As a general point, a lot of computer science is poorly documented or difficult to find in the literature but well-known among the practitioners that use it. A great example is the recent flurry of activity (including patent activity) around "hyperdimensional hashing" and similar concepts. The irony is that the concepts are so old that some of the original patents have expired but the work is sufficiently forgotten that academics are attempting to re-patent those ideas. I am familiar with it because those algorithms are de rigueur for some types of massively parallel systems (I've been designing parallel systems for ages). No one publishes papers on it though because it is not new to those people. A large part of it is that much of the original literature is lost in paper form and cannot be found electronically. I've been burned by this personally, having done work that I later discovered had been previously done in the 1980s by a respected academic but which is not cited in any modern literature; from the standpoint of modern literature, the work is completely forgotten. But it had been done and rare people exist that are familiar with it.

So more specifically (I got back from vacation a couple hours ago, forgive the terseness):

Spanner is an old, old idea in literature and a number of bespoke systems were built on the concept. Ancient transaction processing literature mentions it, I've worked with a few crusty mainframe environments that used variants of it. It was considered a stupid idea for decades because the hardware required for a reasonable implementation was outrageously expensive for most applications. Atomic clocks and satellite-delivered time codes don't cost what they used to. I used to have a transaction processing book from the 1980s that described a couple specific implementations that were quite close to Spanner (written by a couple guys with Indian surnames, wish I remember the book). It had been written off so long ago for inexpensive commodity systems that apparently it took Google to rediscover it but it never fell out of use.

Similarly, I think a lot of people in the HPC world were surprised to discover that MapReduce was patentable. Not every problem in HPC looks like computational fluid dynamics (CFD), which is often the stereotype. If you are actually an academic then you should know better than anyone that the number of cites is misleading. The MapReduce model (though not under that name) was a standard design pattern for some non-CFD problems in HPC in every respect going back to at least the 1990s. The specific novelty claimed by MapReduce has never been clear to me. A lot of the (terrible) compilers for obscure HPC platforms back in the day transformed the code into a canonical MapReduce model. Maybe I am missing what makes Google's version unique?