mmazur / slo-evaluator

Evaluate&Store Engine for OpenSLO–defined Objectives

slo-evaluator

Evaluate&Store Engine for OpenSLO–defined Objectives

How it works

SLO Evaluator is intended to be run at predefined intervals (like a cronjob) and on each perform the following steps:

1. (TODO) Read configs defininig data sources and data sinks.
2. Read all yaml files from the SLOs directory and parse all SLOs defined therein (see example-SLOs dir).
3. For each SLO defined connect to the data source, retrieve required data, evaluate the SLO for the current time window and (TODO) store the results in the appropriate data sinks.

Dev Env

Requires Python >= 3.9.

1. Clone this repo.
2. Make sure you have all the python modules from requirements.txt.

   python3 -m pip install -r requirements.txt
   

3. Generate some fake data like so:

   ./test/datagen.py --id-column-header sensor --ids=cpu,system,gpu --column node=klapek --column instance="localhost:9100" -c 200000 -s 60 -o test/node_temps.csv
   

4. Give it some test files and run the code:

   ln -s example-SLOs SLOs
   ./cronrun
   

You should see some example SLOs being calculated from the fake data you just generated.

Design

• All configs are in yaml format.
• SLO configs are in the OpenSLO format.
  • Deviations from the upstream spec should be kept to a minimum. Working with upstream on extending the spec where needed is emphasized.
• The code is in Python.
  • Data manipulation is done using Pandas DataFrames.
• In principle any combination of data source and data sink should be supported.
  • However the initial design is based around supporting Prometheus and influences of that choice might be visible in places.

TODO

• Unit tests all the things.
  • Start with functions that perform calculations, those are critical to be kept stable.
• Set up CI in the project once the tests are up.
  • Maybe a linter while we're at it (PEP8 at least?).
• Add ratioMetric support.
• Fully support Prometheus as a data source.
  • Waiting on this prometheus-api-client PR to land first.
• Spin off source–agnostic calculation code into its own module.
• Add some basic cli options to cronjob using Click module.
  • Log level, defining where configs are, maybe something else.
• Add an initial data sink.
  • SQL? Prom using pushgateway?
• Morph objectiveMetric in current code into whatever the new metric becomes (discussion).
• Reduce python dependency to at most 3.9, possibly earlier.