Active health checks and monitoring of Nginx upstreams

Disclaimer: this is not an Nginx module in the traditional sense! It compiles to a shared library that gets loaded in Nginx using directive haskell load from Nginx module nginx-haskell-module. Let's call this plugin. The plugin provides support for active health checks and monitoring of peers in normal per-worker and shared upstreams (those declared with directive zone, possibly with dynamic peers), it supports all kinds of balancing models present in Nginx (round-robin, least_conn, hash and so forth). Both health checks and monitoring are optional, meaning that they do not depend on each other, and one feature may be switched off while the other is on.

Table of contents

• What the active health checks here means
• What the monitoring here means
• Examples
  • Normal upstreams, health checks and monitoring
  • Normal upstreams, only health checks
  • Normal upstreams, only monitoring
  • Shared upstreams, health checks and monitoring
• Nginx-based monitoring of normal upstreams
  • Normal upstreams, health checks and monitoring, related changes
• Periodic checks of healthy peers
• Collecting Prometheus metrics
  • Normal upstreams, related changes
  • Shared upstreams, related changes
• Corner cases
• Building and installation

What the active health checks here means

Well, they are not completely active. They are active only in one direction. If all the peers in a checked upstream respond successfully then the active health check does nothing! Successfulness of a peer response is checked against the rules of directive proxy_next_upstream and settings in directive server. This is quite standard in Nginx: a peer gets marked as failed when its N responses do not pass the proxy_next_upstream rules during time period of T. Values of N and T are defined in arguments max_fails and fail_timeout in directive server respectively: by default they are equal to 1 and 10s. When a peer fails, it cannot be chosen for proxying requests during the next time period of T.

And here the active health checks come into play! They merely move the time of the last check of failed peers forward periodically, say, every 5 seconds, which makes Nginx unable to return them back to the valid peers. So, do they keep staying failed forever? No! The active health check tests them every those 5 seconds against its own rules and recover them when the check passes.

What the monitoring here means

There are traditional per-worker and shared upstreams. Monitoring of traditional (or normal) upstreams is built in a dedicated location that returns a JSON object which contains a list of failed peers nested in worker's PID / health check service / upstream hierarchy. Monitoring of shared upstreams is as well built in a dedicated location that returns a JSON object with a list of failed peers nested in health check service / upstream hierarchy. The meaning of the health check service will be explained later.

Examples

In all examples below, the shared library gets built from a simple Haskell code which only exports service handlers from module NgxExport.Healthcheck. The code is saved in file ngx_healthcheck.hs.

module NgxHealthcheck where

import NgxExport.Healthcheck ()

Normal upstreams, health checks and monitoring

user                    nginx;
worker_processes        4;

events {
    worker_connections  1024;
}

http {
    default_type        application/octet-stream;
    sendfile            on;

    upstream u_backend {
        server localhost:8020 fail_timeout=600s;
        server localhost:8030 fail_timeout=600s;
    }

    upstream u_backend0 {
        server localhost:8040 fail_timeout=600s;
        server localhost:8050 fail_timeout=600s backup;
    }

    upstream u_backend1 {
        server localhost:8060 fail_timeout=600s;
        server localhost:8050 fail_timeout=600s backup;
    }

    proxy_next_upstream error timeout http_502;

    haskell load /var/lib/nginx/ngx_healthcheck.so;

    haskell_run_service checkPeers $hs_service_healthcheck
        'hs_service_healthcheck
         Conf { upstreams     = ["u_backend"
                                ,"u_backend0"
                                ,"u_backend1"
                                ]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Just Endpoint { epUrl = "/healthcheck"
                                              , epProto = Http
                                              , epPassRule = DefaultPassRule
                                              }
              , sendStatsPort = Just 8100
              }
        ';

    haskell_service_update_hook updatePeers $hs_service_healthcheck;

    haskell_run_service checkPeers $hs_service_healthcheck0
        'hs_service_healthcheck0
         Conf { upstreams     = ["u_backend"]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Just Endpoint { epUrl = "/healthcheck"
                                              , epProto = Http
                                              , epPassRule =
                                                      PassRuleByHttpStatus
                                                      [200, 404]
                                              }
              , sendStatsPort = Just 8100
              }
        ';

    haskell_service_update_hook updatePeers $hs_service_healthcheck0;

    haskell_run_service statsServer $hs_service_stats
        'StatsServerConf { ssPort          = 8100
                         , ssPurgeInterval = Min 5
                         }
        ';

    haskell_service_var_in_shm stats 64k /tmp $hs_service_stats;

    server {
        listen          8010;
        server_name     main;

        location /pass {
            proxy_pass http://u_backend;
        }

        location /pass0 {
            proxy_pass http://u_backend0;
        }

        location /pass1 {
            proxy_pass http://u_backend1;
        }

        location /stat {
            allow 127.0.0.1;
            deny all;
            proxy_pass http://127.0.0.1:8100;
        }
    }
}

In this configuration two health check services are declared: they are bound to service variables $hs_service_healthcheck and $hs_service_healthcheck0. The services run two instances of an asynchronous Haskell handler checkPeers in every Nginx worker process when the workers start. The services have complementary service update hooks that run a Haskell handler updatePeers synchronously in the main Nginx thread when the service variable updates, i.e. every 5 (or up to 7) seconds as stated in values interval and peerTimeout in constructor Conf. Each service has an associated unique key specified before the Conf declaration: in this example keys are equal to the names of service variables. Service keys also define values of header Host (with port if given) and the host name for validation of server certificates used in health checks over https (without port). The host name gets known from a service key according to the following rule:

1. if the service key contains no slashes: the host name is equal to the service key,
2. otherwise, if the service key contains the only slash at its end: the host name is taken from the name of the server bound to the peer in the upstream configuration,
3. otherwise (if the service key contains slashes on the left of its end): the host name is equal to the part of the service key after the first slash from its beginning, for example, if a service key is equal to 1/healthcheck then the host name is equal to healthcheck.

Besides time intervals in constructor Conf, a list of upstreams to check, an optional endpoint and an optional monitoring port where statistics about failed peers should be sent to are defined. Endpoints contain an URL on which health checks will test failed peers, transport protocol (Http or Https), and passing rules to describe in which cases failed peers must be regarded as recovered. Currently, only two kinds of passing rules are supported: DefaultPassRule (when a peer responds with HTTP status 200) and PassRuleByHttpStatus. Internally, updatePeers calls a C function that reads and updates Nginx data related to peer statuses.

Do not hesitate to declare as many Haskell services as you want, because they are very cheap. When you have to check against a large number of upstreams, it is even recommended to split services because this may potentially shorten duration of the synchronous service hook.

Monitoring service statsServer collects statistics of failed peers. This is a shared service because the associated service variable are declared as shared in directive haskell_service_var_in_shm, and therefore it runs on a single arbitrary Nginx worker all the time. Internally, it is implemented using Snap framework. The server is bound to the local IP address 127.0.0.1 and only accessible from outside via a dedicated location /stat. It will listen to incoming requests on the port specified in option ssPort. Option ssPurgeInterval defines how often stats from died workers will be purged.

Below is a typical response from the monitoring service.

{
  "31493": [
    "2020-10-08T11:27:05.914756785Z",
    {
      "hs_service_healthcheck": {
        "u_backend": [
          "127.0.0.1:8020"
        ],
        "u_backend1": [
          "127.0.0.1:8060",
          "127.0.0.1:8050"
        ]
      },
      "hs_service_healthcheck0": {
        "u_backend": [
          "127.0.0.1:8020"
        ]
      }
    }
  ],
  "31494": [
    "2020-10-08T11:27:05.945503075Z",
    {
      "hs_service_healthcheck": {
        "u_backend": [
          "127.0.0.1:8020"
        ]
      },
      "hs_service_healthcheck0": {
        "u_backend": [
          "127.0.0.1:8020"
        ]
      }
    }
  ],
  "31496": [
    "2020-10-08T11:27:05.946501076Z",
    {}
  ],
  "31497": [
    "2020-10-08T11:27:05.932294049Z",
    {
      "hs_service_healthcheck": {
        "u_backend": [
          "127.0.0.1:8020"
        ]
      },
      "hs_service_healthcheck0": {
        "u_backend": [
          "127.0.0.1:8020"
        ]
      }
    }
  ]
}

When the monitoring service is accessed via URL /stat/merge, its response data gets merged from all worker processes.

{
  "hs_service_healthcheck": {
    "u_backend": [
      [
        "2020-10-08T11:27:05.945503075Z",
        "127.0.0.1:8020"
      ]
    ],
    "u_backend1": [
      [
        "2020-10-08T11:27:05.914756785Z",
        "127.0.0.1:8060"
      ],
      [
        "2020-10-08T11:27:05.914756785Z",
        "127.0.0.1:8050"
      ]
    ]
  },
  "hs_service_healthcheck0": {
    "u_backend": [
      [
        "2020-10-08T11:27:05.945503075Z",
        "127.0.0.1:8020"
      ]
    ]
  }
}

In this merged view, all faulty peers are tagged with times of their latest checks. Notice also that upstreams in the merged view never contain duplicate peers.

Normal upstreams, only health checks

Henceforth I am going to skip unrelated parts of the configuration for brevity.

    haskell_run_service checkPeers $hs_service_healthcheck
        'hs_service_healthcheck
         Conf { upstreams     = ["u_backend"
                                ,"u_backend0"
                                ,"u_backend1"
                                ]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Just Endpoint { epUrl = "/healthcheck"
                                              , epProto = Http
                                              , epPassRule = DefaultPassRule
                                              }
              , sendStatsPort = Nothing
              }
        ';

    haskell_service_update_hook updatePeers $hs_service_healthcheck;

    haskell_run_service checkPeers $hs_service_healthcheck0
        'hs_service_healthcheck0
         Conf { upstreams     = ["u_backend"]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Just Endpoint { epUrl = "/healthcheck"
                                              , epProto = Http
                                              , epPassRule =
                                                      PassRuleByHttpStatus
                                                      [200, 404]
                                              }
              , sendStatsPort = Nothing
              }
        ';

    server {
        listen          8010;
        server_name     main;

        location /pass {
            proxy_pass http://u_backend;
        }

        location /pass0 {
            proxy_pass http://u_backend0;
        }

        location /pass1 {
            proxy_pass http://u_backend1;
        }
    }
}

Values of sendStatsPort in both the health check services are not defined (i.e. they are equal to special value Nothing), service statsServer and location /stats are removed.

Normal upstreams, only monitoring

    haskell_run_service checkPeers $hs_service_healthcheck
        'hs_service_healthcheck
         Conf { upstreams     = ["u_backend"
                                ,"u_backend0"
                                ,"u_backend1"
                                ]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Nothing
              , sendStatsPort = Just 8100
              }
        ';

    haskell_service_update_hook updatePeers $hs_service_healthcheck;

    haskell_run_service checkPeers $hs_service_healthcheck0
        'hs_service_healthcheck0
         Conf { upstreams     = ["u_backend"]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Nothing
              , sendStatsPort = Just 8100
              }
        ';

    haskell_run_service statsServer $hs_service_stats
        'StatsServerConf { ssPort          = 8100
                         , ssPurgeInterval = Min 5
                         }
        ';

    haskell_service_var_in_shm stats 64k /tmp $hs_service_stats;

    server {
        listen          8010;
        server_name     main;

        location /pass {
            proxy_pass http://u_backend;
        }

        location /pass0 {
            proxy_pass http://u_backend0;
        }

        location /pass1 {
            proxy_pass http://u_backend1;
        }

        location /stat {
            allow 127.0.0.1;
            deny all;
            proxy_pass http://127.0.0.1:8100;
        }
    }
}

Endpoints are not defined, the monitoring service and location /stat are intact.

Shared upstreams, health checks and monitoring

    upstream u_backend {
        zone u_backend 64k;
        server localhost:8020 fail_timeout=600s;
        server localhost:8030 fail_timeout=600s;
    }

    upstream u_backend0 {
        zone u_backend 64k;
        server localhost:8040 fail_timeout=600s;
        server localhost:8050 fail_timeout=600s backup;
    }

    upstream u_backend1 {
        zone u_backend 64k;
        server localhost:8060 fail_timeout=600s;
        server localhost:8050 fail_timeout=600s backup;
    }

    haskell_run_service checkPeers $hs_service_healthcheck
        'hs_service_healthcheck
         Conf { upstreams     = ["u_backend"
                                ,"u_backend0"
                                ,"u_backend1"
                                ]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Just Endpoint { epUrl = "/healthcheck"
                                              , epProto = Http
                                              , epPassRule = DefaultPassRule
                                              }
              , sendStatsPort = Nothing
              }
        ';

    haskell_service_update_hook updatePeers $hs_service_healthcheck;

    haskell_run_service checkPeers $hs_service_healthcheck0
        'hs_service_healthcheck0
         Conf { upstreams     = ["u_backend"]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Just Endpoint { epUrl = "/healthcheck"
                                              , epProto = Http
                                              , epPassRule =
                                                      PassRuleByHttpStatus
                                                      [200, 404]
                                              }
              , sendStatsPort = Nothing
              }
        ';

    haskell_service_update_hook updatePeers $hs_service_healthcheck0;

    haskell_service_var_in_shm upstreams 64k /tmp
            $hs_service_healthcheck $hs_service_healthcheck0;

    server {
        listen          8010;
        server_name     main;

        location /pass {
            proxy_pass http://u_backend;
        }

        location /pass0 {
            proxy_pass http://u_backend0;
        }

        location /pass1 {
            proxy_pass http://u_backend1;
        }

        location /stat {
            allow 127.0.0.1;
            deny all;
            haskell_async_content reportPeers;
        }
    }

The upstreams are now associated with a shared memory zone. The health check services become shared, the monitoring service gets replaced with a simple content handler reportPeers because peers are now shared between all Nginx worker processes.

Combinations of only health checks / only monitoring are built trivially like in the case of normal per-worker upstreams.

Below is a sample stats output for shared upstreams.

{
  "hs_service_healthcheck": {
    "u_backend": [
      "127.0.0.1:8020"
    ],
    "u_backend1": [
      "127.0.0.1:8060",
      "127.0.0.1:8050"
    ]
  },
  "hs_service_healthcheck0": {
    "u_backend": [
      "127.0.0.1:8020"
    ]
  }
}

Nginx-based monitoring of normal upstreams

Service statsServer is implemented using Snap framework. Basically, a native Nginx implementation is not easy because the service must listen on a single (not duplicated) file descriptor which is not the case when Nginx spawns more than one worker processes. Running statsServer as a shared service is an elegant solution as shared services guarantee that they occupy only one worker at a time. However, nginx-haskell-module provides directive single_listener which can be used to apply the required restriction in a custom Nginx virtual server. This directive requires that the virtual server listens with option reuseport and is only available on Linux with socket option SO_ATTACH_REUSEPORT_CBPF.

Let's replace statsServer from the example with normal upstreams and monitoring with an Nginx-based monitoring service using single_listener and listening on port 8200.

Normal upstreams, health checks and monitoring, related changes

    haskell_run_service checkPeers $hs_service_healthcheck
        'hs_service_healthcheck
         Conf { upstreams     = ["u_backend"
                                ,"u_backend0"
                                ,"u_backend1"
                                ]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Just Endpoint { epUrl = "/healthcheck"
                                              , epProto = Http
                                              , epPassRule = DefaultPassRule
                                              }
              , sendStatsPort = Just 8200
              }
        ';

# ...


    haskell_run_service checkPeers $hs_service_healthcheck0
        'hs_service_healthcheck0
         Conf { upstreams     = ["u_backend"]
              , interval      = Sec 5
              , peerTimeout   = Sec 2
              , endpoint      = Just Endpoint { epUrl = "/healthcheck"
                                              , epProto = Http
                                              , epPassRule =
                                                      PassRuleByHttpStatus
                                                      [200, 404]
                                              }
              , sendStatsPort = Just 8200
              }
        ';

# ...

    haskell_var_empty_on_error $hs_stats;

# ...

        location /stat {
            allow 127.0.0.1;
            deny all;
            proxy_pass http://127.0.0.1:8200;
        }

# ...

    server {
        listen          8200 reuseport;
        server_name     stats;

        single_listener on;

        location /report {
            haskell_run_async_on_request_body receiveStats $hs_stats "Min 1";

            if ($hs_stats = '') {
                return 400;
            }

            return 200;
        }

        location /stat {
            haskell_async_content sendStats noarg;
        }

        location /stat/merge {
            haskell_async_content sendMergedStats noarg;
        }
    }
}

Handler receiveStats accepts a time interval corresponding to the value of ssPurgeInterval from service statsServer. If the value is not readable (say, noarg) then it is defaulted to Min 5.

Notice that the monitoring virtual server must listen on address 127.0.0.1 because service checkPeers reports stats to this address.

Periodic checks of healthy peers

The plugin checks periodically only faulty peers. Healthy peers become faulty if they do not pass the standard Nginx rules, but this may happen only when client requests trigger forwarding to the corresponding upstreams. Sometimes, it makes sense to check healthy peers unconditionally. One way to achieve this is running periodic active checks of the corresponding upstreams. For this, let's use Haskell module NgxExport.Tools.Subrequest.

Below is the source code of the shared library.

{-# LANGUAGE TemplateHaskell #-}

module NgxHealthcheckPeriodic where

import           NgxExport
import           NgxExport.Tools
import           NgxExport.Tools.Subrequest

import           NgxExport.Healthcheck ()

import           Data.ByteString (ByteString)
import qualified Data.ByteString.Lazy as L

makeRequest :: ByteString -> Bool -> IO L.ByteString
makeRequest = const . makeSubrequest

ngxExportSimpleService 'makeRequest $ PersistentService $ Just $ Sec 10

The periodic services must be declared in the Nginx configuration for all the upstreams to check.

    haskell_run_service simpleService_makeRequest $hs_check_u_backend
            '{"uri": "http://127.0.0.1:8010/Local/check/0/u_backend"}';

    haskell_run_service simpleService_makeRequest $hs_check_u_backend0
            '{"uri": "http://127.0.0.1:8010/Local/check/0/u_backend0"}';

    haskell_run_service simpleService_makeRequest $hs_check_u_backend1
            '{"uri": "http://127.0.0.1:8010/Local/check/0/u_backend1"}';

Location /Local/check/0/ shall proxy to the specified upstream.

        location ~ ^/Local/check/0/(.+) {
            allow 127.0.0.1;
            deny all;
            haskell_run_async makeSubrequest $hs_subrequest
                    '{"uri": "http://127.0.0.1:8010/Local/check/1/$1"}';
            return 200;
        }

        location ~ ^/Local/check/1/(.+) {
            allow 127.0.0.1;
            deny all;
            proxy_pass http://$1/healthcheck;
        }

The intermediate makeSubrequest catches possible 5xx and other bad HTTP statuses received from the upstream to prevent the periodic checks from throwing exceptions. Additionally, checking the response status in the intermediate location can be used for alerting that all servers in the upstream have failed.

Sadly, the subrequests may come to arbitrary worker processes which means that there is no guarantee that faulty peers in normal upstreams will be checked in all worker processes! The worker processes get chosen not only randomly but also not fairly: a single worker process may serve all incoming requests during a very long time. To make load between the workers more uniform, we can forward the subrequests to some dedicated virtual server listening on a port with reuseport and, additionally, randomize the hash for the reuseport by always changing ports of the subrequests with HTTP request header Connection: close.

    haskell_run_service simpleService_makeRequest $hs_check_u_backend
            '{"uri": "http://127.0.0.1:8011/Local/check/0/u_backend"}';

    haskell_run_service simpleService_makeRequest $hs_check_u_backend0
            '{"uri": "http://127.0.0.1:8011/Local/check/0/u_backend0"}';

    haskell_run_service simpleService_makeRequest $hs_check_u_backend1
            '{"uri": "http://127.0.0.1:8011/Local/check/0/u_backend1"}';

    server {
        listen          8011 reuseport;
        server_name     aux_fair_load;

        location ~ ^/Local/check/0/(.+) {
            allow 127.0.0.1;
            deny all;
            haskell_run_async makeSubrequest $hs_subrequest
                    '{"uri": "http://127.0.0.1:8011/Local/check/1/$1"
                     ,"headers": [["Connection", "close"]]
                     }';
            return 200;
        }

        location ~ ^/Local/check/1/(.+) {
            allow 127.0.0.1;
            deny all;
            proxy_pass http://$1/healthcheck;
        }
    }

As soon as faulty servers from normal upstreams may still appear arbitrarily in different worker processes, it makes sense to monitor them using the merged view, i.e. via URL /stat/merge.

Collecting Prometheus metrics

With modules NgxExport.Tools.Prometheus, NgxExport.Tools.Subrequest, and nginx-custom-counters-module, custom Prometheus metrics can be collected. Let's monitor the number of currently failed peers in all checked upstreams by extending examples from section Examples.

Below is the source code of the shared library.

{-# LANGUAGE TemplateHaskell, TypeApplications, ViewPatterns #-}

module NgxHealthcheckPrometheus where

import           NgxExport
import           NgxExport.Tools
import           NgxExport.Tools.Prometheus ()
import           NgxExport.Tools.Subrequest ()

import           NgxExport.Healthcheck

import qualified Data.Map.Strict as M
import qualified Data.Map.Lazy as ML
import           Data.ByteString (ByteString)
import qualified Data.ByteString.Lazy as L
import qualified Data.ByteString.Char8 as C8
import qualified Data.ByteString.Lazy.Char8 as C8L
import qualified Data.Text.Encoding as T
import           Data.Binary
import           Data.Maybe

type MergedStats = MServiceKey AnnotatedFlatPeers
type SharedStats = MServiceKey FlatPeers
type FlatStats = MUpstream Int

toFlatStats :: MServiceKey a -> FlatStats
toFlatStats = ML.foldr (flip $ M.foldrWithKey $
                           \k v -> M.alter (setN $ length v) k
                       ) M.empty
    where setN n Nothing = Just n
          setN n (Just m) = Just $ max n m

mergedFlatStats :: ByteString -> L.ByteString
mergedFlatStats =
    encode . toFlatStats . fromJust . readFromByteStringAsJSON @MergedStats

ngxExportYY 'mergedFlatStats

sharedFlatStats :: ByteString -> L.ByteString
sharedFlatStats =
    encode . toFlatStats . fromJust . readFromByteStringAsJSON @SharedStats

ngxExportYY 'sharedFlatStats

nFailedServers :: ByteString -> L.ByteString
nFailedServers v =
    let (T.decodeUtf8 -> u, decode @FlatStats . L.fromStrict . C8.tail -> s) =
            C8.break (== '|') v
    in C8L.pack $ show $ fromMaybe 0 $ M.lookup u s

ngxExportYY 'nFailedServers

Normal upstreams, related changes

    haskell_run_service simpleService_prometheusConf $hs_prometheus_conf
            'PrometheusConf
                { pcMetrics = fromList
                    [("cnt_upstream_failure"
                      ,"Number of servers which are currently failed")
                    ]
                , pcGauges = fromList
                    ["cnt_upstream_failure@upstream=(u_backend)"
                    ,"cnt_upstream_failure@upstream=(u_backend0)"
                    ,"cnt_upstream_failure@upstream=(u_backend1)"
                    ]
                , pcScale1000 = fromList []
                }';

# ...

        location /stat {
            allow 127.0.0.1;
            deny all;
            proxy_pass http://127.0.0.1:8100;
        }

        location /stat/merge {
            allow 127.0.0.1;
            deny all;

            counter $cnt_upstream_failure@upstream=(u_backend)
                    set $hs_n_u_backend;
            counter $cnt_upstream_failure@upstream=(u_backend0)
                    set $hs_n_u_backend0;
            counter $cnt_upstream_failure@upstream=(u_backend1)
                    set $hs_n_u_backend1;

            haskell_run_async makeSubrequestFull $hs_subrequest
                    '{"uri": "http://127.0.0.1:8100/stat/merge"}';
            haskell_run extractBodyFromFullResponse $hs_subrequest_body
                    $hs_subrequest;
            haskell_run mergedFlatStats $hs_failed_backends
                    $hs_subrequest_body;
            haskell_run nFailedServers $hs_n_u_backend
                    u_backend|$hs_failed_backends;
            haskell_run nFailedServers $hs_n_u_backend0
                    u_backend0|$hs_failed_backends;
            haskell_run nFailedServers $hs_n_u_backend1
                    u_backend1|$hs_failed_backends;

            haskell_content fromFullResponse $hs_subrequest;
        }

        location /metrics {
            allow 127.0.0.1;
            deny all;

            haskell_run_async makeSubrequest $hs_subrequest
                    '{"uri": "http://127.0.0.1:8010/stat/merge"}';
            haskell_async_content prometheusMetrics 
                    '["main", $cnt_collection, {}, {}]';
        }

Shared upstreams, related changes

    haskell_run_service simpleService_prometheusConf $hs_prometheus_conf
            'PrometheusConf
                { pcMetrics = fromList
                    [("cnt_upstream_failure"
                      ,"Number of servers which are currently failed")
                    ]
                , pcGauges = fromList
                    ["cnt_upstream_failure@upstream=(u_backend)"
                    ,"cnt_upstream_failure@upstream=(u_backend0)"
                    ,"cnt_upstream_failure@upstream=(u_backend1)"
                    ]
                , pcScale1000 = fromList []
                }';

# ...

        location /stat {
            allow 127.0.0.1;
            deny all;
            haskell_async_content reportPeers;
        }

        location /stat/shared {
            allow 127.0.0.1;
            deny all;

            counter $cnt_upstream_failure@upstream=(u_backend)
                    set $hs_n_u_backend;
            counter $cnt_upstream_failure@upstream=(u_backend0)
                    set $hs_n_u_backend0;
            counter $cnt_upstream_failure@upstream=(u_backend1)
                    set $hs_n_u_backend1;

            haskell_run_async makeSubrequestFull $hs_subrequest
                    '{"uri": "http://127.0.0.1:8010/stat"}';
            haskell_run extractBodyFromFullResponse $hs_subrequest_body
                    $hs_subrequest;
            haskell_run sharedFlatStats $hs_failed_backends
                    $hs_subrequest_body;
            haskell_run nFailedServers $hs_n_u_backend
                    u_backend|$hs_failed_backends;
            haskell_run nFailedServers $hs_n_u_backend0
                    u_backend0|$hs_failed_backends;
            haskell_run nFailedServers $hs_n_u_backend1
                    u_backend1|$hs_failed_backends;

            haskell_content fromFullResponse $hs_subrequest;
        }

        location /metrics {
            allow 127.0.0.1;
            deny all;

            haskell_run_async makeSubrequest $hs_subrequest
                    '{"uri": "http://127.0.0.1:8010/stat/shared"}';
            haskell_async_content prometheusMetrics
                    '["main", $cnt_collection, {}, {}]';
        }

Below is a typical sample of the Prometheus metrics.

# HELP cnt_upstream_failure Number of servers which are currently failed
# TYPE cnt_upstream_failure gauge
cnt_upstream_failure{upstream="u_backend"} 2.0
cnt_upstream_failure{upstream="u_backend0"} 0.0
cnt_upstream_failure{upstream="u_backend1"} 0.0

Corner cases

• In Nginx before version 1.12.0, there was a so-called quick recovery mechanism, meaning that when all peers in the main or the backup part of an upstream get failed, then they immediately were recovered. As such, health checks are impossible in such a case.

• When the main or the backup part of an upstream contains only one peer, then the peer never fails. For the sake of the health checks, such an upstream could be extended by a fake server marked as down, for example

      upstream u_backend {
          server localhost:8020;
          server localhost:8020 down;
      }

  But there is a better solution proposed in Nginx module nginx-combined-upstreams-module.

      upstream u_backend {
          server localhost:8020;
          extend_single_peers;
      }

  Directive extend_single_peers adds a fake peer only when an upstream part (the main or the backup) contains a single peer.

Building and installation

The plugin contains Haskell and C parts, and thus, it requires ghc, Cabal, gcc, and a directory with the Nginx sources. The build tool also requires patchelf and utility nhm-tool which is shipped with package ngx-export-distribution.

Let's first install the Nginx module. For this, go to the directory with the Nginx source code,

$ cd /path/to/nginx/sources

compile,

$ ./configure --add-dynamic-module=/path/to/nginx-healthcheck-plugin/sources
$ make modules

and install ngx_healthcheck_plugin.so.

$ export NGX_HS_INSTALL_DIR=/var/lib/nginx
$ sudo install -d $NGX_HS_INSTALL_DIR
$ sudo cp objs/ngx_healthcheck_plugin.so $NGX_HS_INSTALL_DIR/libngx_healthcheck_plugin.so

Notice that we added prefix lib to the module's name!

Now let's build the Haskell code. For this, go to one of the directories with Haskell handlers: simple/, periodic/, or prometheus/.

$ cd -
$ cd simple

Before running make, tune the constraints stanza in cabal.project. Currently, it should look similar to

constraints: ngx-export-healthcheck +snapstatsserver +healthcheckhttps

This line enforces building the Snap monitoring server and support for secure connections to endpoints. To disable them, replace +snapstatsserver by -snapstatsserver and +healthcheckhttps by -healthcheckhttps. To let Cabal deduce whether to build these features automatically, remove the constraints.

Now run

$ make PREFIX=$NGX_HS_INSTALL_DIR
$ sudo make PREFIX=$NGX_HS_INSTALL_DIR install

or simply

$ make
$ sudo make install

if installation directory is /var/lib/nginx/.

With ghc older than 9.0.1, build with

$ make LINKRTS=-lHSrts_thr-ghc$(ghc --numeric-version)

By default, package ngx-export-healthcheck gets installed from Hackage. To build it locally, augment stanza packages inside cabal.project according to the commentary attached to it.