Meet the easiest way to generate SLOs for Prometheus.
Sloth generates understandable, uniform and reliable Prometheus SLOs for any kind of service. Using a simple SLO spec that results in multiple metrics and multi window multi burn alerts.
At this moment Sloth is focused on Prometheus, however depending on the demand and complexity we may support other backends.
- Simple, maintainable and understandable SLO spec.
- Reliable SLO metrics and alerts.
- Based on Google SLO implementation and multi window multi burn alerts framework.
- Autogenerates Prometheus SLI recording rules in different time windows.
- Autogenerates Prometheus SLO metadata rules.
- Autogenerates Prometheus SLO multi window multi burn alert rules (Page and warning).
- SLO spec validation (including
validate
command for Gitops and CI). - Customization of labels, disabling different type of alerts...
- A single way (uniform) of creating SLOs across all different services and teams.
- Automatic Grafana dashboard to see all your SLOs state.
- Single binary and easy to use CLI.
- Kubernetes (Prometheus-operator) support.
- Kubernetes Controller/operator mode with CRDs.
- Support different SLI types.
- Support for SLI plugins
- A library with common SLI plugins.
- OpenSLO support.
- Releases
- Multi-arch Docker images
git clone [email protected]:slok/sloth.git && cd ./sloth && make build && ls -la ./bin
Release the Sloth!
sloth generate -i ./examples/getting-started.yml
version: "prometheus/v1"
service: "myservice"
labels:
owner: "myteam"
repo: "myorg/myservice"
tier: "2"
slos:
# We allow failing (5xx and 429) 1 request every 1000 requests (99.9%).
- name: "requests-availability"
objective: 99.9
description: "Common SLO based on availability for HTTP request responses."
labels:
category: availability
sli:
events:
error_query: sum(rate(http_request_duration_seconds_count{job="myservice",code=~"(5..|429)"}[{{.window}}]))
total_query: sum(rate(http_request_duration_seconds_count{job="myservice"}[{{.window}}]))
alerting:
name: "MyServiceHighErrorRate"
labels:
category: "availability"
annotations:
# Overwrite default Sloth SLO alert summmary on ticket and page alerts.
summary: "High error rate on 'myservice' requests responses"
page_alert:
labels:
severity: "pageteam"
routing_key: "myteam"
ticket_alert:
labels:
severity: "slack"
slack_channel: "#alerts-myteam"
This would be the result you would obtain from the above spec example.
At this moment Sloth uses Prometheus rules to generate SLOs. Based on the generated recording and alert rules it creates a reliable and uniform SLO implementation:
1 Sloth spec -> Sloth -> N Prometheus rules
The Prometheus rules that Sloth generates can be explained in 3 categories:
- SLIs: These rules are the base, they use the queries provided by the user to get a value used to show what is the error service level (availability). It creates multiple rules for different time windows, these different results will be used for the alerts.
- Metadata: These are used as informative metrics, like the remaining error budget, the SLO objective percent... These are very handy for SLO visualization, e.g Grafana dashboard.
- Alerts: These are the multiwindow-multiburn alerts that are based on the SLI rules.
Sloth will take the service level spec and for each SLO in the spec will create 3 rule groups with the above categories.
These rules will share the same metric name across SLOs. However the labels are the key to identify the different services, SLO... This is how we obtain a uniform way of describing all the SLOs across different teams and services.
To get all the available metric names created by Sloth, use this query:
count({sloth_id!=""}) by (__name__)
At this moment Sloth can work in two different modes:
- As a CLI application: Load Sloth manifests, return generated SLOs.
- As a Kubernetes controller: React on CRD events, generate SLOs and store them in Kubernetes.
generate
will generate Prometheus rules in different formats based on the specs. This mode only needs the CLI so its very useful for GitOps, CI, scripts or as a CLI on your toolbox.
Currently there are two types of specs supported for generate
command. Sloth will detect the input spec type and generate the output type accordingly:
Check spec here: v1
Will generate the prometheus recording and alerting rules in Standard Prometheus YAML format.
Example:
$ sloth generate -i ./examples/home-wifi.yml -o /tmp/home-wifi.yml
INFO[0000] Generating from Prometheus spec version=v0.1.0-43-g5715af5
INFO[0000] Multiwindow-multiburn alerts generated slo=home-wifi-good-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] SLI recording rules generated rules=8 slo=home-wifi-good-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] Metadata recording rules generated rules=7 slo=home-wifi-good-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] SLO alert rules generated rules=2 slo=home-wifi-good-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] Multiwindow-multiburn alerts generated slo=home-wifi-risk-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] SLI recording rules generated rules=8 slo=home-wifi-risk-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] Metadata recording rules generated rules=7 slo=home-wifi-risk-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] SLO alert rules generated rules=2 slo=home-wifi-risk-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] Prometheus rules written format=yaml groups=6 out=/tmp/home-wifi.yml svc=storage.IOWriter version=v0.1.0-43-g5715af5
Kubernetes CRD (Prometheus-operator)
Check CRD here: v1
Will generate from a Sloth CRD spec into [Prometheus-operator]crd rules. This generates the prometheus operator CRDs based on the Sloth CRD template.
- Sloth CRD is NOT required in the cluster because it happens offline as a CLI. For controller/operator K8s flow, check next section
- Prometheus operator Rule CRD is required in the cluster for the Sloth generated rules: Manifest
Example:
$ sloth generate -i ./examples/k8s-home-wifi.yml -o /tmp/k8s-home-wifi.yml
INFO[0000] Generating from Kubernetes Prometheus spec version=v0.1.0-43-g5715af5
INFO[0000] Multiwindow-multiburn alerts generated slo=home-wifi-good-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] SLI recording rules generated rules=8 slo=home-wifi-good-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] Metadata recording rules generated rules=7 slo=home-wifi-good-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] SLO alert rules generated rules=2 slo=home-wifi-good-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] Multiwindow-multiburn alerts generated slo=home-wifi-risk-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] SLI recording rules generated rules=8 slo=home-wifi-risk-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] Metadata recording rules generated rules=7 slo=home-wifi-risk-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
INFO[0000] SLO alert rules generated rules=2 slo=home-wifi-risk-wifi-client-satisfaction svc=generate.prometheus.Service version=v0.1.0-43-g5715af5
Sloth supports OpenSLO v1alpha spec, however we need to take into account that it has some restrictions:
- OpenSLO timewindow restricted to 30 days.
- Only Objective ratio metrics are supported.
- Only Prometheus and PromQL query types are supported.
- Configuration fields not required by Sloth will be ignored.
Regarding Sloth fatures, OpenSLO spec doesn't support all of the sloth features:
- No Prometheus labels support.
- No alerting support.
- No SLI plugins support.
- No Kubernetes support (at least until official OpenSLO CRDs are released).
Check examples to see some OpenSLO example specs.
Kubernetes Controller (Prometheus-operator)
kubernetes-controller
command runs Sloth as a controller/operator that will react on sloth.slok.dev/v1/PrometheusServiceLevel
CRD. The controller will create the required [Prometheus-operator]crd rules.
In the end the controller mode makes the same work as the CLI however integrates better with a native Kubernetes flow.
- Sloth CRD is required in the cluster: Manifest
- Prometheus operator Rule CRD is required in the cluster: Manifest
# Register Sloth CRD.
$ kubectl apply -f ./pkg/kubernetes/gen/crd/sloth.slok.dev_prometheusservicelevels.yaml
# Prometheus Operator Rules CRD is required.
# kubectl apply -f ./test/integration/crd/prometheus-operator-crd.yaml
# Run sloth controller.
$ kubectl create ns monitoring
$ kubectl apply -f ./deploy/kubernetes/raw/sloth.yaml
# Deploy some SLOs.
$ kubectl apply -f ./examples/k8s-getting-started.yml
# Get CRs.
$ kubectl -n monitoring get slos
NAME SERVICE DESIRED SLOS READY SLOS GEN OK GEN AGE AGE
sloth-slo-my-service myservice 1 1 true 27s 27s
$ kubectl -n monitoring get prometheusrules
NAME AGE
sloth-slo-my-service 38s
Sloth validates the spec on generation, however, on specific steps of the SLO generation process, we only want to validate a group of SLOs. For this purpose Sloth comes with a helpful command called validate
. It will discover all the specs recursively and apply the same generation process as generate
(including plugins, options...) but discarding the result.
Example that validates all SLOs in a directory (including subdirectories) and excludes all in spec files that match _gen
in the spec path.
$ sloth validate --input ./examples --sli-plugins-path ./examples/plugins --fs-exclude _gen
INFO[0000] SLI plugins loaded plugins=1 version=dev
INFO[0000] Validation succeeded slo-specs=13 version=dev
This command is very helpful on Gitops and CI pipelines to have a fast feedback loop, independently of the process you are using for generating the SLOs (Kubernetes controller or CLI).
- Getting started: Getting started example.
- K8s Getting started: Getting started example as a Kubernetes CRD.
- Alerts disabled: Simple example that shows how to disable alerts.
- K8s apiserver: Real example of SLOs for a Kubernetes Apiserver.
- Home wifi: My home Ubiquti Wifi SLOs.
- K8s Home wifi: Same as home-wifi but shows how to generate Prometheus-operator CRD from a Sloth CRD.
- Raw Home wifi: Example showing how to use
raw
SLIs instead of the commonevents
using the home-wifi example. - SLI Plugins: Example showing how to use SLI plugins.
- OpenSLO: Example showing OpenSLO spec.
The resulting generated SLOs are in examples/_gen.
SLI plugins are small Go plugins (using Yaegi) that can be loaded on Sloth start.
These plugins can be referenced as an SLI on the SLO specs and will return a raw SLI type.
Sloth will maintain a library with common SLI plugins that can be used on your SLOs or used as examples to develop your own ones.
Developing a prometheus/v1
SLI plugin is very easy, however you need to meet some requirements:
- The plugin version used as a global called
SLIPluginVersion
. - A plugin ID global called
SLIPluginID
. - A Plugin logic function called
SLIPlugin
. - The plugin must be in a single file named
plugin.go
. - Plugins only can use the Go standard library (
reflect
andunsafe
packages can't be used). - Plugin received options are a
map[string]string
to avoidinterface{}
problems on dynamic execution code, the conversion to specific types are responsibility of the plugin. - Plugins don't depend on go modules, GOPATH or similar (thanks to the previous requirements).
Sloth knows how to autodiscover plugins giving a path (--sli-plugins-path
), and will load all the discovered ones.
A very simple example:
from plugins/x/y/plugin.go
package testplugin
import "context"
const (
SLIPluginVersion = "prometheus/v1"
SLIPluginID = "test_plugin"
)
func SLIPlugin(ctx context.Context, meta, labels, options map[string]string) (string, error) {
return "rate(my_raw_error_ratio_query{}[{{.window}}])", nil
}
Used in SLO spec:
version: "prometheus/v1"
service: "myservice"
slos:
- name: "some-slo"
objective: 99.9
sli:
plugin:
id: "test_plugin"
options:
opt1: "something"
opt2: "something"
alerting:
#...
On spec load, Sloth will execute the referenced plugins with the options and use the result as a Raw SLI type, the one that returns the error ratio query.
Why should I use plugins?
By default you shouldn't unless you have scenarios where they can simplify, add security or improve the SLO adoption on the team/company. Some examples:
- SLI custom validation (parameters, queries...).
- Company wide precreated SLIs for common used libraries.
- Complex Prometheus query SLIs.
- Precreated SLIs for the team or company that normally everyones uses on the SLOs.
- OSS SLI plugins that come with the apps, frameworks or libraries (e.g Kubernetes apiserver, http framework X...).
- The company or the team could have a repository with all the shared plugins and everyone could use them and contribute with new ones.
- Automation power when templates are not enough.
Check CONTRIBUTING.md.
- Why Sloth
- SLI?
- SLO?
- Error budget?
- Burn rate?
- SLO based alerting?
- What are ticket and page alerts?
- Can I disable alerts?
- Grafana dashboard?
- CLI VS K8s controller?
- SLI types on manifests
Creating Prometheus rules for SLI/SLO framework is hard, error prone and is pure toil.
Sloth abstracts this task, and we also gain:
- Read friendlyness: Easy to read and declare SLI/SLOs.
- Gitops: Easy to integrate with CI flows like validation, checks...
- Reliability and testing: Generated prometheus rules are already known that work, no need the creation of tests.
- Centralize features and error fixes: An update in Sloth would be applied to all the SLOs managed/generated with it.
- Standardize the metrics: Same conventions, automatic dashboards...
- Rollout future features for free with the same specs: e.g automatic report creation.
Service level indicator. Is a way of quantify how your service should be responding to user.
TL;DR: What is good/bad service for your users. E.g:
- Requests >=500 considered errors.
- Requests >200ms considered errors.
- Process executions with exit code >0 considered errors.
Normally is measured using events: good/bad-events / total-events
.
Service level objective. A percent that will tell how many SLI errors your service can have in a specific period of time.
An error budget is the ammount of errors (driven by the SLI) you can have in a specific period of time, this is driven by the SLO.
Lets see an example:
- SLI Error: Requests status code >= 500
- Period: 30 days
- SLO: 99.9%
- Error budget: 0.0999 (100-99.9)
- Total requests in 30 days: 10000
- Available error requests: 9.99 (10000 * 0.0999 / 100)
If we have more than 9.99 request response with >=500 status code, we would be burning more error budget than the available, if we have less errors, we would end without spending all the error budget.
The speed you are consuming your error budget. This is key for SLO based alerting (Sloth will create all these alerts), because depending on the speed you are consuming your error budget, it will trigger your alerts.
Speed/rate examples:
- 1: You are consuming 100% of the error budget in the expected period (e.g if 30d period, then 30 days).
- 2: You are consuming 200% of the error budget in the expected period (e.g if 30d period, then 15 days).
- 60: You are consuming 6000% of the error budget in the expected period (e.g if 30d period, then 12h hour).
- 1080: You are consuming 108000% of the error budget in the expected period (e.g if 30d period, then 40 minute).
With SLO based alerting you will get better alerting to a regular alerting system, because:
- Alerts on symptoms (SLIs), not causes.
- Trigger at different levels (warning/ticket and critical/page).
- Takes into account time and quantity, this is: speed of errors and number of errors on specific time.
The result of these is:
- Correct time to trigger alerts (important == fast, not so important == slow).
- Reduce alert fatigue.
- Reduce false positives and negatives.
MWMB type alerting is based on two kinds of alerts, ticket
and page
:
page
: Are critical alerts that normally are used to wake up, notify on important channels, trigger oncall...ticket
: The warning alerts that normally open tickets, post messages on non-important Slack channels...
These are triggered in different ways, page
alerts are triggered faster but require faster error budget burn rate, on the other side, ticket
alerts
are triggered slower and require a lower and constant error budget burn rate.
Yes, use disable: true
on page
and ticket
.
Check grafana-dashboard, this dashboard will load the SLOs automatically.
If you don't have Kubernetes and you need raw prometheus rules, its easy, the CLI (generate
) mode is the only one that supports raw prometheus rules.
On the other side if you have Kubernetes (and most likely prometheus-operator). Using sloth.slok.dev/v1/PrometheusServiceLevel
CRD will output the same result used as a CLI or used as a Kubernetes controller.
The only difference between bot modes is how Sloth application loads the SLOs manifest. On both modes the output will be a Prometheus Operator Rules CRD.
Both have pros and cons:
- The CLI in an advanced gitops flow gives you faster feedback loops because of the generation on the CI.
- Using as a controller the CRD integrates better in helm charts and similar because it removes that generation extra step.
- Having the SLO as CRs in K8s, improves the discovery as you can always do
kubectl get slos --all-namespaces
. - The CLI doesn't require an app running, Sloth CRDs registered... the SLO generation process is simpler, so you have less PoFs.
In a few words, theres no right or wrong answer, pick your own flavour based on your use case: teams size, engineers in the company or development flow...
prometheus/v1
(regular) and sloth.slok.dev/v1/PrometheusServiceLevel
(Kubernetes CRD), support 3 ways of setting SLIs:
- Events: This are based on 2 queries, the one that returns the total/valid number of events and the one that returns the bad events. Sloht will make a query dividing them to get the final error ratio (0-1).
- Raw: This is a single raw prometheus query that when executed will return the error ratio (0-1).
- Plugins: Check plugins section for more information. It reference plugins that will be preloaded and already developed. Sloth will execute them on generation and it will return a raw query. This is the best way to abstract queries from users or having SLOs at scale.