New article exploring hardware limits of AES encryption on a single core with VAES capabilities.
Lots of useful links for those looking to catch-up on the SOTA. Otherwise good resource to learn about AES and/or CPU microarchitecture.
https://t.co/tXS8GosBBK
"Everything is a block" (https://t.co/nZzviOYDoU) has some pros, but its reminiscent of linux' "everything is a file", an idea whose time has come.
It's basically https://t.co/6HqSJFGQSn with a single table (which btw is why it was easy for them to shard), but iiuc it also means they can't use CRDT/OT techniques on blocks of text like everyone else does...
@dexhorthy@parzival1l@chribjel Or rather... that's for v4 which is all random. But there are good reasons to use v7 or others with timestamp bits that aren't pure entropy anyways right? Maybe I'm not understanding the entropy argument.
@dexhorthy@parzival1l@chribjel But... UUID type is using almost all of those bytes for entropy?
Also I guess calling UUID bytes "chars" is a bit confusing here?
@parzival1l@dexhorthy@chribjel Yeah I'm confused also. Solution to "stop storing uuids as strings" should be to use native uuid type (effectively u128), not to store nanoids as strings instead.
Happy to see I'm not the only classifying maxi (aka autist) out there.
> Basic Concepts and Taxonomy of Dependable and Secure Computing
https://t.co/CQNDvUKnAa
Ever since reading DDIA (v2; @martinkl@criccomini) I've been wondering how to think or organize non-functional properties:
- performance
- reliability
- privacy/security
- scalability
- maintainability
Just found the right organization, ordering them by timescale:
- dev scale: maintainability
- devops scale: scaling
- many runs (hyperproperties): most SLAs, privacy
- single run: F(x)=y, isolation, consistency, durability
This is so cool actually... fast/slow-path fallback along 3 separate design axes:
1. read/write register vs RMW (consensus)
2. leader-based vs leaderless
3. quorum size
Can we combine them all into one uber-fast protocol in the happy-path?
@AlexMillerDB@LefKok This makes me wonder whether anyone tried to combine these 2 fast-paths into a single protocol! Combining Gryff and QuePaxa to have fast-paths on the data-side and on the network/fault-tolerance side.
Here's he 3 axes with the fundamental limits that each fast path hits:
- FLP and Herlihy are impossibility results. The hybrid avoids them by taking the expensive path only when needed.
- Lamport's and Martin–Alvisi's bounds are prices. You can't avoid them. You pay with more replicas, or, in Alpenglow's case, with a lower f.
- The intersection rule is a safety condition. People long read it as "every pair of quorums must intersect." Flexible Paxos showed that only quorums from different phases must.
@AlexMillerDB@LefKok This makes me wonder whether anyone tried to combine these 2 fast-paths into a single protocol! Combining Gryff and QuePaxa to have fast-paths on the data-side and on the network/fault-tolerance side.
This framework can be pushed further to explain some of the reasoning in your last talk! For example, you mention that QuePaxa is leaderless consensus. But it's actually leader-based, with leaderless fallback!
And afaiu (@LefKok cmiiw) that's the point of QuePaxa: keep the leader as a fast path latency reduction technique, and use an (asynchronous) leaderless protocol as fallback rather than waiting for clocks like Raft does (only works in partial synchronous setting).
And you can see Gryff as a similar fast-path/fallback-slow-path but for ABD-vs-consensus (aka depends on the data dependencies rather than network conditions).