In other words, MUMPS is a platform and an ecosystem as much as a language. Think of Java or Ruby — for a lot of companies, including MUMPS shops, staying with a specific "sub-ecosystem" is simply the most rational choice because they have so much invested it already.
If you look beyond tech that is currently considered "bleeding edge" — Go, JavaScript, Ruby and so forth — you will find a lot of companies who rely on what you may consider weird or even legacy software. For example, Delphi (a descendant of Borland's Turbo Pascal which is still based on ObjectPascal) is still very popular. In finance, languages like K are still popular. I believe finance still has a ton of stuff based on object databases such as Objectivity/DB, Versant, Matisse and GemStone (Smalltalk), which actually look a lot like today's document-oriented databases. InterSystems Caché, which is based on MUMPS, is a hybrid SQL/OODBMS. In other words, the software market has a lot of aging technology that is still working superbly for the parties involved. Old code is usually proven code.
This system was originally put together in the 80s and while a complete reimplementation from scratch in Python is possible, it wouldn't really give us that many benefits that we don't already have. We don't have to worry about algorithms all that much - we may just write an SQL query which takes care of things for us. One of the more complex things we do in new code is maybe two levels of $order, the equivalent of your for x in y loop.
We're programming on a level quite a bit higher than C, nor are we exposed to the sheer verbosity of Java, thank god. You think Caché ObjectScript is awful? I think Java is awful.
What Caché gives us is tight integration between the language and the database system, so in that way the choice isn't really between Caché ObjectScript and Python, it's more between PL/pgsql or PL/Python and Caché ObjectScript. Once you go there you'll realise that the code that's written in ObjectScript isn't really affected that much by the choice of programming language anyway. And PL/Python isn't really something you'd want to write a production system in anyway. Between that and built-in ORM, things really aren't that bad.
And 100x readability improvement is just wrong. Sorry, it is. Sure, that may apply to some old MUMPS code that survived the 70s when disk space was sparse but even this is fairly straightforward to expand into something quite readable. We do have syntax highlighting, function calls look the same, and assigning a value to an object is 'set object.Property = "blah"' instead of 'object.Property = "blah"', a difference that's quite trivial.
That's not to say the lack of libraries isn't annoying, but everything really important is there in the right places. I've written a decompressor for tar.gz in about 34 quite readable lines, with error handling and all. gzip is built in these days so it's really mostly .tar I needed to worry about. Alternatively, calling out is just a matter of $zf(cmdLine, -1). Similarly, I've just put chosen (https://github.com/harvesthq/chosen) into our web application, based on Caché. It was easy enough to do.
And it's a pretty top-notch fast SQL implementation with a nice built-in language and a nice ORM and lots of other bonuses like full-text search and bitmap indices and OLAP cubes (yes, it speaks MDX even) if nothing else.
If Caché had no strengths, I would agree with you, but as a non-relational database, it's pretty good. Sparse associative arrays are the default data structure, so it's very popular in, e.g., medical applications, where you would want to be able to store thousands of different things, but any given patient will only need a few of them.
We're pretty happy with it. If, theoretically, we'd need to start from scratch, it'd would definitely be in our top very very few choices. Not least of all because of things like DeepSee.
I could certainly think of things much worse than Caché.
Caché has really been moving away from that though... CacheStorage stores data in a big list in the *D globals.
I think the power in CacheStorage really comes from the indices and - very relevantly to the healthcare industry - all the relationships. They've got a very complex schema they need to support and Caché continues to be pretty good at that kind of thing - see the implicit joins in their SQL variety for example or Zen.
We do make a lot of use of indices, though. Finding records can be incredibly fast.
I'm so sorry. Nothing more fun than manually keeping indices in MUMPS :|