For par condicio: I worked at a similar optimization for Redis using different techniques (when I submitted the PR my colleagues informed me that ValKey and another Redis team was working at it, I was not aware: the amoung the three implementation we picked mine after extensive benchmarks) reaching memory savings that sometimes can be tens of percentage of less RAM compared to this ValKey optimization. So if you are exited for less memory for sorted Sets you should try both the implementations. We also got a lot more speed in many use cases compared to the skiplist.
It uses two techniques: B+ tree instead of sorted sets, but also the hash table side, the dictionary, is specialized for sorted sets and stores several items per allocation. The B+ tree leaves hold all the members and scores (also in clusters of many elements, 96 max or 4k: that's a big part of the win), rank lookup can skip many children because of meta data, which is part of the ability to go as fast or faster than Redis even if the memory is so much smaller. The hash table points to leaves in the B+ tree, does not have elements at all, and is heavily optimized to do just the minimum, a single byte tag is used as a speed optimization to optimized lookups in an optimistic way. Basically it is the sum of many different techniques conceived during the years and never adopted for fear of complexity (today no longer a problem), and newely discovered techniques.
P.S. I tried several variants with different representations, so this specific implementation is the Pareto frontier in speed/size of many, many attempts.
I have to say I was always on the fence about Bill Pugh (Mr Skiplist).
Some of his concurrency ideas were pretty good but Skip Lists were about the same time I was realizing that theoretical and practical design for performance had diverged and skip lists were a huge cache miss liability.
It didn't hurt that a few people I was aware had tried them had gone back to something else due to live performance analysis being negative or inconclusive (If you can't prove one thing is better than the other, you use the simpler one, and skiplists are not simple)
If cache misses are a huge concern and we have a highly skewed distribution (mostly touching newer records), the splay tree starts to look interesting.
We don't like the O(n) worst case, but we really like that magical amortization property. There are some simple ways to avoid getting into a linear hot path.
Timsort and powersort both have O(k) operations which rely on k being approximately a cache line in size.
One of the last large dev blog entries for V8 was about how they got pointer compression to be performance neutral and that’s down to getting the translation down below the amortized cache hit penalty for 64 bit pointers. IIRC the JVM gets close to that as well.
B+tree were designed around optimized IO in the era of magnetic storage, they're inherently cache-friendly. Skiplists are the same thing as a b+tree but the structure is inverted, which makes skiplists more to concurrent writer-friendly than b+trees. Different tradeoffs.
The data structure is not a pure B+tree. They also strip common key prefix for node elements and store 4 bytes of remaining key data together with a pointer to the children to avoid key lookup in many cases. Plus they tune node sizes to match jemalloc allocation classes to minimize allocation overhead. I am not sure with such focus on cache friendliness the advantage of skip lists for writes still holds as they will still have more cache misses even on inserts.
Yep. I’ve implemented b trees and skip lists a few times. I find skip lists to be a little easier to implement. But b trees seem to perform ~20% faster. I can’t think of any good reasons to use skip lists instead of b trees in high performance code.
Curious how rank queries (ZRANK, ZRANGE by index) perform after the switch. Skiplists made those cheap via span counters; does the B+tree keep per-node subtree counts to match?
This is the standard LLM wall of irrelevant garbage that always makes me groan when I see it. There's a lot of words but it conveys very little useful information. This is exactly the type of thing that LLMs are terrible at and has negative ROI for human reviewers.
This is an LLM smell and, after having dealt with this sort of thing and realizing it’s usually low-signal high-noise, you end up ignoring this kind of thing outright.
I prefer to go straight to the code when reviewing. I’ll form my own opinion, then read the PR description, then do another, more detailed pass through the code.
Like all things, it depends. If you don't need the latest and greatest features in upstream redis, and you don't want license headaches, Valkey is probably a good default. It's generally cheaper on hosting providers too - AWS will discount you ~20-35% for Valkey vs Redis.
I haven't checked the newest additions, there were some updates about adding login and encryption, but the latest innovation I use from redis is xstream, and almost all clones now support it.
And a moderate LLM subscription can readily generate a useful portion of this project overnight, which makes me concerned about their profitability in the long run.
I seem to remember it was xstream when it was first released. The commands all have a prefix X (XADD, XREAD, XTRIM) perhaps it got rebranded into just stream.
You might also want to try Lavik (https://github.com/eloqdata/lavik). I'm one of the authors. It's an open-source Redis-compatible server that uses NVMe storage while delivering in-memory-class performance. It's particularly fast for strings and hashes, though sorted sets aren't its strong suit yet.
Curious. Skiplists don't need back pointers. You can traverse a skiplist backward almost as fast as forward if you keep the "finger" around. The finger is the set of nodes from level-n down to level-0 that point as close to the goal node as possible without going beyond it. The ideal place to keep these nodes is on the stack. To traverse backwards, you do a conventional search for the starting node but every time you go down a pointer level, you push the last pointer on the stack first. Then back traversal becomes stack pops together with appropriate forward recursive calls when you pop back to a pointer that was too far to the left.
It's easier to convey the idea with code than english but I'm not typing the algorithm with my thumbs into my phone.
Compute speed has been outpacing memory speed for decades now. So when you spend the cycles to fetch a cache line (64 bytes generally), you want to milk it for all its worth, to minimize time spent waiting on memory.
On a cold cache, you pay for each level of indirection, so you want to increase fan-out as much as possible. B-trees do that intrinsically.
Binary trees have worse cache locality. The cost of accessing memory has grown so much compared to computation that supposedly slower "heavy weight" data structures are worth it now (as in the last 10-20 years).
A binary tree would be strictly worse than a skip list for cache line efficiency. Seeing as how skip lists were invented to improve on binary tree cache behavior.
I'm guessing their point is unlike binary trees which have two children per node, B trees can have as many children per node is best for cache locality/SIMD/etc. Which despite being "heavyweight" (wasted space if not all children per node are occupied) is faster in practice because it's cheaper to preload large unused blocks to cache rather than only access memory when it is needed.
Every node in a binary tree needs a value and two child pointers, though. That's asymptotically almost twice as many child pointers per value as a B-tree has.
The comment I replied to claimed that B-trees are "heavyweight" compared to binary trees. I thought it worth pointing out that the claim is only true for quite small trees.
Yes, that hardware is called "the memory cache". B-trees were invented because disk reads were very expensive. Now, memory cache fills are very expensive.
From what I've seen on here many commenters didn't read the article, only the title (I've been guilty of this too). OP is maybe just adapting to HN culture.
As a coding exercise I implemented a redis that utilized skiplists. I was excited to read this article. Turns out it's ai slop. Seriously people, stop it. This hurts to read.
P.S. I tried several variants with different representations, so this specific implementation is the Pareto frontier in speed/size of many, many attempts.
Some of his concurrency ideas were pretty good but Skip Lists were about the same time I was realizing that theoretical and practical design for performance had diverged and skip lists were a huge cache miss liability.
It didn't hurt that a few people I was aware had tried them had gone back to something else due to live performance analysis being negative or inconclusive (If you can't prove one thing is better than the other, you use the simpler one, and skiplists are not simple)
We don't like the O(n) worst case, but we really like that magical amortization property. There are some simple ways to avoid getting into a linear hot path.
One of the last large dev blog entries for V8 was about how they got pointer compression to be performance neutral and that’s down to getting the translation down below the amortized cache hit penalty for 64 bit pointers. IIRC the JVM gets close to that as well.
That cost varies from chip generation to generation, but it does seem to be the case that b trees win in more generations than they lose.
Can you please clarify: Simpler in operation or simple in implementation? Or neither/both?
There is a reviewing guide with suggested reading order, clear design decisions, and scope boundaries for what is intentionally not included.
This is the type of thing LLMs are very good at and high ROI for human reviewers.
[0]: https://github.com/valkey-io/valkey/pull/4206
Not sure about the name reviewing guide, but I do like the contents. Then again I know nothing about code base.
And a moderate LLM subscription can readily generate a useful portion of this project overnight, which makes me concerned about their profitability in the long run.
(Disclosure, worked for them 3 years ago for a few months)
Curious. Skiplists don't need back pointers. You can traverse a skiplist backward almost as fast as forward if you keep the "finger" around. The finger is the set of nodes from level-n down to level-0 that point as close to the goal node as possible without going beyond it. The ideal place to keep these nodes is on the stack. To traverse backwards, you do a conventional search for the starting node but every time you go down a pointer level, you push the last pointer on the stack first. Then back traversal becomes stack pops together with appropriate forward recursive calls when you pop back to a pointer that was too far to the left.
It's easier to convey the idea with code than english but I'm not typing the algorithm with my thumbs into my phone.
At this point have we optimized hardware to assume B trees in a sort of self-fulfilling prophesy?
On a cold cache, you pay for each level of indirection, so you want to increase fan-out as much as possible. B-trees do that intrinsically.
[0] https://en.wikipedia.org/wiki/B-tree
The PR for the article. I'm not convinced it's AI, but I didn't do a deep dive on the commit tree just poked around a bit.