Interesting article, but the choice isn't truly a binary one between a SaaS solution and hand-rolling everything from scratch... Snowplow is a pre-rolled opensource analytics stack which runs on AWS, so it's pretty minimal in terms of developer effort and maintenance, and you get to own all your data.
I started building a custom analytics solution like this, you basically passed all the data you wanted to it (from however disparate sources you like), and you could write custom Lua scripts to process them in whatever way you wanted. It had a pretty simple widget/graph API that allowed you to basically return JSON to be turned into a full dashboard.
I didn't manage to get anyone to want to try it out in the year I was looking for potential customers, so I shut it down in the end.
Sounds like our API (Keen IO)! We offer simple client libraries for sending & querying the data. Still going strong over a year later and developers keep telling us how much they love the API. I think going through an accelerator, having some successful blog posts here on Hacker News, and using twitter effectively are what helped us spread the word and develop early traction.
It was similar, but more powerful. Your API has a some predefined functions one can run on the data, the one I built had fully programmable capabilities, with any metric interacting with any other, from any time frame, etc.
What did you do with the code after you closed down? I'd be curious to see what you're doing with Lua. I wrote a behavioral analytics database (http://skydb.io/) in Go & C that uses LuaJIT as part of the low level query engine.
Are you working with event data? (It sounded like it where you mentioned time frames in a different comment) Maybe we're building something similar.
The code just sits on my private repo, for no particular reason. Sky sounds like where I would have gone if I had more interest. As it stands, all I was doing with Lua(JIT) was pass event data for the last X days to a script which processed them and returned the dashboard structure.
The problem with my use was that user the code might be doing absolutely anything, so parallelizing might have been from trivial to hard to impossible. The next step, as I saw it, would be to separate queries and processing in some sort of map/reduce paradigm, or a custom query DSL and normal Lua processing.
I was just building the MVP and waiting to see where customer feedback/needs took me. Nowhere, it turned out.
https://github.com/snowplow/snowplow | http://snowplowanalytics.com/analytics/index.html
(And it's easy to build Chartio dashboards on top of Snowplow data - we're big fans of Chartio.)