Server(shiny)-less dashboards with R, {htmlwidgets} and {crosstalk}

In this blog post, I want to discuss something that I, personally, have never seen discussed; how to create a “serverless” (or “shinyless” you could say) dashboard using R.

I made one dashboard like that, which you can find here. This dashboard is running on a simple, standard web server. No Shiny involved!

The idea is to create a dashboard with simple tables, graphs, and filters, to communicate results without the need for a Shiny server. The “dashboard” will be a simple html file that only needs a good old web server. Or you could even send the rendered html file per email, and the recipient only needs to open it using a web browser. The shortcoming of that, of course, is that this “dashboard”, which is a simple html file will be static; no computation will be possible (well not quite as you’ll see), so you need to precompute everything that you want to show. It won’t also be possible for the users to enter parameters and then have graphs update automatically. For instance, you cannot let a user choose how many days should be used in a moving average. At best, you can compute three variable, each one with a different number of days, and then let the user choose which of these precomputed variables should be drawn.

But the first question is, why would we want, or need, something so limited?

The advantage of not needing a Shiny server, is that it makes deployment much easier. If you can “deploy” a dashboard that does not need a Shiny server, this means that you don’t need to set up…, well a server. In an institutional setting, this can literally mean you end up saving weeks, sometimes months, of getting the right green lights and signatures. When I worked as a consultant, deployment was definitely the toughest problem to solve (well, toughest maybe after getting access to the data itself). And also, this solution might not be as limited as you think. While it is true that users cannot compute anything on the fly, it is still possible to do a lot of things, which should in all honesty be enough for most use cases. Most users only want or need a glorified Excel with pivot tables and pivot charts. So we’re giving them that, but in a nicer package: the dashboard can be hosted, and users do not have writing rights. That’s honestly all I need in perhaps 90% of the situations.

The solution I’m going to present was in front of me for the longest time; it’s just that I did not put 2 and 2 together. The first part of the solution is {flexdashboard}, which is the framework allowing us to build a dashboard. Dashboards made with {flexdashboard} are simple html files, which can have Shiny elements in them, so for instance an interactive plot that gets generated once the user has entered some input. But these dashboards don’t need to have Shiny elements in them; htmlwidgets are enough. What are htmlwidgets? Take a look at the graph below:

library(plotly)
mtcars %>%
  plot_ly(y = ~hp, x = ~mpg, split = ~am)
## No trace type specified:
##   Based on info supplied, a 'scatter' trace seems appropriate.
##   Read more about this trace type -> https://plotly.com/r/reference/#scatter
## No scatter mode specifed:
##   Setting the mode to markers
##   Read more about this attribute -> https://plotly.com/r/reference/#scatter-mode

You can interact with this visualisation, and it’s 100% running in your web browser. No Shiny involved, even though you can zoom and select different levels in the legend on the top right (try double-clicking on the “0” level for instance). This visualisation was made with the {plotly} package, one of the many htmlwidgets available. My favorite for making such visualisations is {echarts4r} which I’ve used to create the following map (how-to blog post here). htmlwidgets bring JavaScript visualisations (and other goodies) to R, and what’s really cool about them is that they don’t need a Shiny server to run (that’s the whole point of JavaScript, everything runs in the browser). So this means that by combining {flexdashboard} with the right htmlwidgets we can create a simple, yet useful, dashboard that can be deployed as a web page.

To illustrate, I’ve made the following dashboard, which shows tables, graphs, and even a pivot table of COVID-19 cases and deaths of the Greater Region (to know more about the Greater Region and why this interests me currently, you can read this).

Something else I need to talk about: on the very first tab, you can see a sidebar with some inputs that the user can interact with. For instance, the user can choose which country’s data should appear on the table. It is also possible to filter the positive cases data (not the deaths, but this could be added). This interaction between the sidebar and the table (which was made using {DT}) was made possible using the {crosstalk} package. This package makes it possible to link several htmlwidgets together, but they have to be compatible with {crosstalk}. Unfortunately, at the time of writing, not many htmlwidgets are compatible with {crosstalk} (see here), but I would say that the ones that are compatible still make it possible to create some pretty useful stuff.

The only thing you need to do to link htmlwidgets with each other is to convent the dataframe holding your data to a SharedData object:

data_set_shared <- SharedData$new(data_set)

Widgets compatible with {crosstalk} can now use this SharedData object instead of the regular dataframe, and this is how you link them: through this SharedData object.

Another tab that uses {crosstalk} is the last one, where you can take a look at the weekly positive cases and deaths for the countries of the Greater Regions (but only for the sub-regions of these countries composing the Greater Region). Here, the user can choose whether deaths or positive cases should be shown. The plot updates immediately, and it’s also possible to focus on a single country by double-clicking on it in the legend on the top-right. Again, it’s also possible to focus on a particular month. Here I wanted to use a slicer like on the first table, but on the date. This should work (I’m using exactly that on another dashboard I made), but for some reason here, it would not work. The dashboard would compile without any error message but trying to open the html file on my browser would make the browser hang. So I settled for another type of slicer. Something else that is quite cool; if you choose to focus on the cases, you can hover the mouse over the bars and see how many cases there were in the sub regions in each country. For this, I had to change the default behavior of the popup in the {plotly} visualisation.

Now comes the cherry on top of this already delicious cake; on the second tab, you can interact with a pivot table! This makes it possible to, for instance, see how many deaths there were in each country, region or sub-region, on a weekly basis. You can even switch from a table to several types of visualisations! This pivot table is made possible using the very nice {rpivotTable} package. This package is honestly nuts. It feels like it shouldn’t work so well, and yet, it does work beautifully. Seriously, play around with it in the dashboard, it’s pure magic.

One final note; on the top right of the dashboard you can click on “Source Code” and read the dashboard’s source code. You will notice that I use two functions, tar_load() and tar_read() that can be found in the {targets} package. I will be explaining what that is exactly in a subsequent blog post, or perhaps a video on my youtube channel. You can also see how the inputs in the sidebar work, and how they are linked (through the SharedData object) to the visualisations they control.

In any case, I’m quite happy that I found the possibility to develop dashboards without the need of a server, where all the logic is handled client-side by the web browser. I think that this definitely can help many of you that need to communicate results fast to stakeholders without the need to deploy a full server, which can often take quite a long time.