Einstein is often quoted as saying “Make everything as simple as possible, but no simpler.” Last month, I went to a seminar given by Edward Tufte, author of “Beautiful Evidence” and “Envisioning Information” – books that serve as inspiration for many in the field of information visualization. While Einstein advocated for simplicity in describing the universe through mathematics, Tufte argues for allowing data to speak for itself by taking what might be called a minimalist approach to designing visualizations. At the seminar, I learned that Tufte has coined the term sparkline to refer to small yet data-rich line graphs like the following, which take this concept to its limit for one-dimensional time-varying data:
I thought that these sparklines might be an interesting way to represent data on the IMA Dashboard, and so I’m experimenting by bringing this nugget of wisdom to bear on a chart that I haven’t been very pleased with.
This is the chart that we have on the dashboard to display power readings from our machine room (please note that these charts were all generated on Monday, Sept 21st 2009 and thus do not include the readings for that day). The problem here is that I was trying to show three different things, and none of them ended up being communicated very well. Because I’m trying to show individual measurements along with the total, the scale on the vertical axis diminishes the appearance of trends in the data. It is fairly clear that overall, consumption dropped during this period, and that there appears to be a significant decline in usage by the server components. But it is hard to tell if there is also a trend in the cooling system measurements.
The same story about the overall consumption could be told by this sparkline representation of the previous 28 daily totals:
The maximum value over this period is shown by the dot at 1350 kWh, the minimum value is 1222 kWh, and the latest value (from yesterday) is 1227 kWh. This sparkline includes eight values from last month, and you get same story with more precise information added by a single sentence, all in significantly less space than the original graph… pretty potent stuff.
Applying a sparkline to the server readings, 
we can see that after the peak at 807 kWh, consumption dropped and is leveling out around the lastest value at 759 kWh. Okay, so that was an attempt at using the sparkline as a word-like sentence element, which is supposed to be one of it’s strong points. It actually works fairly well, even scaled down to font-size, but here’s the sparkline at the native resolution:
The sparkline clearly reflects a drop in consumption a little over a week ago, which is due to the removal of some units from our server rack.
Similarly, we can look at the cooling system readings, which fell from 545 kWh four weeks ago to 457 kWh when we raised the thermostat setting by 3°F, and has been slowly climbing on average since then to 468 kWh – a fact that is more difficult to derive from the original bar chart.
So, it seems that sparklines can be useful for teasing out information about general trends. But, there are a few things to be careful about. For example, each of these sparklines has it’s own vertical scale, so while trends can be compared, the actual differences in consumption cannot. Also, I might look at those bumps in the cooling system sparkline and wonder if those occurred on warmer days. I may then look at the temperature records and see a pattern that appears similar, but without the actual dates and readings I’d be treading on thin ice in forming any conclusions.
So, I think I’ll add the sparkline to my array of tools, but I’m still deciding how to use it’s principles to improve upon our power consumption chart.
September 24th, 2009 at 4:15 am
The problem with sparklines, as with all simple time series, is the lack of context. Large shifts can be readily seen, but small waves may or may not be relevant, due to inevitable randomness. Sparklines are too small to permit the addition of control lines that would clearly indicate when fluctuation is unusual, and when it’s just the universe jittering. Sparklines work better, I think, when they’re merely pointers and not being used as KPIs in their own right. They should send us to the data for more interpretation rather than hold much meaning in themselves.
September 24th, 2009 at 8:38 am
Tim, thanks for your insight. One good example using the model that you mentioned is the stats sparkline on Flickr, which shows views on your photos over time. Clicking on that sparkline loads up the stats page, where activity can be explored in more detail. While the general trend can be interesting, it’s even more enlightening to drill down and see where an influx of views came from.
This sort of combination might also serve both the general public, interested in trends in general and specifics when something significant happens, and experts who might be more interested in monitoring small changes on a regular basis.
Trackbacks
Leave a Reply