Skip to content

cumulative-max

Input Stack:
expr: TimeSeriesExpr
 ⇨ Output Stack:
TimeSeriesExpr

Compute the running maximum of a time series across the evaluation time window. Each output datapoint is the largest value seen on the input line from the start of the graph up to the time for that datapoint, so the result never decreases.

This is the max analogue of :integral: a running value from the start of the window, rather than a sliding window like :rolling-max.

Parameters

  • expr: The time series expression to compute the running maximum of

Behavior

  • Non-decreasing: Each value is the maximum of the input up to that point, so the output only ever rises or stays flat
  • Missing data: NaN values are ignored and leave the running maximum unchanged
  • Window relative: The running maximum starts over at the beginning of the graph window, so changing the time range changes the result
  • No leading NaN: Unlike the rolling operations, there is no window to fill, so the first datapoint is already the maximum of the input up to that point

Data Processing

Input :cumulative-max
1 1
2 2
0 2
NaN 2
5 5
3 5

Peak So Far

The common use is showing the worst value seen so far next to the current value. The two lines meet whenever the input is setting a new peak, and the running maximum holds flat afterwards:

name,sps,:eq,
:sum,
:dup,
:cumulative-max,
peak+so+far,:legend,
:swap,
sps,:legend

The same shape applied to the number of viewers of a live event, where the peak reached during the event is held after the audience leaves:

name,viewers.concurrent,:eq,
:sum,
:approx-distinct,
:dup,
:cumulative-max,
peak+so+far,:legend,
:swap,
active+viewers,:legend

Distinct Counts

To get a running count of distinct values, use :approx-distinct-cumulative rather than applying :cumulative-max to a distinct count. As shown above, applying it to the output of :approx-distinct gives the highest count seen at any one time, which is not the same as the total number seen over the window.

Since: 1.9