SingleStore (f.k.a. MemSQL) used lock-free skiplists extensively as the backing storage of their rowstore tables and indexes. Adam Prout (ex CTO) wrote about it here: https://www.singlestore.com/blog/what-is-skiplist-why-skipli...
When SingleStore added a Columnar storage option (LSM tree), L0 was simply a rowstore table. Since rowstore was already a highly optimized, durable, and large-scale storage engine, it allowed L0 to absorb a highly concurrent transactional write workload. This capability was a key part of SingleStore's ability to handle HTAP workloads. If you want to learn more, take a look at this paper which documents the entire system end-to-end: https://dl.acm.org/doi/10.1145/3514221.3526055
- More info about skiplists: https://arxiv.org/pdf/2403.04582
- Performance comparison with B-tree ?: https://db.cs.cmu.edu/papers/2018/mod342-wangA.pdf
- Other blog from Anthithesis about writing their own db: https://antithesis.com/blog/2025/testing_pangolin/
Also I find it a bit hard to understand the performance outcome of this setup.
I know formats like parquet and databases like ClickHouse work better when duplicating data instead of doing joins. I guess BigQuery is similar.
The article is great but would be also interesting to learn how performance actually worked out with this.
https://nickziv.wordpress.com/wp-content/uploads/2014/02/vis...
I used whatever I could find on the internet at the time, so the comparison compares both algorithm and implementation (they were all written in C, but even slight changes to the C code can change performance -- uuavl performs much better than all other avl variants, for example). I suspect that a differently-programmed skip-list would not have performed quite so poorly.
The general conclusion from all this, is that any data-structure that can organize itself _around_ page-sizes and cache-sizes, will perform very well compared to structures that cannot.
Skiplists also win over balanced BSTs when it comes to concurrent access. Lock-free implementations are much simplier to reason about and get right. ConcurrentSkipListMap has been in the standard library since Java 6 for exactly this reason and it holds up well under high contention
Balanced Skiplists search better than plain Skiplists which may skew (but balancing itself is expensive). Also, I've have found that finger search (especially with doubly-linked skiplist with million+ entries) instead of always looking for elements from the root/head is an even bigger win.
> ConcurrentSkipListMap has been in the standard library since Java 6 for exactly this reason and it holds up well under high contention
An amusing observation I lifted from OCaml's implementation (which inturn quotes Donald Knuth): MSBs of PRNG values have more "randomness" than LSBs (randomness affects "balance"): https://github.com/ocaml/ocaml/blob/389121d3/runtime/lf_skip...
---
Some neat refs from our codebase:
- Skip lists: Done right (2016), https://ticki.github.io/blog/skip-lists-done-right / https://archive.is/kwhnG
- An analysis of skip lists, https://eugene-eeo.github.io/blog/skip-lists.html / https://archive.is/ffCDr
- Skip lists, http://web.archive.org/web/20070212103148/http://eternallyco... / https://archive.is/nl3G8
I’d just store a table of records with the leaf, associated with the seed. A good fuzzer is entirely deterministic. So you should be able to regenerate the entire run from simply knowing the seed. Just store a table of {leaf, seed}. Then gather all the seeds which generated the leaf you’re interested in and rerun the fuzzer for those seeds at query time to figure out what choices were made.
Yes, this is (more or less) how we regenerate the system state, when necessary. But keep in mind that the fuzzing target is a network of containers, plus a whole Linux userland, plus the kernel. And these workloads often run for many minutes in each timeline. Regenerating the entire state from t=0 would be far too computationally intensive on the "read path", when all you want are the logs leading up to some event. We only do it on the "write path", when there's a need to interact with the system by creating new branching timelines. And even then, we have some smart snapshotting so that you're not always paying the full time cost from t=0; we trade off more memory usage for lower latency.
Oh one other thing: the "fuzzer" component itself is not fully deterministic. It can't be, because it also has to forward arbitrary user input into the simulation component (which is deterministic). If you decide to rewind to some moment and run a shell command, that's an input which can't be recovered from a fixed random seed. So in practice we explicitly store all the inputs that were fed in.
They are extremely good at intersections, as you can use the skip pointers in clever ways to skip ahead and eliminate whole swathes of values. You can kinda do that with b-trees[1] as well, but skip lists can beat them out in many cases.
It's highly dependent on the shape of the data though. For random numbers, it's probably an even match, but for postings lists and the like (where skiplists are often used), they perform extremely well as these often have runs of semiconsecutive values being intersected.
[1] I'll argue that if you squint a bit, a skiplist arguably is a sort of degenerate b-tree.
Things like BSP trees are very good at intersections indeed, and have been used for things since time immemorial, but I think the skiplist/tree tradeoff is not that different in this domain.
A correct skiplist is easier to NIH than a correct red-black tree (which for me was the final boss of the DS class in college), but has performance edge cases a red-black tree doesnt, if you treat it like a search tree.
Lately at work I've done C++ optimization tricks like inplace_map, inplace_string, placement new to inline map-like iterators inside a view adapter's iterators and putting that byte buffer as the first member of the class to not incur std::max_align_t padding with the other members. At a higher architecture level, I wrote a data model binding library that can serialize JSON, YAML and CBOR documents to an output iterator one byte at a time without incurring heap allocation in most cases.
This is because I work on an embedded system with 640 KiB of SRAM and given the sheer amount of run-time data it will have to handle and produce, I'm wary not only about heap usage, but also heap fragmentation.
AI will readily identify such tricks, it can even help implement them, but unless constrained otherwise AI will pick the most expedient solution that answers the question (note that I didn't say answers the problem).
With SOTA models it all depends on how you drive them.
LLMs aren't even close to the level of knowledge distillation capacity a human has yet.
The market is telling us that through increased hardware prices.
LLMs being very powerful means that we need to start being smarter about allocating resources. Should chat apps really eat up gigabytes of RAM and be entitled to cores, when we could use that for inference?
Skip trees/graphs sound interesting, but I can't think of any use case for them off the top of my head.
Not convincing. One can write the bulk data which is at first unused - no need to sync anything. Then one writes to the tree DB using, where each node only stores a key to the relevant data.
In practice, I have found out, nothing much. Their appeal comes from being simpler to implement than self-balancing trees, while claiming to offer the same performance.
But they completely lack a mechanism for rebalancing, and are incredibly pointer heavy (in this implementation at least), and inserts/deletes can involve an ungodly amount of pointer patching.
While I think there are some append-heavy access patterns where it can come up on top, I have found that the gap between using a BST, a hashtable, or just putting stuff in an array and just sorting it when needed is very small.
My first instinctive idea would be that there is an optimal distance, maybe based on absolute distance or by function of list size or frequency of access or whatever. Leaving the promotion to randomness is counter intuitive to me.
If you knew all the items ahead of time and didn't require adding/removing them incrementally you could build optimal skiplist without randomness. But in such situations you likely don't need a skip list or BST at all. Binary search on sorted list will do.
> Skiplists to the rescue! Or rather, a weird thing we invented called a “skiptree”…
I can't help but wonder. The article makes no mention of b-trees if any kind. To me, this sounded like the obvious first step.
If their main requirement was to do sequential access to load data, and their problem was how to speed up tree traversal on an ad-hoc tree data structure that was too deep, then I wonder if their problem was simply having tree nodes with too few children. A B+ tree specifically sounds to be right on the nose for the task.
The not so nice thing about it is that you have to pre-guess how deep your tree will be and allocate as many slots, or otherwise itll degrade into a linked list when you run out of rungs, or you end up wasting space.
This is also why skiplists are attractive in research and in early-stage implementations: the randomized promotion converts write complexity (deterministic rebalancing) into write simplicity (probabilistic promotion). You get a weaker worst-case guarantee, but in practice — especially in distributed systems where latency variance is already high — losing the rebalancing phase is often a good trade. The Redis case is the same dynamic: they didn't choose skiplists over red-black trees for raw speed, they chose them because augmenting a skiplist with custom metadata (like the "span" antirez mentions) is local and composable in a way that augmenting a balanced BST is not.