This document captures the intended design for KusionStack's AppConfiguration model to support Prometheus-related configuration when describing an application.
Section 1-4 is intended for KusionStack users and model developers who aren't familiar with Prometheus. If you are familiar with Prometheus, you can skip the first 4 sections and head directly to Design.
AppConfiguration model provides an interface to describe all attributes that are tied to an application.
More details on AppConfiguration model can be found in the AppConfiguration design doc.
Prometheus is an open-source systems monitoring and alerting toolkit originally built at SoundCloud and has now become the de facto standard for cloud-native monitoring solutions. Prometheus collects and stores its metrics as time series data, i.e. metrics information is stored with the timestamp at which it was recorded, alongside optional key-value pairs called labels[1].
For kusion related concepts such as frontend models, backend models and backbone models, please see here.
Prometheus server: The actual server that runs the Prometheus binary.
Prometheus scrape_config
: The configuration that Prometheus honors when scraping metrics. This is defined in the Prometheus configuration (usually named Prometheus.yml
).The full documentation can be found here.
Static scraping: As part of the scraping configuration, users can define the list of static endpoints Prometheus scrapes from in the static_configs
section in the scrape_config
. More can be find here.
static_configs:
- targets: ['localhost:9090']
Service discovery: As opposed to static scraping, Prometheus also supports automated service discovery mechanisms that can be defined in the xxx_sd_configs
section in the scrape_config
, where xxx
can be a list of options found here. The Kubernetes-related scraping configurations are defined under kubernetes_sd_config
, which is the most relevant configuration being discussed in this documentation.
Prometheus Operator: The Prometheus Operator provides Kubernetes native deployment and management of Prometheus and related monitoring components[2]. It leverages the operator pattern to provide a way to manage Prometheus installation and configurations via CRDs.
KusionStack is application-centric by design and aims to provide a homogeneous experience in application delivery for all kinds of applications.
As of today, the design in this document only captures the scraping configurations Prometheus for a given application that fits into the AppConfiguration model. For example, the steps to install and configure Prometheus itself is not within consideration of this document.
Prometheus can be installed in a Kubernetes cluster in a variety of ways. On a high level, there are a few patterns to follow:
- Installing the Prometheus server (along with its necessary Kubernetes resources such as
ServiceAccount
,ClusterRole
,etc) and managing the configuration directly. - Installing Prometheus binary and run it in the agent mode[3] while connecting to a remote Prometheus server for aggregation. A number of cloud vendors provide monitoring solutions that resemble this approach.
- Using the Prometheus operator and managing the Prometheus deployment and monitoring configuration via Kubernetes CRs.
From kusion's perspective, these can be categorized into Prometheus operator vs. non-operator. Both case should be supported, but certain constraints may exist.
In the operator installation of Prometheus, the application scraping configuration in Kubernetes is managed via CRs, specifically:
ServiceMonitor
in monitoring.coreos.com/v1, for scraping servicesPodMonitor
in monitoring.coreos.com/v1, for scraping workloadsProbe
in monitoring.coreos.com/v1, for scraping static targets and ingresses
These 3 CRDs above will be supported by kusion when describing an application.
The Prometheus operator acts on a few other CRDs that define the installation and configuration of Prometheus components, some of which are optional:
Prometheus
in monitoring.coreos.com/v1, defines a Prometheus deployment.PrometheusAgent
in monitoring.coreos.com/v1alpha1, defines a Prometheus agent deployment.ScrapeConfig
in monitoring.coreos.com/v1alpha1, for namespaced scraping configurationPrometheusRule
in monitoring.coreos.com/v1, defines the recording and alerting rules for a Prometheus instance.AlertManager
in monitoring.coreos.com/v1, defines an AlertManager cluster.ThanosRuler
in monitoring.coreos.com/v1, defines a ThanosRuler deployment.
These 6 CRDs above are not within the scope of this documentation.
In the non-operator installation of Prometheus, the scraping configuration in Kubernetes is directly managed in kubernetes_sd_configs
.
If you are managing Prometheus this way, please make sure you have the ability to directly modify this configuration. It exists in the Prometheus configuration, in the scrape_config
section, and depending on your actual Prometheus setup, the values might be set in the Prometheus.yml
file, a ConfigMap, or command-line arguments. More information can be found in the Prometheus configuration documentation.
The kubernetes_sd_config
allows customized filtering based on different criteria. Specifically, it allows Prometheus to retrieve the scraping targets(services, endpoints or workloads) by the presence of a set of annotations on the said Kubernetes resources.
These annotations can technically be anything as long as the service discovery configuration in scrape_config
section in the Prometheus configuration reflects the same. The most widely-used annotations are:
annotations:
prometheus.io/scrape: "true"
prometheus.io.scheme: "https"
prometheus.io/path: "/metrics"
prometheus.io/port: "9191"
The corresponding kubernetes_sd_configs
looks like (in relabel_configs
section):
...
scrape_configs:
- job_name: 'kubernetes-service-endpoints'
scrape_interval: 1s
scrape_timeout: 1s
kubernetes_sd_configs:
- role: endpoints
relabel_configs:
- source_labels: [__meta_kubernetes_service_annotation_prometheus_io_scrape]
action: keep
regex: true
- source_labels: [__meta_kubernetes_service_annotation_prometheus_io_scheme]
action: replace
target_label: __scheme__
regex: (https?)
- source_labels: [__meta_kubernetes_service_annotation_prometheus_io_path]
action: replace
target_label: __metrics_path__
regex: (.+)
- source_labels: [__address__, __meta_kubernetes_service_annotation_prometheus_io_port]
action: replace
target_label: __address__
regex: ([^:]+)(?::\d+)?;(\d+)
replacement: $1:$2
...
The source_labels
fields determine the actual key of the annotation Prometheus uses when assembling the scraping target via relabeling. __scheme__
, __metrics_path__
, __address__
together points to the actual URL where Prometheus scrapes from:
__scheme__://__address____metrics_path__
However, annotation-based approach has certain limitations. A few examples of things that are impossible with this approach[4]:
- any targets with multiple ports
- any targets that don't all have the exact same authentication, http and TLS configuration
- selection is an all or nothing, not based on the well known label selection paradigm in Kubernetes
The advise from the Prometheus team is to use the ServiceMonitor
or PodMonitor
CRs via the Prometheus operator to manage scrape configs going forward[5].
While acknowledging the Prometheus operator is the longer term approach, kusion should still be able to support setting an application level Prometheus scraping config via the AppConfiguration model, by making the assumption that the Prometheus installation uses the community most-adopted annotations, namely these four:
annotations:
prometheus.io/scrape: "true"
prometheus.io.scheme: "https"
prometheus.io/path: "/metrics"
prometheus.io/port: "9191"
The AppConfiguration model should automatically add the above annotations to the workloads/services/ingresses deployed via kusion when the Prometheus related configuration is present (with a set of default values such as http
, /metrics
and 8080
).
If the Prometheus installation is configured to recognize any other annotations on the Kubernetes objects, you can still customize the scraping config by adding those customized annotations on your workloads or services. For example:
annotations:
customized.prometheus/scrape: "true"
customized.prometheus/scheme: "https"
customized.prometheus/path: "/metrics"
customized.prometheus/port: "9191"
- Prometheus is installed and properly configured with prometheus.yml in the cluster. This includes either the server, agent OR the operator installation.
- Users are able to declare Prometheus-related monitoring attributes in AppConfiguration frontend model
- The backend model is able to convert the user-defined attributes to relevant configurations while rendering the Kubernetes manifests
If the Prometheus operator is used to manage the monitoring configuration in the cluster (this should be indicated as a flag in the frontend model), kusion should be able to render and apply the ServiceMonitor
, PodMonitor
and Probe
CRDs.
If non-operator Prometheus (server/agent) is used, kusion should automatically inject the community-standard prometheus.io/xxxx
annotations into the workload, making the assumption that the Prometheus server/agent is configured to look for that.
If Istio is also installed in the cluster and metrics merging is enabled, it should look for the community-standard annotations on the workloads and expose the merged metrics at /stats/prometheus
on Istio-standard port 15020
. This is Istio default behavior and does not require user participation besides making sure the annotations are in place.
The frontend models should expose monitoring related attributes. Below is a sample design for Prometheus.Scraping
(not finalized):
scheme
- (string, optional) The protocol scheme used for scraping. Possible values arehttp
,https
. Default tohttp
.path
- (string, optional) The path to scrape metrics from. Default to/metrics
.port
- (k8s.io/apimachinery/pkg/util/intstr.IntOrString, optional) The port to scrape metrics from. This can be a named port or a port number. Default to8080
.interval
- (protobuf.Duration, optional) The interval to scrape metrics. Only applicable ifoperatorMode
is on. Default to15s
.timeout
- (protobuf.Duration, optional) The time until a scrape request times out. Only applicable ifoperatorMode
is on. Default to5s
.operatorMode
- (bool, optional) Whether or not to apply scrape configs using the Prometheus operator. Requires the Prometheus operator to be present in the cluster. Default tofalse
.monitorType
- (string, optional) The type of resources to scrape from. Possible values areservice
,pod
,ingress
. Default topod
.
The backend models should provide two paths for rendering the Prometheus-related resources, based on the input of the operatorMode
flag:
- If
operatorMode
is set totrue
, the backend model should produce one or more rendered CR(s), namelyServiceMonitor
,PodMonitor
orProbe
. - Alternative, if
operatorMode
is set tofalse
, the backend model should automatically inject a set of annotations into the application pods:
annotations:
prometheus.io/scrape: "xxxx"
prometheus.io.scheme: "xxxx"
prometheus.io/path: "/xxxx"
prometheus.io/port: "xxxx"
To support the Prometheus-related CRDs watched by the Prometheus operator, the backbone models need to be generated from Prometheus-operator CRDs using the kcl-openapi
tool. More info can be found here.
By default, Prometheus scrapes metrics over HTTP in cleartext. If scraping over TLS is required, or if the metrics endpoint is protected behind authentication, the TLS settings and/or authentication details must be present in the scraping configuration (or in the case of Prometheus operator, in the ServiceMonitor
or PodMonitor
definition). Please see the tlsSettings
, basic_auth
, oauth2
and authorization
in scraping config and relevant CRD API Reference for more details.
When Istio is installed in the cluster that Prometheus is scraping metrics from, there could be some complications.
By default, Istio-proxy configures itself in the permissive
mode where while it expects mTLS traffic coming into the application pod, it also accepts clear-text traffic such as Prometheus scraping requests.
However, if Istio-proxy is configured to use the strict
mTLS mode, the proxy stops accepting non-mTLS traffic, causing issues for Prometheus scraping.
To solve the problem, Istio have since added the capability known as metrics merging. Metrics merging leverages the istio-agent process to scrape both the application metrics and Istio proxy metrics, and combine them together in one place for Prometheus to scrape from without needing mTLS.
The Istio agent process takes note of the aforementioned prometheus.io/xxxx
annotations and passes them as input to the Istio agent process while invoking the Istio injection mutating webhook so that Istio agent knows where to find the application metrics to combine with.
If you have the Istio mTLS mode set to strict
, the prometheus.io/xxxx
annotations must be present to allow Istio to locate application metrics.
Istio metrics merging is by default on since Istio 1.7.0. More on metrics merging can be found here.
Metrics merging have certain limitations, it does not serve the following needs[6]:
- The scraping traffic needs to be TLS-encrypted (Metrics merging serves merged metrics over cleartext scrape by Istio-agent)
- The application exposes metrics with the same name as Istio metrics
- The Prometheus server is NOT configured to look for the community-standard
prometheus.io/xxxx
annotations
If strict mTLS scraping is needed (meaning the scraping traffic needs to be encrypted over mTLS), custom TLS settings need to be present in the Prometheus scraping configuration. More info can be found here.
monitoring: Prometheus.Scraping{
interval = 5s
timeout = 3s
path = "/actuator/metrics"
port = 9080
scheme = "http"
operatorMode = true
monitorType = "pod"
}
The above example should be rendered into the following PodMonitor
CR:
apiVersion: monitoring.coreos.com/v1
kind: PodMonitor
metadata:
name: sample-app-pod-monitor
namespace: sample-app-ns
spec:
selector:
matchLabels:
app: sample-app
podMetricsEndpoints:
- port: 9080
path: "/actuator/metrics"
scheme: "http"
interval: 5s
scrapeTimeout: 3s
Alternatively, if operatorMode
is set to false
(or omitted), kusion will automatically inject the following annotations into the application workload:
apiVersion: apps/v1
kind: Deployment
metadata:
...
spec:
...
template:
metadata:
annotations:
prometheus.io/scrape: "true"
prometheus.io.scheme: "http"
prometheus.io/path: "/actuator/metrics"
prometheus.io/port: "9080"
...
The above annotations should be recognized by Prometheus provided that Prometheus configs have the __meta_kubernetes_service_annotation_prometheus_io_xxxx
meta labels defined in the relabel_configs
, which should be on by default if you are using the extremely popular Prometheus community helm chart while setting up Prometheus.
- Prometheus: https://prometheus.io/docs/introduction/overview/
- Prometheus Operator: https://github.com/prometheus-operator/prometheus-operator
- Prometheus Agent mode: https://prometheus.io/blog/2021/11/16/agent/
- Annotation-based approach limitations: prometheus-operator/prometheus-operator#1547 (comment)
- Prometheus team advise: prometheus-operator/prometheus-operator#1547 (comment)
- Istio integration with Prometheus: https://istio.io/latest/docs/ops/integrations/prometheus/