I would suspect an instrument failure - it has happened before even in published data which then is tidied up later using other data.
I was thinking more like this
error in HadCRUT discovered by someone who is a software guy, not a climate scientist.