OP here, apologies if this is not suitable for HN. I posted this not as much for highlighting the fact that datadog does this. But rather in the hopes that there might some discussion about alternative approaches and maybe some critique.
So from the look of this, this is something I have been trying to find for some time. I guess only way to tell if it's as good as it seems is to try it. If anyone else knows any alternatives for doing anomaly detection in stats, open source or other services it would be great if you could share it here.
What's your use case? The way the blog post is written suggests that their method performs a decomposition of the timeseries into trend, seasonality, and error. They then create a confidence interval, that's probably based off moving smoothed mean and sd. The interface looks nice, but the math behind it is probably nothing special.
It's probably worth noting that the video and sample graphs show very smooth movement. I'd garner that many of the metrics you care about will not show this behavior in practice.
2 comments
[ 3.4 ms ] story [ 10.3 ms ] threadSo from the look of this, this is something I have been trying to find for some time. I guess only way to tell if it's as good as it seems is to try it. If anyone else knows any alternatives for doing anomaly detection in stats, open source or other services it would be great if you could share it here.
It's probably worth noting that the video and sample graphs show very smooth movement. I'd garner that many of the metrics you care about will not show this behavior in practice.