Not exactly a datastructure alone, but bitslicing is a neat trick to turn some variable-time operations into constant-time operations. Used in cryptography for “substitution box” (S-box) operations, which can otherwise leak secrets via data-dependent timing variations.

The datastructure side of it is breaking up n words into bits and interleaving them within n variables (usually machine registers), so that the first variable contains the first bit from each word, the second variable the second bit, etc. It’s also called “SIMD within a register”.

I’m also not sure if this is obscure, but Bloom Filters! It’s a structure that you can add elements to then ask it if it has seen the element before with the answer being either “no” or “probably yes”. There’s a trade-off between confidence of a “probably yes”, how many elements you expect to add, and how big the Bloom Filter is, but it’s very space and time efficient. And it uses hash functions which always make for a fun time.

Not really a data structure per say, but just knowing LISP and the interesting structures it uses internally.

The results of LISP operations CAR, CDR, CADR and the other one I can’t remember now.

Conflict free replicated data types, I don’t know if I’d call them obscure but they’re definitely cool and less often used. They’re for shared state across computers, like in collaborative apps

B trees are cool but not obscure necessarily. I didn’t learn about them in college. It sounds like binary tree and it’s similar but it’s different. It’s a data structure to take advantage of the way disk reads work.

XOR lists are obscure and cursed but cool. And not useful on modern hardware as the CPU can’t predict access patterns. They date from a time when every byte of memory counted and CPUs didn’t have pipelines.

(In general, all linked lists or trees are terrible for performance on modern CPUs. Prefer vectors or btrees with large fanout factors. There are some niche use cases still for linked lists in for example kernels, but unless you know exactly what you are doing you shouldn’t use linked data structures.)

EDIT: Fixed spelling

Skew binary trees. They’re an immutable data structure combining the performance characteristics of lists (O(1) non-amortized push/pop) and b-trees (log(N) lookup and updates)
They use a sequence of complete trees, cleverly arranged using skew binary numbers so that adding an element never causes cascading updates.
In practice they’re superseded by relaxed radix balanced trees.

@protein

Finger Tree!

A persistent, purely functional workhorse. Amortized O(1) access at both ends, O(log n) concatenation/splitting.

It generalizes elegantly to build sequences, priority queues, and more. Powers Haskell’s main Data.Sequence. A functional programmer’s secret weapon.

I’ve been knee-deep in these lately so I’m a big fan

Theta sketches!

Do you want to approximately count a large volume of items, but save the state for later so you can UNION , INTERSECT and even DIFF them? Then Thetas are right for you!

Or basically anything in the Apache Datasketches lbrary.

The CSR (compressed sparse row) format is a very simple but efficient way of storing sparse matrices, meaning matrices with a large amount of zero entries, which should not all occupy memory. It has three arrays: one holds all non-zero entries in order, read row by row, the next array contains the column indices of each non-zero element (and therefore has the same length as the first array), the third array indices into the first array for the first element of each row, so we can tell where a new row starts.

On sparse matrices it has optimal memory efficiency and fast lookups, the main downside is that adding or removing elements from the matrix requires shifting all three arrays, so it is mostly useful for immutable data.

Fibonacci heaps are pretty cool. Not used very often b/c they’re awful to implement, but better complexity than many other heaps.

Also Binary Lifting is closer to an algorithm than a data structure but it’s used in Competitive Programming a fair bit, and isn’t often taught: https://cp-algorithms.com/graph/lca_binary_lifting.html

And again closer to an algo tham a data structure, but Sum over Subsets DP in 3^n also has a cool little bit of math in it: https://cp-algorithms.com/algebra/all-submasks.html

I came up with a kind of clever data type for storing short strings in a fixed size struct so they can be stored on the stack or inline without any allocations.
It’s always null-terminated so it can be passed directly as a C-style string, but it also stores the string length without using any additional data (Getting the length would normally have to iterate to find the end).
The trick is to store the number of unused bytes in the last character of the buffer. When the string is full, there are 0 unused bytes and the size byte overlaps the null terminator.
(Only works for strings < 256 chars excluding null byte)

Implementation in C++ here: https://github.com/frustra/strayphotons/blob/master/src/common/common/InlineString.hh

Edit: Since a couple people don’t seem to understand the performance impact of this vs regular std::string, here’s a demo: https://godbolt.org/z/34j7obnbs This generates 10000 strings like “Hello, World! 00001” via concatenation. The effect is huge in debug mode, but still has performance benefits with optimizations turned on:

With -O3 optimization
std::string: 0.949216ms
char[256] with strlen: 0.88104ms
char[256] without strlen: 0.684734ms

With no optimization:
std::string: 3.5501ms
char[256] with strlen: 0.885888ms
char[256] without strlen: 0.687733ms

(You may need to run it a few times to get sample numbers due to random server load on godbolt)
Changing the buffer size to 32 bytes has a negligible performance improvement over 256 bytes in this case, but might be slightly faster due to the whole string fitting in a cache line.

Locality Sensitive Hashing. Not sure if it’s more algorithm than data structure, or even if it’s obscure, but it is very cool.

I get way more use out of Doubly Connected Edge Lists (DCEL) than I ever thought I would when I first learned about them in school.

When I want to render simple stuff to the screen, built-in functions like ‘circle’ or ‘line’ work. But for any shapes more complicated than that, I often find that it’s useful to work with the data in DCEL form.

IMO, circular buffers with two advancing pointers are an awesome data structure for high performance compute. They’re used in virtualized network hardware (see virtio) and minimizing Linux syscalls (see io_uring). Each ring implements a single producer, single consumer queue, so two rings are usually used for bidirectional data transfer.

It’s kinda obscure because the need for asynchronous-transfer queues doesn’t show up that often unless dealing with hardware or crossing outside of a single CPU. But it’s becoming relevant due to coprocessors (ie small ARM CPUs attached to a main CPU) that process offloaded requests and then quickly return the result when ready.