modules/nixos/monitoring: add alertmanager, matrix-alertmanager
This commit is contained in:
zowoq
parent
0764cd9792
commit
0db08b6881
6 changed files with 308 additions and 4 deletions
|
|
@ -1,4 +1,5 @@
|
|||
ssh_host_ed25519_key: ENC[AES256_GCM,data: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,iv:Z6SfUFsjfRaVc23CNM1NE4/c92MLmbdEXilPJomX9qM=,tag:xknd9rqBVvUg69ICvhXHcA==,type:str]
|
||||
nix-community-matrix-bot-token: ENC[AES256_GCM,data:uWqZe3tC5bGSC78SfVH0uwjnHBU0RdgglOlVUQo99AYykdzeAH9toTk2ZgtP97L1kKdyQo70f7QcNSCqtbmB0Q==,iv:xPob8PjdF0ha9u4lwFOWRJvSD+xUMIL6uw6OogMVP+g=,tag:gx0tIiiH6eNiapcEVZaiBw==,type:str]
|
||||
sops:
|
||||
kms: []
|
||||
gcp_kms: []
|
||||
|
|
@ -59,8 +60,8 @@ sops:
|
|||
QnJZZzN1a1M5b1dwa3hvL3ZHYkpxQUkK1g9sQB0UHl9coaznjIn4WDpQv21Y8cl9
|
||||
LNqnv0Q6KrxNliq2JEJoEpjD5+xTcqV/5FgylKhtdNWUZ0eAX8taog==
|
||||
-----END AGE ENCRYPTED FILE-----
|
||||
lastmodified: "2023-07-29T05:26:34Z"
|
||||
mac: ENC[AES256_GCM,data:HS8Jr5pHtANiytEOAYYja3b+FxyCb858pTFZvi1ZZ2NBkjRmkOY8UEzoL+dEJQ2RQ49l3GktIwu3oBwkjjoBHC7cqo5VfwB7a23u28iWwfiHduZMOOu4xHg6vsUCtScu1tr3bJexfVu47RHI/su/ds3UWk6eJKBm49MzcGTI7lU=,iv:J1lgkh9mSmd3iUf9pkvJAAsPgDZQsNtjMeBTwm+nhdQ=,tag:3IMkozS5d6jRoz+Gl8K0BA==,type:str]
|
||||
lastmodified: "2023-08-12T01:00:57Z"
|
||||
mac: ENC[AES256_GCM,data:GqsV7ULY9kWLooThcpJulxkJgxP92kY9OEWETuvO1Do9uiUaJUpmmPcwu7mhAEZzys4U+wZEfpRvfrrJanCotGCeNsfGh9zAbgHn4T5P8wqOyTWns5qaQDQki1Fs83CyGuCUjwPUWcH6euLYUlWANYcUbIZms9sb1l0bo7MxZ18=,iv:xqqpUsc5nfHmB+DX9S5fNPrdDqPBd+IPtlr3GnS/jfU=,tag:pPLlZ2q8ZxQJFHBxGTJRdA==,type:str]
|
||||
pgp: []
|
||||
unencrypted_suffix: _unencrypted
|
||||
version: 3.7.3
|
||||
|
|
|
|||
191
modules/nixos/monitoring/alert-rules.nix
Normal file
191
modules/nixos/monitoring/alert-rules.nix
Normal file
|
|
@ -0,0 +1,191 @@
|
|||
{ lib }:
|
||||
lib.mapAttrsToList
|
||||
(name: opts: {
|
||||
alert = name;
|
||||
expr = opts.condition;
|
||||
for = opts.time or "2m";
|
||||
labels = { };
|
||||
annotations.description = opts.description;
|
||||
# for matrix alert-receiver
|
||||
annotations.summary = opts.description;
|
||||
})
|
||||
((lib.genAttrs [
|
||||
"borgbackup-job-github-org.service"
|
||||
"borgbackup-job-nixpkgs-update.service"
|
||||
]
|
||||
(name: {
|
||||
condition = ''absent_over_time(task_last_run{name="${name}"}[1d])'';
|
||||
description = "status of ${name} is unknown: no data for a day";
|
||||
})
|
||||
) // {
|
||||
prometheus_too_many_restarts = {
|
||||
condition = ''changes(process_start_time_seconds{job=~"prometheus|pushgateway|alertmanager|telegraf"}[15m]) > 2'';
|
||||
description = "Prometheus has restarted more than twice in the last 15 minutes. It might be crashlooping";
|
||||
};
|
||||
|
||||
alert_manager_config_not_synced = {
|
||||
condition = ''count(count_values("config_hash", alertmanager_config_hash)) > 1'';
|
||||
description = "Configurations of AlertManager cluster instances are out of sync";
|
||||
};
|
||||
|
||||
#alert_manager_e2e_dead_man_switch = {
|
||||
# condition = "vector(1)";
|
||||
# description = "Prometheus DeadManSwitch is an always-firing alert. It's used as an end-to-end test of Prometheus through the Alertmanager.";
|
||||
#};
|
||||
|
||||
prometheus_not_connected_to_alertmanager = {
|
||||
condition = "prometheus_notifications_alertmanagers_discovered < 1";
|
||||
description = "Prometheus cannot connect the alertmanager\n VALUE = {{ $value }}\n LABELS = {{ $labels }}";
|
||||
};
|
||||
|
||||
prometheus_rule_evaluation_failures = {
|
||||
condition = "increase(prometheus_rule_evaluation_failures_total[3m]) > 0";
|
||||
description = "Prometheus encountered {{ $value }} rule evaluation failures, leading to potentially ignored alerts.\n VALUE = {{ $value }}\n LABELS = {{ $labels }}";
|
||||
};
|
||||
|
||||
prometheus_template_expansion_failures = {
|
||||
condition = "increase(prometheus_template_text_expansion_failures_total[3m]) > 0";
|
||||
time = "0m";
|
||||
description = "Prometheus encountered {{ $value }} template text expansion failures\n VALUE = {{ $value }}\n LABELS = {{ $labels }}";
|
||||
};
|
||||
|
||||
filesystem_full_80percent = {
|
||||
condition = ''disk_used_percent{mode!="ro"} >= 80'';
|
||||
time = "10m";
|
||||
description = "{{$labels.instance}} device {{$labels.device}} on {{$labels.path}} got less than 20% space left on its filesystem";
|
||||
};
|
||||
|
||||
filesystem_inodes_full = {
|
||||
condition = ''disk_inodes_free / disk_inodes_total < 0.10'';
|
||||
time = "10m";
|
||||
description = "{{$labels.instance}} device {{$labels.device}} on {{$labels.path}} got less than 10% inodes left on its filesystem";
|
||||
};
|
||||
|
||||
daily_task_not_run = {
|
||||
# give 6 hours grace period
|
||||
condition = ''time() - task_last_run{state="ok",frequency="daily"} > (24 + 6) * 60 * 60'';
|
||||
description = "{{$labels.host}}: {{$labels.name}} was not run in the last 24h";
|
||||
};
|
||||
|
||||
daily_task_failed = {
|
||||
condition = ''task_last_run{state="fail"}'';
|
||||
description = "{{$labels.host}}: {{$labels.name}} failed to run";
|
||||
};
|
||||
|
||||
nixpkgs_out_of_date = {
|
||||
condition = ''(time() - flake_input_last_modified{input="nixpkgs"}) / (60*60*24) > 7'';
|
||||
description = "{{$labels.host}}: nixpkgs flake is older than a week";
|
||||
};
|
||||
|
||||
swap_using_30percent = {
|
||||
condition = ''mem_swap_total - (mem_swap_cached + mem_swap_free) > mem_swap_total * 0.3'';
|
||||
time = "30m";
|
||||
description = "{{$labels.host}} is using 30% of its swap space for at least 30 minutes";
|
||||
};
|
||||
|
||||
# user@$uid.service and similar sometimes fail, we don't care about those services.
|
||||
systemd_service_failed = {
|
||||
condition = ''systemd_units_active_code{name!~"user@\\d+.service"} == 3'';
|
||||
description = "{{$labels.host}} failed to (re)start service {{$labels.name}}";
|
||||
};
|
||||
|
||||
ram_using_95percent = {
|
||||
condition = "mem_buffered + mem_free + mem_cached < mem_total * 0.05";
|
||||
time = "1h";
|
||||
description = "{{$labels.host}} is using at least 95% of its RAM for at least 1 hour";
|
||||
};
|
||||
|
||||
load15 = {
|
||||
condition = ''system_load15 / system_n_cpus{instance!~"build.*.nix-community.org:9273"} >= 2.0'';
|
||||
time = "10m";
|
||||
description = "{{$labels.host}} is running with load15 > 1 for at least 5 minutes: {{$value}}";
|
||||
};
|
||||
|
||||
reboot = {
|
||||
condition = "system_uptime < 600";
|
||||
description = "{{$labels.host}} just rebooted";
|
||||
};
|
||||
|
||||
uptime = {
|
||||
condition = ''system_uptime > (60*60*24*14)'';
|
||||
description = "Uptime monster: {{$labels.host}} has been up for more than 14 days";
|
||||
};
|
||||
|
||||
telegraf_down = {
|
||||
condition = ''min(up{job=~"telegraf"}) by (source, job, instance, org) == 0'';
|
||||
time = "3m";
|
||||
description = "{{$labels.instance}}: {{$labels.job}} telegraf exporter from {{$labels.source}} is down";
|
||||
};
|
||||
|
||||
http = {
|
||||
condition = "http_response_result_code != 0";
|
||||
description = "{{$labels.server}} : http request failed from {{$labels.instance}}: {{$labels.result}}";
|
||||
};
|
||||
|
||||
http_match_failed = {
|
||||
condition = "http_response_response_string_match == 0";
|
||||
description = "{{$labels.server}} : http body not as expected; status code: {{$labels.status_code}}";
|
||||
};
|
||||
|
||||
connection_failed = {
|
||||
condition = "net_response_result_code != 0";
|
||||
description = "{{$labels.server}}: connection to {{$labels.port}}({{$labels.protocol}}) failed from {{$labels.instance}}";
|
||||
};
|
||||
|
||||
zfs_errors = {
|
||||
condition = "zfs_arcstats_l2_io_error + zfs_dmu_tx_error + zfs_arcstats_l2_writes_error > 0";
|
||||
description = "{{$labels.instance}} reports: {{$value}} ZFS IO errors";
|
||||
};
|
||||
|
||||
zpool_status = {
|
||||
condition = "zpool_status_errors > 0";
|
||||
description = "{{$labels.instance}} reports: zpool {{$labels.name}} has {{$value}} errors";
|
||||
};
|
||||
|
||||
mdraid_degraded_disks = {
|
||||
condition = "mdstat_degraded_disks > 0";
|
||||
description = "{{$labels.instance}}: raid {{$labels.dev}} has failed disks";
|
||||
};
|
||||
|
||||
# ignore devices that disabled S.M.A.R.T (example if attached via USB)
|
||||
# Also ignore build02, build03
|
||||
smart_errors = {
|
||||
condition = ''smart_device_health_ok{enabled!="Disabled", instance!~"(build02|build03).nix-community.org:9273"} != 1'';
|
||||
description = "{{$labels.instance}}: S.M.A.R.T reports: {{$labels.device}} ({{$labels.model}}) has errors";
|
||||
};
|
||||
|
||||
oom_kills = {
|
||||
condition = "increase(kernel_vmstat_oom_kill[5m]) > 0";
|
||||
description = "{{$labels.instance}}: OOM kill detected";
|
||||
};
|
||||
|
||||
unusual_disk_read_latency = {
|
||||
condition = "rate(diskio_read_time[1m]) / rate(diskio_reads[1m]) > 0.1 and rate(diskio_reads[1m]) > 0";
|
||||
description = "{{$labels.instance}}: Disk latency is growing (read operations > 100ms)";
|
||||
};
|
||||
|
||||
unusual_disk_write_latency = {
|
||||
condition = "rate(diskio_write_time[1m]) / rate(diskio_write[1m]) > 0.1 and rate(diskio_write[1m]) > 0";
|
||||
description = "{{$labels.instance}}: Disk latency is growing (write operations > 100ms)";
|
||||
};
|
||||
|
||||
ipv6_dad_check = {
|
||||
condition = "ipv6_dad_failures_count > 0";
|
||||
description = "{{$labels.host}}: {{$value}} assigned ipv6 addresses have failed duplicate address check";
|
||||
};
|
||||
|
||||
host_memory_under_memory_pressure = {
|
||||
condition = "rate(node_vmstat_pgmajfault[1m]) > 1000";
|
||||
description = "{{$labels.instance}}: The node is under heavy memory pressure. High rate of major page faults: {{$value}}";
|
||||
};
|
||||
|
||||
ext4_errors = {
|
||||
condition = "ext4_errors_value > 0";
|
||||
description = "{{$labels.instance}}: ext4 has reported {{$value}} I/O errors: check /sys/fs/ext4/*/errors_count";
|
||||
};
|
||||
|
||||
alerts_silences_changed = {
|
||||
condition = ''abs(delta(alertmanager_silences{state="active"}[1h])) >= 1'';
|
||||
description = "alertmanager: number of active silences has changed: {{$value}}";
|
||||
};
|
||||
})
|
||||
|
|
@ -1,5 +1,6 @@
|
|||
{
|
||||
imports = [
|
||||
./matrix-alertmanager.nix
|
||||
./prometheus.nix
|
||||
./telegraf.nix
|
||||
];
|
||||
|
|
|
|||
46
modules/nixos/monitoring/matrix-alertmanager.nix
Normal file
46
modules/nixos/monitoring/matrix-alertmanager.nix
Normal file
|
|
@ -0,0 +1,46 @@
|
|||
{ config, pkgs, ... }:
|
||||
let
|
||||
matrix-alertmanager-receiver = pkgs.buildGoModule rec {
|
||||
pname = "matrix-alertmanager-receiver";
|
||||
version = "0.1.2";
|
||||
src = pkgs.fetchFromSourcehut {
|
||||
owner = "~fnux";
|
||||
repo = "matrix-alertmanager-receiver";
|
||||
rev = version;
|
||||
hash = "sha256-F6Cn0lmASAjWGEBCmyLdfz4r06fDTEfZQcynfA/RRtI=";
|
||||
};
|
||||
vendorHash = "sha256-7tRCX9FzOsLXCTWWjLp3hr1kegt1dxsbCKfC7tICreo=";
|
||||
};
|
||||
in
|
||||
{
|
||||
sops.secrets.nix-community-matrix-bot-token = { };
|
||||
sops.templates."config.toml".content = ''
|
||||
Homeserver = "https://matrix-client.matrix.org"
|
||||
TargetRoomID = "!cBybDCkeRlSWfuaFvn:numtide.com"
|
||||
MXID = "@nix-community-matrix-bot:matrix.org"
|
||||
MXToken = "${config.sops.placeholder.nix-community-matrix-bot-token}"
|
||||
HTTPPort = 9088
|
||||
HTTPAddress = "localhost"
|
||||
'';
|
||||
sops.templates."config.toml".owner = "matrix-alertmanager-receiver";
|
||||
|
||||
users.users.matrix-alertmanager-receiver = {
|
||||
isSystemUser = true;
|
||||
group = "matrix-alertmanager-receiver";
|
||||
};
|
||||
users.groups.matrix-alertmanager-receiver = { };
|
||||
|
||||
systemd.services.matrix-alertmanager-receiver = {
|
||||
description = "Matrix Alertmanager Receiver";
|
||||
after = [ "network.target" ];
|
||||
wantedBy = [ "multi-user.target" ];
|
||||
serviceConfig = {
|
||||
Type = "simple";
|
||||
ExecStart = "${matrix-alertmanager-receiver}/bin/matrix-alertmanager-receiver -config ${config.sops.templates."config.toml".path}";
|
||||
Restart = "always";
|
||||
RestartSec = "10";
|
||||
User = "matrix-alertmanager-receiver";
|
||||
Group = "matrix-alertmanager-receiver";
|
||||
};
|
||||
};
|
||||
}
|
||||
|
|
@ -1,6 +1,17 @@
|
|||
{ pkgs, ... }:
|
||||
{
|
||||
services.prometheus = {
|
||||
enable = true;
|
||||
ruleFiles = [
|
||||
(pkgs.writeText "prometheus-rules.yml" (builtins.toJSON {
|
||||
groups = [
|
||||
{
|
||||
name = "alerting-rules";
|
||||
rules = import ./alert-rules.nix { inherit (pkgs) lib; };
|
||||
}
|
||||
];
|
||||
}))
|
||||
];
|
||||
webExternalUrl = "https://prometheus.nix-community.org";
|
||||
scrapeConfigs = [
|
||||
{
|
||||
|
|
@ -26,6 +37,15 @@
|
|||
];
|
||||
}
|
||||
];
|
||||
alertmanagers = [
|
||||
{
|
||||
static_configs = [
|
||||
{
|
||||
targets = [ "localhost:9093" ];
|
||||
}
|
||||
];
|
||||
}
|
||||
];
|
||||
};
|
||||
|
||||
services.nginx.virtualHosts."prometheus.nix-community.org" = {
|
||||
|
|
@ -33,4 +53,45 @@
|
|||
forceSSL = true;
|
||||
locations."/".proxyPass = "http://localhost:9090";
|
||||
};
|
||||
|
||||
services.prometheus.alertmanager = {
|
||||
enable = true;
|
||||
webExternalUrl = "https://alertmanager.nix-community.org";
|
||||
listenAddress = "[::1]";
|
||||
configuration = {
|
||||
route = {
|
||||
receiver = "default";
|
||||
routes = [
|
||||
{
|
||||
group_by = [ "host" ];
|
||||
match_re.org = "nix-community";
|
||||
group_wait = "5m";
|
||||
group_interval = "5m";
|
||||
repeat_interval = "4h";
|
||||
receiver = "nix-community";
|
||||
}
|
||||
];
|
||||
};
|
||||
receivers = [
|
||||
{
|
||||
name = "nix-community";
|
||||
webhook_configs = [
|
||||
{
|
||||
url = "http://localhost:9088/alert";
|
||||
max_alerts = 5;
|
||||
}
|
||||
];
|
||||
}
|
||||
{
|
||||
name = "default";
|
||||
}
|
||||
];
|
||||
};
|
||||
};
|
||||
|
||||
services.nginx.virtualHosts."alertmanager.nix-community.org" = {
|
||||
enableACME = true;
|
||||
forceSSL = true;
|
||||
locations."/".proxyPass = "http://localhost:9093";
|
||||
};
|
||||
}
|
||||
|
|
|
|||
|
|
@ -11,6 +11,10 @@ nix-community-alert-bot:
|
|||
username: ENC[AES256_GCM,data:nA+jbVz3HtDgfDC+y2J4818=,iv:8Q7roobuWuXMnJXWKEJRjKlMYmvFUm5eodX5htLbVHY=,tag:r0FEOL9PByNYtL17hU7ApQ==,type:str]
|
||||
home-server: ENC[AES256_GCM,data:k6rsrEvyCUQj,iv:TJBl7dYioATGVYOOETIajkaw+7LKsyrJDI0zmey3Uo4=,tag:ZuSahyrPlmj2qGp8vvuyOw==,type:str]
|
||||
password: ENC[AES256_GCM,data:a0DcDpeUNDZpPxDiLzk3,iv:fXpMBTbp5tmjWlhe5EieyvQ4+tVbuowTf26CevOj0oc=,tag:7VIFyHbia6aR8WabLOz9TA==,type:str]
|
||||
nix-community-matrix-bot:
|
||||
username: ENC[AES256_GCM,data:p37MEk8kk2lOoBDgxY4tgXhOAHs3vRDY,iv:xvRX+k/V2HwAF3Df4oXtRf/XzS1rriJJPZ8z1B2/fLo=,tag:NNovFZGV8XFlP8pCC5mmrQ==,type:str]
|
||||
home-server: ENC[AES256_GCM,data:5W8LDIlG8RThrQ==,iv:DjCo2cs0NoYBno1XuQ8cM9B1l6l6ZudDgOV15fZhxBo=,tag:wgYNxPXcpejteURF1IdXoQ==,type:str]
|
||||
password: ENC[AES256_GCM,data:4q4RHDDzPfoK8cqbr6K6crheEdur5V1YiddVJP35939vmXP+Sg==,iv:R9dDa1K3UW1wByRLazNJTz1GdT5nhux+q39X/UrL0t4=,tag:Cwt6ffjNCiGGt4atBs3TgQ==,type:str]
|
||||
nix-infra-bot:
|
||||
URL: ENC[AES256_GCM,data:ZKphMNdc0TjZTA==,iv:ufiUj89KXkDjbsFiKs4hBky2n/fZ6Y1oafKrNMPC2k4=,tag:z1A18GPaf6gSKgi8JORNtw==,type:str]
|
||||
User: ENC[AES256_GCM,data:TK7VamXfCTELo38vZQ==,iv:koHFX5tWA4kEOvLdXy8WkFF5GQX6NZpo8p5aGUc3WWo=,tag:AJNpBhXiTxoQ+h3UTgB1GQ==,type:str]
|
||||
|
|
@ -83,8 +87,8 @@ sops:
|
|||
MkcvL1JyVFBJV0Y5RFFCMGN1OUFXdU0Kdx1wy6ZOOTg1a6VKaq52SMBvC26lMsW/
|
||||
oMP+hmXc2WtoqZp+jZ9rrXz6cZW6/dO7CPqxl3aUEKg6BkXIwgyKeg==
|
||||
-----END AGE ENCRYPTED FILE-----
|
||||
lastmodified: "2023-07-18T06:28:59Z"
|
||||
mac: ENC[AES256_GCM,data:HpdZCHgZmxvNtgeAFhnwz1gL+MuJWZJiZ75p3+heiNftAhiR1CSIWB1uAcAm0SVN5/mDOWs9SKH2IQv2clra4hkVnELtNjZVt576+OSq7Dy1LXvFEyafp8Sb6nJfMN3FgT+Mx/Y4YcGiLhllX2crkLBfiJu32n++Aejj5I0RxNA=,iv:qjXkgjB1PBB2cGPsI5pELlRONfaDL1cT3NzFPdmMQUU=,tag:BAF2WUF7v/Tp5IkEQ3oVZA==,type:str]
|
||||
lastmodified: "2023-08-12T00:50:11Z"
|
||||
mac: ENC[AES256_GCM,data:E1MYlrlHS2fsCapkIPXx1ljOTgXnAqHqvOvk40Aubd6sJqYwGPVsOuH4uLZHTIrpeZYqX7e8VHaboV1qwobcIIUpvSWJV2AJGHj4cJpL43JAktp7ANJqMV4NoGjPqfIcNLCi6hJjM0wJDvNLglkQ7TPUQUoxNyp9knuQuaQNcgY=,iv:s+BU1mtqx14WZawZUxvYj4tmNtLGleFxkvsvpm/LIww=,tag:90eIIWQ8WoYDEDLzlqOIQg==,type:str]
|
||||
pgp: []
|
||||
unencrypted_suffix: _unencrypted
|
||||
version: 3.7.3
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue