关于运维:如何精简-Prometheus-的指标和存储占用

前言

随着 Prometheus 监控的组件、数量、指标越来越多，Prometheus 对计算性能的要求会越来越高，存储占用也会越来越多。

在这种状况下，要优化 Prometheus 性能, 优化存储占用. 第一工夫想到的可能是各种 Prometheus 的兼容存储计划, 如 Thanos 或 VM、Mimir 等。然而实际上尽管集中存储、长期存储、存储降采样及存储压缩能够肯定水平解决相干问题，然而治标不治本。

真正的本，还是在于指标量（series）过于宏大。
治标之法，应该是缩小指标量。有 2 种方法：
- Prometheus 性能调优 - 解决高基数问题
- 依据理论应用状况，只保留（keep）展现（Grafana Dashboards）和告警（prometheus rules）会用到的指标。

本次重点介绍第二种方法：如何依据理论的应用状况精简 Prometheus 的指标和存储占用？

思路

剖析以后 Prometheus 中存储的所有的 metric name（指标项）；
剖析展现环节用到的所有 metric name，即 Grafana 的 Dashboards 用到的所有指标；
剖析告警环节用到的所有 metric name，即 Prometheus Rule 配置中用到的所有指标；
（可选）剖析诊断环境用到的所有 metric name，即常常在 Prometheus UI 上 query 的指标；
通过 relabel 在 metric_relabel_configs 或 write_relabel_configs 仅 keep 2-4 中的指标, 以此大幅缩小 Prometheus 须要存储的指标量.

要具体实现这个思路, 能够通过 Grafana Labs 出品的 mimirtool 来搞定.

我这里有个前后的比照成果, 可供参考这样做成果有多惊人:

精简前: 270336 流动 series
精简后: 61055 流动 series
精简成果: 将近 5 倍的精简率!

Grafana Mimirtool

Grafana Mimir 是一款以对象存储为存储形式的 Prometheus 长期存储解决方案, 从 Cortex 演变而来. 官网号称反对亿级别的 series 写入存储和查问.

Grafana Mimirtool 是 Mimir 公布的一个实用工具, 可独自应用.

Grafana Mimirtool 反对从以下方面提取指标:

Grafana 实例中的Grafana Dashboards(通过 Grafana API)
Mimir 实例中的 Prometheus alerting 和 recording rules
Grafana Dashboards JSON文件
Prometheus记alerting 和 recording rules 的 YAML文件

而后，Grafana Mimirtool能够将这些提取的指标与Prometheus或Cloud Prometheus实例中的流动 series 进行比拟，并输入一个 used 指标和 unused 指标的列表。

Prometheus 精简指标实战

假如

假设:

通过kube-prometheus-stack 装置 Prometheus
已装置 Grafana 且作为展现端
已配置相应的告警规定
除此之外, 无其余须要额定保留的指标

前提

Grafana Mimirtool 从 releases 中找到 mimirtool 对应平台的版本下载即可应用;
已创立 Grafana API token
Prometheus已装置和配置.

第一步: 剖析 Grafana Dashboards 用到的指标

通过 Grafana API

具体如下:

# 通过 Grafana API剖析 Grafana 用到的指标# 前提是当初 Grafana上创立 API Keysmimirtool analyze grafana --address http://172.16.0.20:32651 --key=eyJrIjoiYjBWMGVoTHZTY3BnM3V5UzNVem9iWDBDSG5sdFRxRVoiLCJuIjoibWltaXJ0b29sIiwiaWQiOjF9

阐明:

http://172.16.0.20:32651 是 Grafana 地址
--key=eyJr 是 Grafana API Token. 通过如下界面取得:

获取到的是一个 metrics-in-grafana.json, 内容概述如下:

{    "metricsUsed": [        ":node_memory_MemAvailable_bytes:sum",        "alertmanager_alerts",        "alertmanager_alerts_invalid_total",        "alertmanager_alerts_received_total",        "alertmanager_notification_latency_seconds_bucket",        "alertmanager_notification_latency_seconds_count",        "alertmanager_notification_latency_seconds_sum",        "alertmanager_notifications_failed_total",        "alertmanager_notifications_total",        "cluster",        "cluster:namespace:pod_cpu:active:kube_pod_container_resource_limits",        "cluster:namespace:pod_cpu:active:kube_pod_container_resource_requests",        "cluster:namespace:pod_memory:active:kube_pod_container_resource_limits",        "cluster:namespace:pod_memory:active:kube_pod_container_resource_requests",        "cluster:node_cpu:ratio_rate5m",        "container_cpu_cfs_periods_total",        "container_cpu_cfs_throttled_periods_total",        "..."    ],    "dashboards": [        {            "slug": "",            "uid": "alertmanager-overview",            "title": "Alertmanager / Overview",            "metrics": [                "alertmanager_alerts",                "alertmanager_alerts_invalid_total",                "alertmanager_alerts_received_total",                "alertmanager_notification_latency_seconds_bucket",                "alertmanager_notification_latency_seconds_count",                "alertmanager_notification_latency_seconds_sum",                "alertmanager_notifications_failed_total",                "alertmanager_notifications_total"            ],            "parse_errors": null        },        {            "slug": "",            "uid": "c2f4e12cdf69feb95caa41a5a1b423d9",            "title": "etcd",            "metrics": [                "etcd_disk_backend_commit_duration_seconds_bucket",                "etcd_disk_wal_fsync_duration_seconds_bucket",                "etcd_mvcc_db_total_size_in_bytes",                "etcd_network_client_grpc_received_bytes_total",                "etcd_network_client_grpc_sent_bytes_total",                "etcd_network_peer_received_bytes_total",                "etcd_network_peer_sent_bytes_total",                "etcd_server_has_leader",                "etcd_server_leader_changes_seen_total",                "etcd_server_proposals_applied_total",                "etcd_server_proposals_committed_total",                "etcd_server_proposals_failed_total",                "etcd_server_proposals_pending",                "grpc_server_handled_total",                "grpc_server_started_total",                "process_resident_memory_bytes"            ],            "parse_errors": null        },        {...}    ]}

(可选)通过 Grafana Dashboards json 文件

如果无奈创立 Grafana API Token, 只有有 Grafana Dashboards json 文件, 也能够用来剖析, 示例如下:

# 通过 Grafana Dashboard json 剖析 Grafana 用到的指标mimirtool analyze dashboard grafana_dashboards/blackboxexporter-probe.jsonmimirtool analyze dashboard grafana_dashboards/es.json

失去的 json 构造和上一节相似, 就不赘述了.

第二步: 剖析 Prometheus Alerting 和 Recording Rules 用到的指标

具体操作如下:

# (可选)通过 kubectl cp 将用到的 rule files 拷贝到本地kubectl cp <prompod>:/etc/prometheus/rules/<releasename>-kube-prometheus-st-prometheus-rulefiles-0 -c prometheus ./kube-prometheus-stack/rulefiles/# 通过 Prometheus rule files 剖析 Prometheus Rule 用到的指标(波及 recording rule 和 alert rules)mimirtool analyze rule-file ./kube-prometheus-stack/rulefiles/*

后果如下 metrics-in-ruler.json:

{  "metricsUsed": [    "ALERTS",    "aggregator_unavailable_apiservice",    "aggregator_unavailable_apiservice_total",    "apiserver_client_certificate_expiration_seconds_bucket",    "apiserver_client_certificate_expiration_seconds_count",    "apiserver_request_terminations_total",    "apiserver_request_total",    "blackbox_exporter_config_last_reload_successful",    "..."  ],  "ruleGroups": [    {      "namspace": "default-monitor-kube-prometheus-st-kubernetes-apps-ae2b16e5-41d8-4069-9297-075c28c6969e",      "name": "kubernetes-apps",      "metrics": [        "kube_daemonset_status_current_number_scheduled",        "kube_daemonset_status_desired_number_scheduled",        "kube_daemonset_status_number_available",        "kube_daemonset_status_number_misscheduled",        "kube_daemonset_status_updated_number_scheduled",        "..."      ]      "parse_errors": null    },    {      "namspace": "default-monitor-kube-prometheus-st-kubernetes-resources-ccb4a7bc-f2a0-4fe4-87f7-0b000468f18f",      "name": "kubernetes-resources",      "metrics": [        "container_cpu_cfs_periods_total",        "container_cpu_cfs_throttled_periods_total",        "kube_node_status_allocatable",        "kube_resourcequota",        "namespace_cpu:kube_pod_container_resource_requests:sum",        "namespace_memory:kube_pod_container_resource_requests:sum"      ],      "parse_errors": null    },     {...}  ]}

第三步: 剖析没用到的指标

具体如下:

# 综合剖析 Prometheus 采集到的 VS. (展现(Grafana Dashboards) + 记录及告警(Rule files))mimirtool analyze prometheus --address=http://172.16.0.20:30090/ --grafana-metrics-file="metrics-in-grafana.json" --ruler-metrics-file="metrics-in-ruler.json"

阐明:

--address=http://172.16.0.20:30090/ 为 prometheus 地址
--grafana-metrics-file="metrics-in-grafana.json" 为第一步失去的 json 文件
--ruler-metrics-file="kube-prometheus-stack-metrics-in-ruler.json" 为第二步失去的 json 文件

输入后果prometheus-metrics.json 如下:

{  "total_active_series": 270336,  "in_use_active_series": 61055,  "additional_active_series": 209281,  "in_use_metric_counts": [    {      "metric": "rest_client_request_duration_seconds_bucket",      "count": 8855,      "job_counts": [        {          "job": "kubelet",          "count": 4840        },         {          "job": "kube-controller-manager",          "count": 1958        },        {...}      ]    },    {      "metric": "grpc_server_handled_total",      "count": 4394,      "job_counts": [        {          "job": "kube-etcd",          "count": 4386        },        {          "job": "default/kubernetes-ebao-ebaoops-pods",          "count": 8        }      ]    },    {...}  ],  "additional_metric_counts": [        {      "metric": "rest_client_rate_limiter_duration_seconds_bucket",      "count": 81917,      "job_counts": [        {          "job": "kubelet",          "count": 53966        },        {          "job": "kube-proxy",          "count": 23595        },        {          "job": "kube-scheduler",          "count": 2398        },        {          "job": "kube-controller-manager",          "count": 1958        }      ]    },      {      "metric": "rest_client_rate_limiter_duration_seconds_count",      "count": 7447,      "job_counts": [        {          "job": "kubelet",          "count": 4906        },        {          "job": "kube-proxy",          "count": 2145        },        {          "job": "kube-scheduler",          "count": 218        },        {          "job": "kube-controller-manager",          "count": 178        }      ]    },    {...}  ]}

第四步: 仅 `keep` 用到的指标

在 `write_relabel_configs` 环节配置

如果你有应用 remote_write, 那么间接在 write_relabel_configs 环节配置 keep relabel 规定, 简略粗犷.

能够先用 jp 命令失去所有须要 keep 的metric name:

jq '.metricsUsed' metrics-in-grafana.json \| tr -d '", ' \| sed '1d;$d' \| grep -v 'grafanacloud*' \| paste -s -d '|' -

输入后果相似如下:

instance:node_cpu_utilisation:rate1m|instance:node_load1_per_cpu:ratio|instance:node_memory_utilisation:ratio|instance:node_network_receive_bytes_excluding_lo:rate1m|instance:node_network_receive_drop_excluding_lo:rate1m|instance:node_network_transmit_bytes_excluding_lo:rate1m|instance:node_network_transmit_drop_excluding_lo:rate1m|instance:node_vmstat_pgmajfault:rate1m|instance_device:node_disk_io_time_seconds:rate1m|instance_device:node_disk_io_time_weighted_seconds:rate1m|node_cpu_seconds_total|node_disk_io_time_seconds_total|node_disk_read_bytes_total|node_disk_written_bytes_total|node_filesystem_avail_bytes|node_filesystem_size_bytes|node_load1|node_load15|node_load5|node_memory_Buffers_bytes|node_memory_Cached_bytes|node_memory_MemAvailable_bytes|node_memory_MemFree_bytes|node_memory_MemTotal_bytes|node_network_receive_bytes_total|node_network_transmit_bytes_total|node_uname_info|up

而后间接在 write_relabel_configs 环节配置 keep relabel 规定:

remote_write:- url: <remote_write endpoint>  basic_auth:    username: <按需>    password: <按需>  write_relabel_configs:  - source_labels: [__name__]    regex: instance:node_cpu_utilisation:rate1m|instance:node_load1_per_cpu:ratio|instance:node_memory_utilisation:ratio|instance:node_network_receive_bytes_excluding_lo:rate1m|instance:node_network_receive_drop_excluding_lo:rate1m|instance:node_network_transmit_bytes_excluding_lo:rate1m|instance:node_network_transmit_drop_excluding_lo:rate1m|instance:node_vmstat_pgmajfault:rate1m|instance_device:node_disk_io_time_seconds:rate1m|instance_device:node_disk_io_time_weighted_seconds:rate1m|node_cpu_seconds_total|node_disk_io_time_seconds_total|node_disk_read_bytes_total|node_disk_written_bytes_total|node_filesystem_avail_bytes|node_filesystem_size_bytes|node_load1|node_load15|node_load5|node_memory_Buffers_bytes|node_memory_Cached_bytes|node_memory_MemAvailable_bytes|node_memory_MemFree_bytes|node_memory_MemTotal_bytes|node_network_receive_bytes_total|node_network_transmit_bytes_total|node_uname_info|up    action: keep

在 `metric_relabel_configs` 环节配置

如果没有应用 remote_write, 那么只能在 metric_relabel_configs 环节配置了.

以 etcd job 为例: (以 prometheus 配置为例, Prometheus Operator 请自行按需调整)

- job_name: serviceMonitor/default/monitor-kube-prometheus-st-kube-etcd/0  honor_labels: false  kubernetes_sd_configs:  - role: endpoints    namespaces:      names:      - kube-system  scheme: https  tls_config:    insecure_skip_verify: true    ca_file: /etc/prometheus/secrets/etcd-certs/ca.crt    cert_file: /etc/prometheus/secrets/etcd-certs/healthcheck-client.crt    key_file: /etc/prometheus/secrets/etcd-certs/healthcheck-client.key  relabel_configs:  - source_labels:    - job    target_label: __tmp_prometheus_job_name  - ...  metric_relabel_configs:   - source_labels: [__name__]    regex: etcd_disk_backend_commit_duration_seconds_bucket|etcd_disk_wal_fsync_duration_seconds_bucket|etcd_mvcc_db_total_size_in_bytes|etcd_network_client_grpc_received_bytes_total|etcd_network_client_grpc_sent_bytes_total|etcd_network_peer_received_bytes_total|etcd_network_peer_sent_bytes_total|etcd_server_has_leader|etcd_server_leader_changes_seen_total|etcd_server_proposals_applied_total|etcd_server_proposals_committed_total|etcd_server_proposals_failed_total|etcd_server_proposals_pending|grpc_server_handled_total|grpc_server_started_total|process_resident_memory_bytes|etcd_http_failed_total|etcd_http_received_total|etcd_http_successful_duration_seconds_bucket|etcd_network_peer_round_trip_time_seconds_bucket|grpc_server_handling_seconds_bucket|up    action: keep

不必 `keep` 而应用 `drop`

同样滴, 不必 keep 而改为应用 drop 也是能够的. 这里不再赘述.

总结

本文中，介绍了精简 Prometheus 指标的需要, 而后阐明如何应用 mimirtool analyze 命令来确定Grafana Dashboards 以及 Prometheus Rules 中用到的指标。而后用 analyze prometheus 剖析了展现和告警中used 和 unused 的流动 series，最初配置了 Prometheus 以仅 keep 用到的指标。

联合这次实战, 精简率能够达到 5 倍左右, 成果还是非常明显的. 举荐试一试. ️️️

️ 参考文档

grafana/mimir: Grafana Mimir provides horizontally scalable, highly available, multi-tenant, long-term storage for Prometheus. (github.com)
Analyzing and reducing metrics usage with Grafana Mimirtool | Grafana Cloud documentation

三人行, 必有我师; 常识共享, 天下为公. 本文由东风微鸣技术博客 EWhisper.cn 编写.

前言

思路