benjaminheng/tetrio-metrics-web

tetrio-metrics-web's Issues

Compute percentiles using a sliding window

Currently percentiles are calculated for discreet buckets; 0-100 games is one bucket, 100-200 is another, and so on. I would like to calculate percentiles using a sliding window instead, similar to the summary type in Prometheus. Prometheus uses this algorithm: http://dimacs.rutgers.edu/%7Egraham/pubs/slides/bquant-long.pdf

I'm not keen on implementing the described algorithm in javascript though, nor am I keen on importing a package to do so. I'll probably just go with a O(n log n) approach I have in mind, which I expect should be much simpler to implement.

Plot P95/P90/P50 over time, aggregated per week

Currently we plot each individual game, which causes the graph to look quite spiky. Visualizing progression might be easier if we calculate the percentiles for every week and plot it over time. The graph will contain 3 series: P95/P90/P50. For simplicity we can calculate the percentiles for each discrete week, ignoring the current partial week. A rolling weekly aggregation would be nice, but could take a while to calculate.

Add histogram plot of 40L times

The histogram would give an idea of what our average timings would be. I presume mine would have a mode around 55s right now.

I would like to have the date range restricted to perhaps the last 3 months. Consider having multiple series on the same graph if it can be made to look good, for example to compare performance over different months.

Bucket size will probably be 1 second.