Yes, but it's pretty rare (I assume by data structure you mean like something out of an algorithms textbook, not the structs & classes that most programmers use all the time). I did a string trie for an autocompletion system once (and then found that binary search over a sorted array was faster, go figure), and a binary space partitioning tree for a game creation engine. Also stacks, queues, state machines, and augmented trees come up all the time, but they're not particularly difficult to implement.

I think that one very underrated skillset is the ability to implement a data structure or algorithm on top of an existing platform. For example, can you implement reservoir sampling and a state machine on top of SQL? If you can, there's a good chance your Hadoop analytics cluster can be replaced with a single DB instance and a work-queue system. Can you implement a graph algorithm on top of MapReduce? That's how Google's Pregel system got started. Can you implement a diff algorithm on top of the DOM? React. Three-way merge on S3? DropBox. Binpacking over Linux containers? Borg/Omega/Kubernetes.