/slo-generator

SLO Generator is a tool to compute SLIs, SLOs, Error Budgets and Burn rate and export an SLO report to supported exporters.

Primary LanguagePythonApache License 2.0Apache-2.0

SLO Generator

build cloudbuild

slo-generator is a tool to compute and export Service Level Objectives (SLOs), Error Budgets and Burn Rates, using policies written in JSON or YAML format.

Description

slo-generator will query metrics backend and compute the following metrics:

  • Service Level Objective defined as SLO (%) = GOOD_EVENTS / VALID_EVENTS
  • Error Budget defined as ERROR_BUDGET = 100 - SLO (%)
  • Burn Rate defined as BURN_RATE = ERROR_BUDGET / ERROR_BUDGET_TARGET

Local usage

Requirements

  • Python 3
  • gcloud SDK installed

Installation

slo-generator is published on PyPI. To install it, run:

pip3 install slo-generator

Run the slo-generator

slo-generator -f <SLO_CONFIG_PATH> -b <ERROR_BUDGET_POLICY> --export

  • <SLO_CONFIG_PATH> is the SLO config file or folder. If a folder path is passed, the SLO configs filenames should match the pattern slo_*.yaml to be loaded.

  • <ERROR_BUDGET_POLICY> is the Error Budget Policy file.

  • --export enables exporting data using the exporters defined in the SLO configuration file.

Use slo-generator --help to list all available arguments.

To enable debug logs, set the environment variable DEBUG to 1 before running the slo-generator:

export DEBUG=1

Configuration

The slo-generator requires two configuration files to run, the SLO configuration file and the Error budget policy file.

SLO configuration

The SLO configuration (JSON or YAML) is composed of the following fields:

  • SLO metadata:

    • slo_name: Name of this SLO.
    • slo_description: Description of this SLO.
    • slo_target: SLO target (between 0 and 1).
    • service_name: Name of the monitored service.
    • feature_name: Name of the monitored subsystem.

  • SLI configuration:

  • Exporter configuration:
    • exporters: A list of exporters to export results to. Specific documentation is available for each supported exporters:
      • Cloud Pub/Sub to stream SLO reports.
      • BigQuery to export SLO reports to BigQuery for historical analysis and DataStudio reporting.
      • Stackdriver Monitoring to export the error_budget_burn_rate metric to Stackdriver Monitoring.
      • Prometheus to export the error_budget_burn_rate metric to Prometheus.

Note: you can use environment variables in your SLO configs by using ${} syntax to avoid having sensitive data in version control. Environment variables will be replaced at run time.

==> An example SLO configuration file is available here.

Error Budget policy

The Error Budget policy (JSON or YAML) is a list of multiple error budgets, each one composed of the following fields:

  • window: Rolling time window for this error budget.
  • alerting_burn_rate_threshold: Target burnrate threshold over which alerting is needed.
  • urgent_notification: boolean whether violating this error budget should trigger a page.
  • overburned_consequence_message: message to show when the error budget is above the target.
  • achieved_consequence_message: message to show when the error budget is within the target.

==> An example Error Budget policy is available here.

Reporting on SLO achievements

The PUBLIC SLO achievements report template can be cloned following instructions on the report page named README. It provide a basic dashboard with:

  1. Morning snapshot: last error budget and SLI achievement, support decisions agreed in the Error Budget Policy
  2. Trends: SLI vs SLO by service/feature over a period of time
  3. Alerting on burnrate: visualize when alerting engage and fade off by sliding window sizes

Extending the SLO generator

The slo-generator tool is designed to add more backends and exporters as it moves forward. Users, customers and Google folks should be able to easily add the metrics backend or the exporter of their choosing.

To prepare for development, you need to fork this repository and work on your own branch so that you can later submit your changes as a GitHub Pull Request.

Once you have forked the repo on GitHub, clone it locally and install the slo-generator in a Python virtual environment:

git clone github.com/google/slo-generator
cd slo-generator
python3 -m venv venv/
source venv/bin/activate

Install slo-generator locally in development mode, so that you can start making changes to it:

python setup.py develop

New backend

To add a new backend, one must:

  • Add a new file slo-generator/backends/<backend>.py

  • Write a new Python class called <Backend> (CamlCase)

  • Add unit tests

Example with a fake Datadog backend:

  • Add a new backend file:

    touch slo-generator/backends/datadog.py

  • Fill the content of datadog.py:

    from slo_generator.backends.base import MetricBackend

class Datadog(MetricBackend):
  def __init__(self, **kwargs):
    # instantiate your client here, or do nothing if your backend
    # doesn't need it.
    url = kwargs['url']
    self.client = DatadogClient(url)

  def query(self, *args, **kwargs):
    # add code to query your backend here.
    return self.client.query(*args, **kwargs)

  @staticmethod
  def count(timeseries):
    # add code to count the number of events in the timeseries returned

  def good_bad_ratio(self, *kwargs):
    # this should return a tuple `(good_event_count, bad_event_count)`
    # or a float `SLI value`.
    good_event_query = kwargs['measurement']['filter_good']
    bad_event_query = kwargs['measurement']['filter_bad']
    good_timeseries = self.query(good_event_query)
    bad_timeseries = self.query(bad_event_query)
    good_count = Datadog.count(good_timeseries)
    bad_count = Datadog.count(bad_timeseries)
    return (good_count, bad_count)

  def my_random_method():
    # this should return a tuple `(good_event_count, bad_event_count)`
    # or a float `SLI value`.
    my_sli_value = self.compute_random_stuff()
    return my_sli_value

  • Write a sample SLO configs (slo_test.yaml):

    service_name: test
feature_name: test
slo_name: datadog
slo_description: Test Datadog SLO
backend:
  class: Datadog
  method: good_bad_ratio
  url: datadog.mycompany.com
  measurement:
    filter_good: avg:system.disk.free{*}.rollup(avg, {window})
    filter_valid: avg:system.disk.used{*}.rollup(avg, {window})

  • Run a test with the SLO generator:

    slo-generator -f slo_test.yaml -b samples/error_budget_target.yaml

  • Create a directory samples/<backend> for your backend samples.

  • Add some YAML samples to show how to write SLO configs for your backend. Samples should be named slo_<service_name>_<feature_name>_<function>.yaml

  • Add a unit test: in the tests/unit/test_compute.py, simply add a method called test_compute_<backend>. Take the other backends an example when writing the test.

The steps above are similar for adding a new exporter, but the exporter code will go to the exporters/ directory and the unit test will be named test_export_<exporter>.

Usage in pipelines

Once the SLO measurement has been tested locally, it's a good idea to deploy a pipeline that will automatically compute SLO reports. This pipeline can be triggered on a schedule, or by specific events.

Cloud Functions

slo-generator is frequently used as part of an SLO Reporting Pipeline made of:

  • A Cloud Scheduler triggering an event every X minutes.
  • A PubSub topic, triggered by the Cloud Scheduler event.
  • A Cloud Function, triggered by the PubSub topic, running slo-generator.
  • A PubSub topic to stream computation results.

Other components can be added to make results available to other destinations:

  • A Cloud Function to export SLO reports (e.g: to BigQuery and Stackdriver Monitoring), running slo-generator.
  • A Stackdriver Monitoring Policy to alert on high budget Burn Rates.

Below is a diagram of what this pipeline looks like:

Architecture

Benefits:

  • Frequent SLO / Error Budget / Burn rate reporting (max 1 every minute) with Cloud Scheduler.

  • Real-time visualization by streaming results to DataStudio.

  • Historical analytics by analyzing SLO data in Bigquery.

  • Real-time alerting by setting up Stackdriver Monitoring alerts based on wanted SLOs.

An example of pipeline automation with Terraform can be found here.

Cloud Build

slo-generator can also be triggered in a Cloud Build pipeline. This can be useful if we want to compute some SLOs as part of a release process (e.g: to calculate a metric on each git commit or push)

To do so, you need to build an image for the slo-generator and push it to Google Container Registry in your project.

To build / push the image, run:

gcloud builds submit --config=cloudbuild.yaml . -s _PROJECT_NAME=<YOUR_PROJECT_NAME>

Once the image is built, you can call the SLO generator using the following snippet in your cloudbuild.yaml:

---
steps:

- name: gcr.io/${_PROJECT_NAME}/slo-generator
  args:
    - -f
    - ${_SLO_CONFIG_FILE}
    - -b
    - ${_ERROR_BUDGET_POLICY_FILE}
    - --export

Then, in another repo containing your SLO definitions, simply run the pipeline, substituting the needed variables:

gcloud builds submit . --config=cloudbuild.yaml --substitutions \
  _SLO_CONFIG_FILE=<YOUR_SLO_CONFIG_FILE> \
  _ERROR_BUDGET_POLICY_FILE=<_ERROR_BUDGET_POLICY_FILE> \
  _WORKSPACE=<ENV>

If your repo is a Cloud Source Repository, you can also configure a trigger for Cloud Build, so that the pipeline is run automatically when a commit is made:

resource "google_cloudbuild_trigger" "dev-trigger" {
  trigger_template {
    branch_name = "dev"
    repo_name   = "my-repo"
  }

  substitutions = {
    _SLO_CONFIG_FILE = "slo.json"
    _ERROR_BUDGET_POLICY_FILE = "error_budget_policy.json"
    _WORKSPACE = "dev"
  }

  filename = "cloudbuild.yaml"
}

resource "google_cloudbuild_trigger" "prod-trigger" {
  trigger_template {
    branch_name = "master"
    repo_name   = "my-repo"
  }

  substitutions = {
    _SLO_CONFIG_FILE = "slo.json"
    _ERROR_BUDGET_POLICY_FILE = "error_budget_policy.json"
  }

  filename = "cloudbuild.yaml"
}