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。

Health Checks and Circuit Breakers Reference

You can make an API proxied by Kong use a ring-balancer, configured by adding an upstream entity that contains one or more target entities, each target pointing to a different IP address (or hostname) and port. The ring-balancer will balance load among the various targets, and based on the upstream configuration, will perform health checks on the targets, marking them as healthy or unhealthy based on whether they are responsive or not. The ring-balancer will then only route traffic to healthy targets.

Kong supports two kinds of health checks, which can be used separately or in conjunction:

active checks, where a specific HTTP or HTTPS endpoint in the target is periodically requested and the health of the target is determined based on its response;
passive checks (also known as circuit breakers), where Kong analyzes the ongoing traffic being proxied and determines the health of targets based on their behavior responding to requests.

Note: This feature is not supported in hybrid mode.

Defining healthy and unhealthy

Targets

The objective of the health checks functionality is to dynamically mark targets as healthy or unhealthy, for a given Kong node. There is no cluster-wide synchronization of health information: each Kong node determines the health of its targets separately. This is desirable since at a given point one Kong node may be able to connect to a target successfully while another node is failing to reach it: the first node will consider it healthy, while the second will mark it as unhealthy and start routing traffic to other targets of the upstream.

Either an active probe (on active health checks) or a proxied request (on passive health checks) produces data which is used to determine whether a target is healthy or unhealthy. A request may produce a TCP error, timeout, or produce an HTTP status code. Based on this information, the health checker updates a series of internal counters:

If the returned status code is one configured as “healthy”, it will increment the “Successes” counter for the target and clear all its other counters;
If it fails to connect, it will increment the “TCP failures” counter for the target and clear the “Successes” counter;
If it times out, it will increment the “timeouts” counter for the target and clear the “Successes” counter;
If the returned status code is one configured as “unhealthy”, it will increment the “HTTP failures” counter for the target and clear the “Successes” counter.

If any of the “TCP failures”, “HTTP failures” or “timeouts” counters reaches their configured threshold, the target will be marked as unhealthy.

If the “Successes” counter reaches its configured threshold, the target will be marked as healthy.

The list of which HTTP status codes are “healthy” or “unhealthy”, and the individual thresholds for each of these counters are configurable on a per-upstream basis. Below, we have an example of a configuration for an Upstream entity, showcasing the default values of the various fields available for configuring health checks. A description of each field is included in the Admin API reference documentation.

{
    "name": "service.v1.xyz",
    "healthchecks": {
        "active": {
            "concurrency": 10,
            "healthy": {
                "http_statuses": [ 200, 302 ],
                "interval": 0,
                "successes": 0
            },
            "http_path": "/",
            "timeout": 1,
            "unhealthy": {
                "http_failures": 0,
                "http_statuses": [ 429, 404, 500, 501,
                                   502, 503, 504, 505 ],
                "interval": 0,
                "tcp_failures": 0,
                "timeouts": 0
            }
        },
        "passive": {
            "healthy": {
                "http_statuses": [ 200, 201, 202, 203,
                                   204, 205, 206, 207,
                                   208, 226, 300, 301,
                                   302, 303, 304, 305,
                                   306, 307, 308 ],
                "successes": 0
            },
            "unhealthy": {
                "http_failures": 0,
                "http_statuses": [ 429, 500, 503 ],
                "tcp_failures": 0,
                "timeouts": 0
            }
        },
        "threshold": 0
    },
    "slots": 10
}

If an upstream is unhealthy (the available capacity % is less than the configured threshold), Kong will respond to requests to the upstream with 503 Service Unavailable.

Note:

Health checks operate only on active targets and do not modify the active status of a target in the Kong database.
Unhealthy targets will not be removed from the load balancer, and hence will not have any impact on the balancer layout when using the hashing algorithm (they will just be skipped).
The DNS caveats and balancer caveats also apply to health checks. If using hostnames for the targets, then make sure the DNS server always returns the full set of IP addresses for a name, and does not limit the response. Failing to do so might lead to health checks not being executed.

Upstreams

Along with health check functionality on individual targets, Upstreams also have a notion of health. The health of an Upstream is determined based on the status of its Targets.

Configuration of the Upstream’s health is done though the property healthchecks.threshold. This is a percentage of minimum available target “weight” (capacity) for the Upstream to be considered healthy.

Here is a simple example:

Assume an Upstream configured with healthchecks.threshold=55.
It has 5 targets, each with weight=100, so the total weight in the ring-balancer is 500.

When failures start to occur, the circuit-breaker for the first target trips. It is now considered unhealthy. This means that in the ring-balancer, 20% of the capacity is now unhealthy (100 weight out of 500). This is still above the threshold of 55, so the remaining targets will serve the traffic of the failed one.

When a second failure occurs, another target fails, and another 100 weight is lost as unhealthy. Now the ring-balancer operates at 60% of its capacity, but still within the configured threshold.

If we assume that the 2 failures occurred due to a system overload, we can now assume that the remaining 60% will also not be able to cope with the full load and soon a third node will fail, reducing healthy capacity to 40%. At this point, the Upstream health will be less than its threshold, and it will be marked as unhealthy itself.

Once it enters an unhealthy state, the Upstream will only return errors. This lets the targets/services recover from the cascading failure they were experiencing.

Once the Targets start recovering and the Upstream’s available capacity passes the threshold again, the health status of the ring-balancer will automatically be updated.

Types of health checks

Active health checks

Active health checks, as the name implies, actively probe targets for their health. When active health checks are enabled in an upstream entity, Kong will periodically issue HTTP or HTTPS requests to a configured path at each target of the upstream. This allows Kong to automatically enable and disable targets in the balancer based on the probe results.

The periodicity of active health checks can be configured separately for when a target is healthy or unhealthy. If the interval value for either is set to zero, the checking is disabled at the corresponding scenario. When both are zero, active health checks are disabled altogether.

Note: Active health checks only support HTTP/HTTPS targets. They do not apply to upstreams assigned to services with the protocol attribute set to tcp or TLS.

Passive health checks (circuit breakers)

Note: This feature is not supported in hybrid mode.

Passive health checks, also known as circuit breakers, are checks performed based on the requests being proxied by Kong (HTTP/HTTPS/TCP), with no additional traffic being generated. When a target becomes unresponsive, the passive health checker will detect that and mark the target as unhealthy. The ring-balancer will start skipping this target, so no more traffic will be routed to it.

Once the problem with a target is solved and it is ready to receive traffic again, the Kong administrator can manually inform the health checker that the target should be enabled again, via an Admin API endpoint:

curl -i -X PUT http://localhost:8001/upstreams/my_upstream/targets/10.1.2.3:1234/healthy

Response:

HTTP/1.1 204 No Content

This command will broadcast a cluster-wide message so that the “healthy” status is propagated to the whole Kong cluster. This will cause Kong nodes to reset the health counters of the health checkers running in all workers of the Kong node, allowing the ring-balancer to route traffic to the target again.

Passive health checks have the advantage of not producing extra traffic, but they are unable to automatically mark a target as healthy again: the “circuit is broken”, and the target needs to be re-enabled again by the system administrator.

Summary of pros and cons

Active health checks can automatically re-enable a target in the ring balancer as soon as it is healthy again. Passive health checks cannot.
Passive health checks do not produce additional traffic to the target. Active health checks do.
An active health checker demands a known URL with a reliable status response in the target to be configured as a probe endpoint (which may be as simple as "/"). Passive health checks do not demand such configuration.
By providing a custom probe endpoint for an active health checker, an application may determine its own health metrics and produce a status code to be consumed by Kong. Even though a target continues to serve traffic which looks healthy to the passive health checker, it would be able to respond to the active probe with a failure status, essentially requesting to be relieved from taking new traffic.

It is possible to combine the two modes. For example, one can enable passive health checks to monitor the target health based solely on its traffic, and only use active health checks while the target is unhealthy, in order to re-enable it automatically.

Enabling and disabling health checks

Enabling active health checks

To enable active health checks, you need to specify the configuration items under healthchecks.active in the Upstream object configuration. You need to specify the necessary information so that Kong can perform periodic probing on the target, and how to interpret the resulting information.

You can use the healthchecks.active.type field to specify whether to perform HTTP or HTTPS probes (setting it to "http" or "https"), or by simply testing if the connection to a given host and port is successful (setting it to "tcp").

For configuring the probe, you need to specify:

healthchecks.active.http_path - The path that should be used when issuing the HTTP GET request to the target. The default value is "/".
healthchecks.active.timeout - The connection timeout limit for the HTTP GET request of the probe. The default value is 1 second.
healthchecks.active.concurrency - Number of targets to check concurrently in active health checks.

You also need to specify positive values for intervals, for running probes:

healthchecks.active.healthy.interval - Interval between active health checks for healthy targets (in seconds). A value of zero indicates that active probes for healthy targets should not be performed.
healthchecks.active.unhealthy.interval - Interval between active health checks for unhealthy targets (in seconds). A value of zero indicates that active probes for unhealthy targets should not be performed.

This allows you to tune the behavior of the active health checks, whether you want probes for healthy and unhealthy targets to run at the same interval, or one to be more frequent than the other.

If you are using HTTPS healthchecks, you can also specify the following fields:

healthchecks.active.https_verify_certificate - Whether to check the validity of the SSL certificate of the remote host when performing active health checks using HTTPS.
healthchecks.active.https_sni - The hostname to use as an SNI (Server Name Identification) when performing active health checks using HTTPS. This is particularly useful when Targets are configured using IPs, so that the target host’s certificate can be verified with the proper SNI.

Note that failed TLS verifications will increment the “TCP failures” counter; the “HTTP failures” refer only to HTTP status codes, whether probes are done through HTTP or HTTPS.

Finally, you need to configure how Kong should interpret the probe, by setting the various thresholds on the health counters, which, once reached will trigger a status change. The counter threshold fields are:

healthchecks.active.healthy.successes - Number of successes in active probes (as defined by healthchecks.active.healthy.http_statuses) to consider a target healthy.
healthchecks.active.unhealthy.tcp_failures - Number of TCP failures or TLS verification failures in active probes to consider a target unhealthy.
healthchecks.active.unhealthy.timeouts - Number of timeouts in active probes to consider a target unhealthy.
healthchecks.active.unhealthy.http_failures - Number of HTTP failures in active probes (as defined by healthchecks.active.unhealthy.http_statuses) to consider a target unhealthy.

Enabling passive health checks

Note: This feature is not supported in hybrid mode.

Passive health checks do not feature a probe, as they work by interpreting the ongoing traffic that flows from a target. This means that to enable passive checks you only need to configure its counter thresholds:

healthchecks.passive.healthy.successes - Number of successes in proxied traffic (as defined by healthchecks.passive.healthy.http_statuses) to consider a target healthy, as observed by passive health checks. This needs to be positive when passive checks are enabled so that healthy traffic resets the unhealthy counters.
healthchecks.passive.unhealthy.tcp_failures - Number of TCP failures in proxied traffic to consider a target unhealthy, as observed by passive health checks.
healthchecks.passive.unhealthy.timeouts - Number of timeouts in proxied traffic to consider a target unhealthy, as observed by passive health checks.
healthchecks.passive.unhealthy.http_failures - Number of HTTP failures in proxied traffic (as defined by healthchecks.passive.unhealthy.http_statuses) to consider a target unhealthy, as observed by passive health checks.

Disabling health checks

In all counter thresholds and intervals specified in the healthchecks configuration, setting a value to zero means that the functionality the field represents is disabled. Setting a probe interval to zero disables a probe. Likewise, you can disable certain types of checks by setting their counter thresholds to zero. For example, to not consider timeouts when performing healthchecks, you can set both timeouts fields (for active and passive checks) to zero. This gives you a fine-grained control of the behavior of the health checker.

In summary, to completely disable active health checks for an upstream, you need to set both healthchecks.active.healthy.interval and healthchecks.active.unhealthy.interval to 0.

To completely disable passive health checks, you need to set all counter thresholds under healthchecks.passive for its various counters to zero.

All counter thresholds and intervals in healthchecks are zero by default, meaning that health checks are completely disabled by default in newly created upstreams.