gist.github.com/thisismiller...

After ~2015, the focus seemed to shift to looking at stats on SSD failures from large deployments, but that's no longer a "does this SSD work right?" but a "how long until it dies?", and so I don't get why the latter replaced the former.

I had once started compiling SSD powerfault testing papers, and found that academia testing SSDs stopped ~2015. 😱

If you still have any notes of all the sources you found and looked at, I’d greatly appreciate a copy to update the posts with anything I’ve missed!

I’ve compiled a survey of all the different strategies for torn writes (case #1) on transactional.blog/blog/2025-to... . A different set of 2&3’s durability discussion is covered on transactional.blog/how-to-learn... too.

Starting from not knowing either, I am finding @jayaprabhakar.bsky.social’s fizzbee.io to be much easier to get something running in

I’ll throw it on my queue of potential projects to see if I can show that as correct a bit more formally sometime though. (Maybe finally a good excuse to give fizzbee a try.)

I expect it should be fine to query all nodes “what was your most recently accepted value?” and if you get a quorum of the same answer (by ballot), then you’re welcome to use that as the read value (even if there’s a prepared value). The paper suggesting to propose f(x)=x is only simplest, I think.

Okay, I shouldn’t have said “all” because there’s at least one more trick of pre-ordering your proposals by stamping the requests with a physical timestamp of when to apply them, as in Tempo or Accord, but let’s not get into that right now.

Pretend I said “almost all” :p

Livelock exists in all forms of leaderless Paxos, and you need to either find a way to negotiate your way out of it, or accept that the things can commute and generalized Paxos your way into tolerating it.

Rystsov also proposed a cute trick around this in web.archive.org/web/20231207...

Neon architecture indeed. Unexciting, but a reasonable choice.

How much have you been keeping up with new WCOJ papers, and how much should I interpret this as a strong vote that generic join (/ free join) is a good/practical choice out of the increasing number of alternatives?

I can’t seem to think of blog posts I’ve seen over time that reflect all the things I had to learn by failing at them.

I’m thinking of things like:
* Driving consensus between conflicting (busy) approvers.
* Ensuring you get proper feedback from design docs.
* Doing good work is as important as performatively doing good work.
* What your manager *should* be doing for you.
* How to make the case for a new project.

Socrates was sort of the most unusual one. Would HorizonDB strongly base its architecture on features only Azure Blob Storage has? 🤔

Maybe? The “stateless” XLOG is kind of cool, but also I think their on-disk cache is very important for that. I await details!

I had also assumed that SAL drives replication to page servers so that the compute knows exactly which version a page server is caught up to so that it can avoid sending reads to servers which don’t have the data yet. They wanted to batch updates to page servers, which adds replication delay, so…

I’m surprised you especially weren’t offended by their “our reconfiguration-based replication protocol has 100% availability” argument :p