Mini-Thread: AWS Auto-Scaling adding support for predictive scaling today. aws.amazon.com/blogs/aws/new-…
I don't work on Auto-Scaling ... but we've using forecast-driven models for a long time. It's super interesting how they work! It's a mix of signals analysis and statistics.
You can't just look at day-over-day patterns. Most workloads have weekly components - the weekends are quieter. Others have seasonal components; back to school, summer holidays, peak shopping seasons for the holidays, etc ...
There's also a general trend for more patterns. For example, a pattern might be "today is like the same day last week, except that there's a baseline of 2% week over week growth". You get the idea.
You can approximate the periodic trends by using a Fourier transform. That takes a signal, like load or requests/sec, and breaks it down into periodic components. It can find a daily period, weekly, and even some seasonal ones.
But it does a bad job with growth. Also with exceptions, like holidays that move around year over year. ML can do a much better job here, especially since a lot of the trends are cross-industry. Exciting to see how it will develop!
Once you have a model, doing the forecast itself is surprisingly nuanced. Check out the Holt-Winters algorithm if you want to go deep! But basically every input component has predictive error-bars, and you need to be careful about summing them.
Otherwise you end up with massively useless prediction bands! We've been refining these for years. A lot of our internal alarms are based on forecasts. For example if we see the ta rate of requests is significantly below prediction, we investigate. Even if no other errors!
I didn't even know this was coming! I don't work on auto-scaling. Massively cool to see this tech go public. /end-of-thread
