Path>/test</Path>
  </HTTPTargetConnection>
</TargetEndpoint>

Weighted

The weighted load balancing algorithm enables you to configure proportional traffic loads for your target servers. The weighted LoadBalancer distributes requests to your target servers in direct proportion to each target server's weight. Therefore, the weighted algorithm requires you to set a weight attribute for each target server. For example:

<TargetEndpoint name="default">
  <HTTPTargetConnection>
    <LoadBalancer>
      <Algorithm>Weighted</Algorithm>
      <Server name="target1">
        <Weight>1</Weight>
      </Server>
      <Server name="target2">
        <Weight>2</Weight>
      </Server>
    </LoadBalancer>
    <Path>/test</Path>
  </HTTPTargetConnection>
</TargetEndpoint>

In this example, two requests will be routed to target2 for every one request routed to target1.

Least Connection

LoadBalancers configured to use the least connection algorithm route outbound requests to the target server with fewest open HTTP connections. For example:

<TargetEndpoint name="default">
  <HTTPTargetConnection>
      <LoadBalancer>
        <Algorithm>LeastConnections</Algorithm>
        <Server name="target1" />
        <Server name="target2" />
      </LoadBalancer>
  </HTTPTargetConnection>
  <Path>/test</Path>
</TargetEndpoint>

Maximum failures

The maximum number of failed requests from the API proxy to the target server that results in the request being redirected to another target server.

A response failure means Apigee doesn't receive any response from a target server. When this happens, the failure counter increments by one.

However, when Apigee does receive a response from a target, even if the response is an HTTP error (such as 404 or 500), that counts as a response from the target server, and the failure counter is reset. To ensure that bad HTTP responses (such as 500) also increment the failure counter to take an unhealthy server out of load balancing rotation as soon as possible, you can add the <ServerUnhealthyResponse> element with <ResponseCode> child elements to your load balancer configuration. Apigee will also count responses with those codes as failures.

In the following example, target1 will be removed from rotation after five failed requests, including 404 and some 5XX responses from the target server.

<TargetEndpoint name="default">
  <HTTPTargetConnection>
      <LoadBalancer>
        <Algorithm>RoundRobin</Algorithm>
        <Server name="target1" />
        <Server name="target2" />
        <MaxFailures>5</MaxFailures>
        <ServerUnhealthyResponse>
            <ResponseCode>404</ResponseCode>
            <ResponseCode>500</ResponseCode>
            <ResponseCode>502</ResponseCode>
            <ResponseCode>503</ResponseCode>
        </ServerUnhealthyResponse>
      </LoadBalancer>
      <Path>/test</Path>
  </HTTPTargetConnection>
</TargetEndpoint>

The MaxFailures default is 0. This means that Apigee always tries to connect to the target for each request and never removes the target server from the rotation.

It is best to use MaxFailures > 0 with a health monitor. If you configure MaxFailures > 0, the target server is removed from rotation when the target fails the number of times you indicate. When a health monitor is in place, Apigee automatically puts the target server back in rotation after the target is up and running again, according to the configuration of that health monitor. See Health monitoring for more information.

Alternatively, if you configure MaxFailures > 0 and you do not configure a health monitor, Apigee will automatically take the target server out of rotation when the first failure is detected. Apigee will check the health of the target server every five minutes and return it to the rotation when it responds normally.

Retry

If retry is enabled, a request will be retried whenever a response failure (I/O error or HTTP timeout) occurs or the response received matches a value set by <ServerUnhealthyResponse>. See Maximum failures above for more on setting <ServerUnhealthyResponse>.

By default <RetryEnabled> is set to true. Set to false to disable retry. For example:

<RetryEnabled>false</RetryEnabled>

IsFallback

One (and only one) target server can be set as the fallback server. The fallback target server is not included in load balancing routines until all other target servers are identified as unavailable by the load balancer. When the load balancer determines that all non-fallback target servers are unavailable, all traffic is routed to the fallback server. For example:

<TargetEndpoint name="default">
  <HTTPTargetConnection>
      <LoadBalancer>
        <Algorithm>RoundRobin</Algorithm>
        <Server name="target1" />
        <Server name="target2" />
        <Server name="target3">
          <IsFallback>true</IsFallback>
        </Server>
      </LoadBalancer>
      <Path>/test</Path>
  </HTTPTargetConnection>
</TargetEndpoint>

The configuration above results in round robin load balancing between target1 and target2 until both target1 and target2 are unavailable. When target1 and target2 are unavailable, all traffic is routed to target3.

If the fallback server is unhealthy, Apigee will not route traffic to it. Instead, API calls will return a 503 "Service is temporarily unavailable" error.

Path

Path defines a URI fragment that will be appended to all requests issued by the target server to the backend server.

This element accepts a literal string path or a message template. A message template lets you perform variable string substitution at runtime. For example, in the following target endpoint definition, the value of {mypath} is used for the path:

<HTTPTargetConnection>
    <SSLInfo>
      <Enabled>true</Enabled>
    </SSLInfo>
    <LoadBalancer>
      <Server name="testserver"/>
    </LoadBalancer>
    <Path>{mypath}</Path>
</HTTPTargetConnection>

Configuring a target server for TLS/SSL

If you are using a target server to define the backend service, and the backend service requires the connection to use the HTTPS protocol, then you must enable TLS/SSL in the target server definition. This is necessary because the host tag does not let you specify the connection protocol.

Configure one-way TLS/SSL

Use the following configuration to define a target server with one-way TLS/SSL. In this example, Apigee makes HTTPS requests to the backend service:

{
  "name": "target1",
  "host": "mocktarget.apigee.net",
  "protocol": "http",
  "port": "443",
  "isEnabled": "true",
  "sSLInfo": {
    "enabled": "true"
  }
}

Configure strict SSL enforcement

To specify strict SSL enforcement in the target server definition, set enforce to true in the sSLInfo block, as shown in the following example:

  {
    "name": "target1",
    "host": "mocktarget.apigee.net",
    "protocol": "http",
    "port": "443",
    "isEnabled": "true",
    "sSLInfo": {
      "enabled": "true",
      "enforce": "true"
    }
  }
  

Configure two-way TLS/SSL

If the backend service requires two-way, or mutual, TLS/SSL, then you can configure the target server with the same TLS/SSL configuration settings as target endpoints:

{
  "name": "TargetServer 1",
  "host": "www.example.com",
  "protocol": "http",
  "port": "443",
  "isEnabled": "true",
  "sSLInfo": {
    "enabled": "true",
    "clientAuthEnabled": "true",
    "keyStore": "keystore-name",
    "keyAlias": "keystore-alias",
    "trustStore": "truststore-name",
    "ignoreValidationErrors": "false",
    "ciphers": []
  }
}

Configure strict SSL using the API

To create a target server with strict SSL enforcement using an API call:

curl "https://apigee.googleapis.com/v1/organizations/$ORG/environments/$ENV/targetservers" \
  -X POST \
  -H "Content-Type:application/json"  \
  -H "Authorization: Bearer $TOKEN" \
  -d '
  {
    "name": "TargetServer 1",
    "host": "www.example.com",
    "protocol": "http",
    "port": 443,
    "isEnabled": true,
    "sSLInfo":
    {
      "enabled":true,
      "enforce":true,
      "clientAuthEnabled":true,
      "keyStore":"keystore-name",
      "keyAlias":"keystore-alias",
      "ciphers":[],
      "protocols":[],
      "clientAuthEnabled":true,
      "ignoreValidationErrors":false,
      "trustStore":"truststore-name"
    }
  }'

Health monitoring

Health monitoring enables you to enhance load balancing configurations by actively polling the backend service URLs defined in the target server configurations. With health monitoring enabled, target servers that are unreachable or return an error response are marked unhealthy. A failed target server is automatically put back into rotation when the health monitor determines that the target server is active. No proxy re-deployments are required.

A health monitor acts as a simple client that invokes a backend service over TCP or HTTP:

• A TCP client simply ensures that a socket can be opened.
• You configure the HTTP client to submit a valid HTTP request to the backend service. You can define HTTP GET, PUT, POST, or DELETE operations. The response of the HTTP monitor call must match the configured settings in the <SuccessResponse> block.

Successes and failures

When you enable health monitoring, Apigee begins sending health checks to your target server. A health check is a request sent to the target server that determines whether or not the target server is healthy.

A health check can have one of two possible results:

• Success: The target server is considered healthy when a successful health check occurs. This is typically the result of one or more of the following:
  • The target server accepts a new connection to the specified port, responds to a request on that port, and then closes the port within the specified timeframe. The response from the target server contains Connection: close
  • The target server responds to a health check request with a 200 (OK) or other HTTP status code that you determine is acceptable.
  • The target server responds to a health check request with a message body that matches the expected message body.

  When Apigee determines that a server is healthy, Apigee continues or resumes sending requests to it.

• Failure: The target server can fail a health check in different ways, depending on the type of check. A failure can be logged when the target server:
  • Refuses a connection from Apigee to the health check port.
  • Does not respond to a health check request within a specified period of time.
  • Returns an unexpected HTTP status code.
  • Responds with a message body that does not match the expected message body.

  When a target server fails a health check, Apigee increments that server's failure count. If the number of failures for that server meets or exceeds a predefined threshold (<MaxFailures>), Apigee stops sending requests to that server.

Enabling a health monitor

To create a health monitor for an API proxy, add the <HealthMonitor> element to the target endpoint's <HTTPTargetConnection element configuration for the proxy.

Health monitors cannot be created in the UI. Instead, you create a proxy configuration and upload it as a ZIP file to Apigee. A proxy configuration is a structured description of all aspects of an API proxy. Proxy configurations consist of XML files in a pre-defined directory structure. For more information, see API Proxy Configuration Reference.

<HealthMonitor> configuration elements

The following table describes the <HealthMonitor> configuration elements:

In addition to these elements, enabling a health monitor requires that you set the <MaxFailures> element in the <HTTPTargetConnection> block of the <TargetEndpoint> element to a value greater than 0. <MaxFailures> is used to specify the maximum number of failed requests from the API proxy to the target server that can occur before the request is redirected to another target server.

By default <MaxFailures> is 0, which means Apigee performs no corrective action. When configuring a health monitor, ensure that you set <MaxFailures> to a non-zero value.

Health monitor using TCP monitor

The sample health monitor described in the following configuration uses a TCP monitor to poll each target server by opening a connection on port 80 every five seconds. (Port is optional. If not specified, the TCP monitor port is the target server port.)

In this sample health monitor configuration:

• If the connection fails or takes more than 10 seconds to connect, then the failure count increments by 1 for that target server.
• If the connection succeeds, then the failure count for the target server is reset to 0.

You can add a health monitor as a child element of the target endpoint's <HTTPTargetConnection> element, as shown below:

<TargetEndpoint name="default">
  <HTTPTargetConnection>
      <LoadBalancer>
        <Algorithm>RoundRobin</Algorithm>
        <Server name="target1" />
        <Server name="target2" />
        <MaxFailures>5</MaxFailures>
      </LoadBalancer>
      <Path>/test</Path>
      <HealthMonitor>
        <IsEnabled>true</IsEnabled>
        <IntervalInSec>5</IntervalInSec>
        <TCPMonitor>
            <ConnectTimeoutInSec>10</ConnectTimeoutInSec>
            <Port>80</Port>
        </TCPMonitor>
      </HealthMonitor>
  </HTTPTargetConnection>
</TargetEndpoint>

<TCPMonitor> configuration elements

The following table describes the <TCPMonitor> configuration elements:

Health monitor using HTTP monitor

The sample health monitor described in the following configuration uses an HTTP monitor that submits a GET request to the backend service once every five seconds and adds an HTTP Basic Authorization header to the request message.

In this sample health monitor configuration:

• The expected response from the backend service is an HTTP response code 200.
• The expected value for the custom HTTP Basic Authorization header ImOK is YourOK.
• The <UseTargetServerSSLInfo> element is set to true to use the SSL parameters from the target server's <SSLInfo> block.

With this configuration, the expected response codes and header values will be compared against actual responses from the backend service. If the response does not match, then the request will treated as a failure by the load balancer configuration.

By default, health monitors cannot access keystores, truststores, or other SSL parameters from their target servers. To enable access to these parameters for your health monitor in order to support mutual TLS, for example, add <UseTargetServerSSLInfo>true</UseTargetServerSSLInfo> in the <Request> block.

    <TargetEndpoint name="default">
      <HTTPTargetConnection>
        <LoadBalancer>
          <Algorithm>RoundRobin</Algorithm>
          <Server name="target1" />
          <Server name="target2" />
          <MaxFailures>5</MaxFailures>
        </LoadBalancer>
        <Path>/test</Path>
        <HealthMonitor>
          <IsEnabled>true</IsEnabled>
          <IntervalInSec>5</IntervalInSec>
          <HTTPMonitor>
            <Request>
              <UseTargetServerSSLInfo>true</UseTargetServerSSLInfo>
              <ConnectTimeoutInSec>10</ConnectTimeoutInSec>
              <SocketReadTimeoutInSec>30</SocketReadTimeoutInSec>
              <Port>80</Port>
              <Verb>GET</Verb>
              <Path>/healthcheck</Path>
              <Header name="Authorization">Basic 12e98yfw87etf</Header>
              <IncludeHealthCheckIdHeader>true</IncludeHealthCheckIdHeader>
            </Request>
            <SuccessResponse>
              <ResponseCode>200</ResponseCode>
              <Header name="ImOK">YourOK</Header>
            </SuccessResponse>
          </HTTPMonitor>
        </HealthMonitor>
      </HTTPTargetConnection>
    </TargetEndpoint>
  

<HTTPMonitor> configuration elements

The following table describes the top-level <HTTPMonitor> configuration elements:

<HTTPMonitor>/<Request> configuration elements

Configuration options for the outbound request message sent by the health monitor to the target servers in the rotation. Note that <Request> is a required element.

<HTTPMonitor>/<SuccessResponse> configuration elements

(Optional) Matching options for the inbound HTTP response message generated by the polled backend service. Responses that do not match increment the failure count by 1.