The Case for Learned Index Structures:
https://www.sigarch.org/the-unreasonable-ineffectiveness-of-...
However, it's certainly possible that the time for this idea has come. Google is probably in the best position to apply it.
I will say that after having worked at Google for over a decade, some of it on data center performance, there are plenty of inefficiencies that have nothing to do with algorithms or code, but have more to do with very coarse-grained resource allocation. In other words, they are "boring management problems" rather than sexy CS problems.
I should also add that you're expanding the claims made in the paper, which appears to be a common mistake. If you look at the related work and citations, you'll see that it's mostly about B-Trees, bloom filters, and indices. But I have seen people generalizing this to "a revolution in systems research", which I don't believe is justified.
Indexed data structures are important but a revolution in them is not going to cause a revolution in all of systems. Off the top of my head, the general area of "scheduling" (in time and space) is probably more important and more potentially impactful.
Look at it this way. As the paper points out, a Hashtable is just a heuristic that works fairly well in the worst case and reasonably well in the average case. No one would argue that you couldn't hand-roll an algorithm that is better for a specific, narrowly defined data set. This paper is demonstrating that you don't have to do it by hand, you can train a model automatically without the expected caveats like: size (doesn't have to be huge), latency (the model trains very quickly) or specialized hardware (they used bog-standard CPUs, not GPUs or TPUs).
This is obviously just my opinion, but I think it's pretty big. It's not big because of the paper itself, although that's amazing in isolation (they improved on hashtables!). As I said above, it _points the way_. They have a throw-away sentence about co-optimizing data layout along with lookup and not showing those result. My guess is that beats a hashtable in every way. More importantly, if a model can beat bloom filters, b-trees, and hashtables, you'd better hold onto your socks because most algorithms are just waiting to fall. To me this paper is Deep Blue vs Kasparov. We all know what comes next.
Also, predicting power usage is a different problem than scheduling. It's not "managing Borg" by any stretch of the imagination. I'm not saying that's impossible, but it's not what was claimed.
Also, what did Deep Blue lead to? I'm not saying it wasn't impressive, but it didn't generalize AFAIK. You are making a huge leap and claiming that a technique that isn't even deployed generalizes.
What did Watson lead to?
https://www.technologyreview.com/s/607965/a-reality-check-fo...
https://respectfulinsolence.com/2017/09/18/ibm-watson-not-li...
https://www.theverge.com/2016/7/21/12246258/google-deepmind-...
The future is here. :)
https://research.google.com/pubs/pub43438.html
My overall point is that even if learned indexes replace traditional data structures (which is a big if), plenty of important systems problems that remain. Some that I could think of:
- Fixing front end performance as mentioned in a sibling comment.
- System administration of a billion Android phones. The state of these seems pretty sorry.
- Auditing app stores for security bugs.
- Eliminating security bugs by construction. Hell, even just eliminating XSS.
- Software that runs transparently in multiple data centers. (Spanner is a big step in that direction, but not a panacea.)
I could name like 10 more... none of these have anything to do with learned indexes. The point is that the OP is generalizing from a really specific technique to some kind of hand-wavy "machine learning writes all the code" secnario.
AFAIK storage is not the system bottle it used to be. We always want more, but network and cores are relatively plentiful.
If we could magically (and safely) modify the software stack, which areas could give x2 or x3 improvements ?
Or are there also plenty server-to-server scenarios?
e.g. devs using remote build system ? A local workload which access a mostly-remote db ?
I'm being totally serious. Backends are generally fast, and the backend engineers are performance-minded.
Front end engineers are not as cognizant of performance (somewhat necessarily, since arguably they have a harder problem to solve). Back in the mid-2000's, Gmail/Maps/Reader days Google had a lot of great JS talent, but it seems to have ceded some of that ground to Facebook and Microsoft.
If you have heard Steve Souders speak, he always mentions that he was a backend guy. Until he actually measured latency, and realized that the bottleneck is the front end. That was at Yahoo, but it's very much true for Google too.
http://stevesouders.com/bio.php
I would like to see a machine learning system rewrite JavaScript code to perform better and make UI more usable. I believe that's beyond the state of the art now, but it's probably not out of the question in the near future.
-----
As far as scheduling that was just one example of an important systems problem that hasn't been solved with machine learning. Not saying it can't be, of course. Just that this is a research direction and not a deployed system.
It's also important to note that there are plenty of other feedback-based/data-driven algorithms for resource management that are not neural nets. If neural nets work, then some simpler technique probably works too.
Well sure, we all want a God compiler.
https://research.google.com/pubs/pub43146.html
This is from the team at Google with the longest experience deploying machine learning (early 2000's, pre-deep learning).
(Contrary to popular belief, machine learning was basically unused in Google's search ranking until pretty recently.)
https://jobs.zalando.com/tech/blog/accelerating-warehouse-op...
Reading again of the abstract, allow me to add that it's an optimization beyond structures such as Btrees which are already not really decipherable datasets to the naked eyes. So any optimization in a space that's not designed to be human-legible can take whatever shape, and this is certainly a welcome achievement!
Can you expand on this bit? I'm not sure I understand what you mean
Since writing this comment I have text-to-speech'd the whole of the article and I listened to it one afternoon over smooth classical music and I have come to appreciate the fact that many computers will have neural network processors like tensor processors or at the very least really good GPU, making it all the more likely that such algorithms [especially used in the specific case of fault intolerant data organization and file hierarchy] will actually not require so much computational effort as traditionally believed.
Besides HN how can one find great papers like this? How did you find out about it?
Additionally, there is http://www.arxiv-sanity.com/ which sorts new machine learning papers by popularity.
http://learningsys.org/nips17/assets/slides/dean-nips17.pdf
You can generally find a paper for each section.
2) This is my personal ritual. Every weekday I check out recent entries in "Hardware architectures" and "Emerging Technologies" section of arXiv.org Many other sections are gaining traction with me, like Database.
Hope this helps.
Whether or not it applies to online databases probably depends on several factors. The ones I can think of off the top of my head are the read/write ratio, the write rate in general, and obviously size of your data set (although the approach works at GB, not TB, so most can take advantage).
Here's a way, I, a systems engineer who doesn't design databases, could imagine using it immediately:
1) Do I have any indexed queries? If not, nothing to do here. 2) Does the data change rapidly? If so, nothing to do here. 3) Okay, now I'm in the zone of profitability:
1) investigate just letting the model do all the work. This is unlikely to be successful in many cases, but it's pretty easy to try.
2) investigate "model-assist", in this mode I let the model give me an answer and, if I don't like it, I use the B-tree. The model is fast enough that this is not a huge hit to my overall performance.
3) investigate "model-replace", in this mode I build my b-tree and do everything like normal. In the meantime I'm training my model. Once my model seems well-trained (I could hand it a randomized subset of queries and, once the accuracy and performance beats the b-tree I consider it well-trained), I switch to using it. As data changes I could potentially go back to the b-tree, update the model (literally throw it away and re-build) and do it again.
4) Investigate "co-opetition" -- make the b-tree a component of the model. If you read the paper, you see how they layer different models. A b-tree is just a particular kind of model that happens to be computationally expensive but has high accuracy. You could include it in the model graph and let it compete for queries like everything else.
Anyway, I'm speculating here, but it seems obvious this has immediate impact in one large space relating to databases (data warehousing) and potential impact in many of the rest.
I am not too worried about data warehouses. Data warehouses can be fed by batch processes, which you can run, say, once every day or once every week. This leads to the design of algorithms that optimize aggregate measures of performance (e.g. amortized complexity, average complexity, throughput) rather than microscopic measures of performance (e.g. worst-case complexity, latency). Machine learning techniques have a track record of delivering good aggregate results.
However, online databases require optimal worst-case performance. The occasional massively slow operation is not okay - we'd rather every operation be a tiny bit slower instead.
> Do I have any indexed queries? If not, nothing to do here.
Of course I do. Needless to say, complex schemata come with lots of indices that are necessary to speed up queries that pull data from 10-15 tables each.
> Does the data change rapidly? If so, nothing to do here.
Of course the data changes rapidly. Crucially, an online transactional database's input comes from external parties (e.g., public-facing web servers) that cannot be considered trustworthy. Many machine learning techniques are vulnerable to so-called “adversarial examples”. Feeding adversarial examples to an ML-powered index could cause the database's performance to drop drastically in ways that are provably not possible with B-trees.
Because I would take the other side of that bet. :)
I'm just stating my concerns.