Recently, I completed a project creating an analytics dashboard for a client. I used Bokeh for the visualizations and wanted to share my experiences.
I'd like to describe the dashboard I created and then talk about some of the steps it took to create. You can see the finished product below:
In the top left corner we have statistics for three servers with line graphs representing CPU, memory, and disk usage. If any graph has a reading above 75% it, its title, and its plot turn red like this. When the readings drop, things go back to normal. Moving downward, we have two line graphs plotting users currently connected to our system and how fast our system is processing their messages over time. At the bottom is a live Google Map of users in the field that are transmitting, with colors representing different user states.
On the top right we have various tables showing running processes, versioning, jar info, uptime, and open file statistics (when available). Below that, we have an alerts area displaying faults, overflows, and server errors. The figures change to red when they exceed certain parameters and the drop downs are shown as needed. Finally, we have the 'firehose' on the bottom right that's fed from tails of various logs from two boxes, colored coded to keep it straight. It's not shown in the photo, but the log scrolls fairly like a terminal, albeit quickly (my client said they wanted it 'for peace of mind', even though the logs fly by at light speed).
You can see some of the elements close up in the images below:
And here is a shot showing the dahsboard updating over time:
I tried to stay with the "server app" design similar to those shown here and kept my folder structure as prescribed here to try and keep things simple, with all my Python in a common folder and a HTML file in a "templates" folder. This was done so Bokeh's Tornado server would find the HTML file and use it as a template for the app. I also used threads and unlocked callbacks for pretty much every element in the app to keep things responsive. The data for all modules is sourced from three daemonized threads, responsible for information from SSH, HTTP, and SQL. They are started by server lifecycle hooks described here, init when the server is launched, gather data from their respective source, and make it availble to the dashboard. And since the dashboard gets data from the a general cache, N users are able to share info cached from a single database, SSH, and HTTP connection. Awesome. This also allows the dashboard to scale nicely, especially with Tornado's help.
I didn't have enough time to give the dashboard a memory, so when you log in you can't see anything from previous runs. My client didn't see this as a dealbreaker, having dedicated plasma screens and projectors for this kind of thing (with some instances showing months of cached data). But it's kind of problematic if anything should happen to those systems, or even the browsers on them. To remedy this, I'd like to save data using locally to a SQlite database and initialize the dashboard with cached data from it to show a week or more of data.