nixos-config/machines/renge/services/prometheus.nix

{ config, lib, pkgs, ... }:
let
  cfg = config.services.prometheus;

  mkStaticTargets = targets: lib.singleton { inherit targets; };
  mkStaticTarget = target: mkStaticTargets (lib.singleton target);
in
{
  services.prometheus = {
    enable = true;
    listenAddress = "127.0.0.1";
    webExternalUrl = "https://prometheus.sbruder.de";
    globalConfig = {
      scrape_interval = "15s";
      evaluation_interval = "15s";
    };
    extraFlags = [
      "--storage.tsdb.retention.time=90d"
      "--web.enable-admin-api"
    ];

    alertmanagers = [
      {
        static_configs = mkStaticTarget "${cfg.alertmanager.listenAddress}:${toString cfg.alertmanager.port}";
        path_prefix = "/alertmanager/";
      }
    ];
    alertmanager = {
      enable = true;
      listenAddress = "127.0.0.1";
      webExternalUrl = "https://prometheus.sbruder.de/alertmanager";
      configuration = {
        global.resolve_timeout = "2m";

        route = {
          receiver = "matrix";
          group_by = [ "alertname" ];
          group_wait = "3m";
        };

        receivers = [
          {
            name = "matrix";
            webhook_configs = lib.singleton {
              url = (lib.elemAt
                (lib.filter
                  ({ ID, ... }: ID == "alertmanager_service")
                  config.services.go-neb.config.services)
                0).Config.webhook_url;
            };
          }
        ];
      };
    };

    scrapeConfigs = [
      {
        job_name = "prometheus";
        static_configs = mkStaticTarget "localhost:${toString cfg.port}";
      }
      {
        job_name = "node";
        static_configs = mkStaticTargets [
          "fuuko.vpn.sbruder.de:9100"
          "mayushii.vpn.sbruder.de:9100"
          "nunotaba.vpn.sbruder.de:9100"
          "renge.vpn.sbruder.de:9100"
          "hitagi.vpn.sbruder.de:9100"
          "vueko.vpn.sbruder.de:9100"
          "okarin.vpn.sbruder.de:9100"
          "shinobu.vpn.sbruder.de:9100"
          "nazuna.vpn.sbruder.de:9100"
        ];
      }
      {
        job_name = "qbittorrent";
        static_configs = mkStaticTargets [
          "fuuko.vpn.sbruder.de:9561"
        ];
        relabel_configs = lib.singleton {
          target_label = "instance";
          source_labels = lib.singleton "__address__";
          regex = "(.*)\\.vpn\\.sbruder\\.de:9561";
        };
      }
      (
        let
          listenerCfg = (lib.elemAt config.services.matrix-synapse.settings.listeners 0);
        in
        {
          job_name = "synapse";
          static_configs = mkStaticTarget "${lib.elemAt listenerCfg.bind_addresses 0}:${toString listenerCfg.port}";
          metrics_path = "/_synapse/metrics";
          relabel_configs = lib.singleton {
            target_label = "instance";
            replacement = "matrix.sbruder.de";
          };
        }
      )
      {
        job_name = "dnsmasq";
        static_configs = mkStaticTarget "shinobu.vpn.sbruder.de:${toString config.services.prometheus.exporters.dnsmasq.port}";
        relabel_configs = lib.singleton {
          target_label = "instance";
          replacement = "shinobu.home.sbruder.de";
        };
      }
      {
        job_name = "hcloud";
        static_configs = mkStaticTarget config.services.hcloud_exporter.listenAddress;
      }
      {
        job_name = "co2";
        static_configs = mkStaticTarget "shinobu.vpn.sbruder.de:9672";
      }
      {
        job_name = "rspamd";
        static_configs = mkStaticTarget "vueko.vpn.sbruder.de";
        metrics_path = "/rspamd/metrics";
        relabel_configs = lib.singleton {
          target_label = "instance";
          replacement = "vueko.sbruder.de";
        };
      }
    ];

    rules =
      let
        mkAlert = { name, expr, for ? "1m", description ? null }: {
          alert = name;
          inherit expr for;
          annotations = lib.optionalAttrs (description != null) { inherit description; };
        };
      in
      [
        (lib.generators.toYAML { } {
          groups = lib.singleton {
            name = "alert.rules";
            rules = map mkAlert [
              {
                name = "InstanceDown";
                expr = ''up{instance!~"(nunotaba|hitagi|mayushii|fuuko).vpn.sbruder.de:.*"} == 0'';
                description = "Instance {{ $labels.instance }} of job {{ $labels.job }} has been down for more than 1 minutes.";
              }
              {
                name = "SystemdUnitFailed";
                expr = ''node_systemd_unit_state{state="failed"} == 1'';
                description = "Systemd unit {{ $labels.name }} on {{ $labels.instance }} has state failed.";
              }
              {
                name = "NodeHighLoad";
                expr = ''sum by (instance) (node_load15) / count by (instance) (node_cpu_seconds_total{mode="system"}) > 2'';
                for = "15m";
                description = "Node {{ $labels.instance }} is having a per-core load ≥ 2 for the last 15 minutes.";
              }
              {
                name = "NodeHighMemory";
                expr = ''(node_memory_MemTotal_bytes - node_memory_MemAvailable_bytes) / node_memory_MemTotal_bytes > 0.9'';
                for = "2m";
                description = "Node {{ $labels.instance }} is using more than 90 % of available RAM.";
              }
              {
                name = "TP440ACPIBroken";
                expr = ''node_hwmon_temp_celsius{chip="thermal_thermal_zone0",instance="nunotaba.vpn.sbruder.de:9100",job="node",sensor="temp1"} == 48'';
                for = "10m";
                description = "Thinkpad T440’s ACPI temperature is broken. Its reported temperature is 48 °C for the last 10 minutes. That doesn’t seem right. Try suspending";
              }
              {
                name = "TorrentNoPeers";
                expr = "sum by (instance) (qBittorrent_torrent_connected_leechs) == 0";
                description = "qBittorrent instance {{ $labels.instance }} has no peers. There might be a network connectivity problem";
              }
              # <40% is to account for /boot being full (which causes ugly errors on rebuild)
              {
                name = "DiskFull";
                expr = ''node_filesystem_free_bytes{fstype!~"ramfs|tmpfs", mountpoint!~"/nix/store"} / node_filesystem_size_bytes{fstype!~"ramfs|tmpfs", mountpoint!~"/nix/store"} < 0.4 and node_filesystem_free_bytes{fstype!~"ramfs|tmpfs", mountpoint!~"/nix/store"} < 4*1024^3'';
                description = "Device {{ $labels.device }} on {{ $labels.instance }}:{{ $labels.mountpoint }} has less than 4GiB free space while being used over 40%";
              }
            ];
          };
        })
      ];
  };

  # exporters that are not part of nixpkgs’ prometheus infrastructure
  services.hcloud_exporter = {
    enable = true;
    listenAddress = "127.0.0.1:9501";
    environmentFile = config.sops.secrets.hcloud_exporter-environment.path;
  };
  sops.secrets.hcloud_exporter-environment.sopsFile = ../secrets.yaml;

  sops.secrets.prometheus-htpasswd = {
    owner = "nginx";
    sopsFile = ../secrets.yaml;
  };

  services.nginx.virtualHosts."prometheus.sbruder.de" = {
    enableACME = true;
    forceSSL = true;

    basicAuthFile = config.sops.secrets.prometheus-htpasswd.path;

    locations = {
      "/".proxyPass = "http://${cfg.listenAddress}:${toString cfg.port}";

      "/alertmanager/".proxyPass = "http://${cfg.alertmanager.listenAddress}:${toString cfg.alertmanager.port}";
    };
  };
}