Presenters

Source

Rate Smoothing & Anchoring: A Deep Dive into Prometheus Rate Function Evolution 🚀

Hey everyone! 👋 Let’s talk about something that’s been buzzing in the monitoring world: the Prometheus rate function. It’s a fundamental piece of the puzzle for calculating rates from time series data, but it’s also a surprisingly complex and often misunderstood component. Today, we’re going to break down the recent changes and explore the new approaches of “anchored” and “smooth” rates – and why they matter for your observability stack.

The Current State of Affairs 🧐

For a long time, the standard rate function has been… well, it’s been okay. But it’s not perfect. The speaker, who clearly has a deep understanding of this area, jokingly refers to it as “the mother of all problems.” Let’s unpack why:

  • Range Limitations: The original rate function only looks within a five-minute window. This can lead to gaps in coverage, especially when graphing data over longer periods.
  • Extrapolation Challenges: To get a complete picture, you often need to extrapolate beyond the boundaries of the range – which introduces potential inaccuracies.
  • Non-Integer Results: Extrapolation results in non-integer values, which can be jarring for users expecting precise rate calculations.
  • Composability Issues: A critical, often overlooked issue is that you can’t simply add two rates together directly. This fundamental limitation, known as composibility, was a significant hurdle. 👾

The speaker highlights that the Prometheus maintainers were initially conservative about changes, but this year saw a surge of proposals – X rate, Y rate, H rate, and many more! Thankfully, Julian stepped in to distill this chaos into a coherent design.

The Rise of Anchored and Smooth Rates 🛠️

Recognizing that a single “better” rate function wouldn’t solve everything, the team took a strategic approach:

  • Anchored Rates: This is a radical simplification. It takes the last symbol in the range and the last symbol before the range and directly calculates the difference. It’s incredibly straightforward – and surprisingly, it works well with integers! However, it’s important to note that it only considers a single point in the range. 🎯

  • Smooth Rates: This approach takes a different tack. Instead of extrapolation, it interpolates between the points before and after the range. It essentially “fills in the gaps” between the data points. This sounds fantastic, but it comes with a crucial caveat: it requires a point from the future to function correctly. 🤖

Key Takeaways & Practical Considerations 💾

  • Prometheus Version: To experiment with these new features, you’ll need Prometheus 3.7 or later and the enable feature promote extended range selectors setting enabled.
  • Native Histograms: Currently, these changes don’t work with native histograms. The team is actively working on integrating them.
  • Experimental Nature: Both anchored and smooth rates are still experimental. Expect ongoing improvements and refinements.
  • Trade-offs: As with any technical solution, there are trade-offs. Anchored rates offer simplicity but may lack full coverage. Smooth rates provide interpolation but require future data.

Resources for Further Exploration 🌐

  • Design Document: The full design document, including all the nuances and subtleties, is available [link to design document – placeholder].
  • Brain Dump: The speaker’s infamous “brain dump” – a massive collection of context and proposals – can be found [link to brain dump – placeholder].
  • KubeCon EU Talk: The speaker is submitting a talk for KubeCon EU in Amsterdam next year. Keep an eye out for updates! 📡

The Bottom Line: The evolution of the Prometheus rate function is a testament to the iterative nature of software development. By acknowledging the limitations of the existing approach and embracing experimentation, the team has created a more flexible and potentially more accurate way to calculate rates. It’s a reminder that even seemingly small changes can have a significant impact on your monitoring and observability practices. ✨

Appendix