Thanks again for making this!
275 karma · joined May 8, 2012
Thanks again for making this!
I have a tangentially related question, hoping that someone knowledgeable can answer. What are the main bottlenecks to scaling testing, in particular in the following stages:
- the supply of virus-specific reagents (primers, perhaps virus-specific reporting probes - are these necessary, or can just generic reporters be used?)
- the supply of other reagents (in case of PCR, polymerase, nucleotide material, buffer)
- the supply of disposable testing equipment. I know that swabs are in short supply for some reason, anything else?
- safety in handling both the samples and the amplified product (can you basically flush the processed sample down the toilet, or do you need to handle it with extreme precautions)
The other benefit of masks (and probably the main one of DIY masks) would be that it decreases the chance of the wearer infecting others.
I'm not assuming that they are 50% effective either, I pulled the 50% number out of my ass, but so is your 0%, and I find it hard to imagine a physical reality where me keeping a 6 feet distance to an infected person decreases my infection probability significantly more than that person wearing a mask. Anecdotally, South Korea, Japan, Hong Kong, and Taiwan, where wearing masks is quite commonplace all seem to be doing better in transmission rate than the Western world, and all that without killing their economies. And considering that we're producing around 2 billion disposable plastic bottles a day, I'd currently bet money that we should be able to ramp up the supply of cheap masks for a sum that's a lot less than that of halting 50% of the Western economy for several months. Obviously, it can't happen in a day, but I'm not really seeing anyone try, except for perhaps Trump after his recent epiphany that Corona is not a Democrats' hoax. I'm willing to be dissuaded by someone who's actually thought it through, but sadly I haven't seen any such analyses.
Also, where do you draw the numbers of primer production capacity from? IIUC, the only other things that an amplification needs are enzymes (just DNA polymerase for PCR) and "food" (nucleotides), but that I assume is not a bottleneck?
How difficult is it to safely dispose of amplified samples? Can you basically just dump them in the toilet?
Of course, all it takes is a UK to screw everyone over by not playing along and providing a population pool for the virus to mutate. And then a ton of poor countries who probably can't do anything even if they wanted, where it's just starting.
Disclaimer: I work there. Very exciting to see it open-sourced!
Also, when using a system "with consensus" there is often no need to actually invoke consensus on the read side of the system, in which case you don't have to pay the throughput and latency penalties. I know you've sort of said this already, but it might be helpful to mention explicitly.
So I suppose that your definition of nondeterministic must differ, but I can't figure out how.
Assuming that "deterministic" should be defined on lists of inputs, then I don't understand how the approach can be applied to nondeterministic databases. The crux seems of the approach seems to be in consistent distributed reads (which you can get by globally ordering transactions and MVCC). But for fault-tolerance, these must also be repeatable, and I don't see how to make them repeatable if the state of replicas in a shard is allowed to diverge.
Other, less obvious ones included the inherent difficulty in writing the specs. Namely, FM is about verifying that a system model conforms to a specification. For this to bring value, the specification has to be significantly simpler than the model that we can deem it "obviously correct". But such specifications can be extremely difficult to find. On one hand, they are abstractions of the model, and finding good abstractions is one of the hardest parts of software engineering. Another way to think about it is Kolmogorov complexity; at some point you cannot "compress" the description of a problem any further, i.e., some problems are just difficult by nature.
On the incentives front, one of the interesting viewpoints brought up was that having formal specs for your product can make you more liable for defects, and that some (in particular, hardware vendors) might prefer to have wiggle room there. E.g., consider the Spectre/Meltdown fiasco and Intel's response. Another thing brought up there was that software vendors often don't directly feel the cost of the bugs in their software, at least not short-term. This is different in hardware; e.g. ARM uses FM extensively, though I don't know if they commit to specs publicly. The RISC-V architecture comes with a formal spec, though.
On the other hand, I think that many of the current trends in the computing world are making FM more aligned with the incentives of a core part of the industry. The cloud brings centralization, with few large companies (Amazon, Google, etc) providing a lot of complicated software infrastructure that's critical to their success. An outage of an AWS service directly costs them significant money, and can easily completely halt the customers' operations. Moreover, they operate in an open and potentially hostile environment; a corner case software bug in MS Office wasn't a big deal, and maybe could be "fixed" by providing the customer a workaround. But such a bug in a cloud service can become an exploitable security vulnerability. And in fact you do see AWS having a dedicated formal methods team. Also, the TLS 1.3 development included formal verification. On the other hand, AFAIK Google doesn't employ FMs that much, even though they are in the same game as Amazon with GCP; so there's definitely a cultural component to it as well. But I still believe that it's likely that we end up with a small core of companies that provides a lot of critical infrastructure that heavily uses FM, and that we see little adoption of FM outside of this circle.
As for the article, I agree with most points made, though it unsurprisingly has a bias for using TLA with TLC. I agree that TLA/TLC can often be the best solution, but there are caveats. For example, models can be awkward (e.g., recursive data structures), the specification language is limited, and "just throwing more hardware at it" doesn't really work that great for exponential-time problems.
[1]: https://www.amazon.com/Proofs-BOOK-Martin-Aigner/dp/36420085...
Moreover, the leaders are lease-based in the author's proposal, so I believe that what he describes correctly solves consensus. However, his proposal doesn't really give much of a resilience improvement over the original Paxos, since the system, in the worst case, tolerates only 2 failures (leader + one more process), same as a 5-process Paxos. Nevertheless, an interesting idea, one that I don't think I've seen elsewhere.
But other than that, wow! I was already very happy whenever I could get ghc-mod and haskell-mode to work, but this takes the responsiveness to a whole different level.
Interesting, do you have any references for this? I thought that the primary reason for purity was to enable equational reasoning, but I have no sources for this. Also, AFAIK, there are no formal semantics for Haskell?
I do have a personal gripe with that, though. I'm hoping a "high up" might read it so I am posting it here. The only people in my contact list who use the app are crypto nerds. It's really hard to get traction. I thought that the TextSecure people had a great idea to solve this chicken and egg problem; a unified messaging app that would handle your sms, and send secure messages wherever possible. One feature that I find is missing though, is the ability to semd broadcast (unencrypted) sms texts. In my experience (Europe, people around 30) this is used a lot, and a drop-in replacement for the built-in messenger would need to have this.
Actually, I would like to see this happen so much that I would be willing to do the work, if someone is willing to provide the handholding...