mjb

20 ms·

181.
by mjb 4y ago
My team at AWS has been using P heavily for modelling and reasoning about distributed storage and database protocols and designs. I'm a big fan. The sweet spot for P in my mind comes from the fact that it's a much more approachabl
182.
by mjb 4y ago
It's constantly surprising to me how, given how airline safety has improved significantly decade after decade (see https://www.bts.gov/content/us-air-carrier-safety-data , for example), that people still have this
183.
Lambda Snapstart, and snapshots as a tool for system builders(brooker.co.za)
2 points by mjb 4y ago|0 comments
184.
by mjb 4y ago
I really love this point. For most applications the optimal size of a (time-based) cache is around the point that keeping an item around in the cache is equal to the cost of accessing it again. This is a point that Jim Gray made in the clas
185.
by mjb 4y ago
Indeed. See "Why INSAG has still got it wrong" by Anatoly Dyatlov himself. The money quote: > How and why should the operators have compensated for design errors they did not know about? Dyatlov is clearly biased here, but he r
186.
by mjb 4y ago
But why? Why have an additional component if your database can do it?
187.
by mjb 4y ago
The idea in ZAMM that has stuck with me the longest, and affects my daily thinking most, is gumption traps. Those little things, little frustrations, that make you lose enthusiasm for a project or task. Hard things can be enjoyable. Even ch
188.
by mjb 4y ago
I enjoyed reading this analysis, even if it is quite specific to a single situation. I agree with the writer's larger point that you need to do this kind of analysis for your workload, in your conditions, to draw the right conclusions.
189.
by mjb 4y ago
Thanks! One more: my understanding is that in Nudge I need to know processing time, but in FCFS I don't. How sensitive is your optimality result to errors in processing time estimates (I don't recall this being covered in the pape
190.
by mjb 4y ago
> For seasonality, if the arrival rates fluctuate over a long time period relative to the typical waiting time, it makes sense to just do separate calculations for the different conditions you experience. If the fluctuation is very fast,
191.
by mjb 4y ago
Simulation, as kqr said. I like to use the classic queue theory results (e.g. results on M/M/1 and M/M/c) as unit tests for the simulators I build. I've found that application of queue theory very helpful in avoidin
192.
by mjb 4y ago
That really is a great book, and very approachable! All the examples are practical, results derived clearly, and a nice build up. Here's my question for you: queue theory is a nice tool, but most of the classic results are about system
193.
by mjb 4y ago
> And is that model common across other proof systems like TLA? Yes, it's pretty common, and is a really nice abstraction. TLA+ doesn't use that model, although a lot of TLA+ code is implemented that way. TLA+'s model is b
194.
by mjb 4y ago
> For this reason, formal proofs of correctness are valuable for distributed algorithms. I have a massive amount of respect for Martin and his work, but I think that this emphasis is the wrong one (if our goal is to increase the correctn
195.
by mjb 4y ago
A while ago I wrote this as a simple introduction to applying MC methods in distributed systems: https://brooker.co.za/blog/2022/04/11/simulation.html
196.
by mjb 4y ago
> the Monte Carlo method lets you punch above your weight class in terms of measuring and predicting phenomena that are too complex, or too expensive to deterministically model. Yes, this is exactly why I like it. At AWS, we've used
197.
by mjb 4y ago
I agree with all of that. The defensiveness that comes with writing for HN (and Twitter) makes me a worse writer too. It's not just about being careful and precise, but about being super defensive to avoid the most uncharitable possibl
198.
by mjb 4y ago
It's always great to see different ways folks are using Firecracker. Snapshot-and-restore is a particularly cool capability, especially if you solve the data movement problems (like these folks have). One challenge to clone-and-restore
199.
What Is Backoff For?(brooker.co.za)
3 points by mjb 4y ago|0 comments
200.
by mjb 4y ago
To clarify, I am not a kernel expert.
201.
by mjb 4y ago
Making providing high-quality randomness without compromises a first-class OS feature seems like a great idea. Especially because it reduces the chances of using a bad/incorrectly seeded/inadequately seeded/cloned from snapsh
202.
by mjb 4y ago
As expected from Daniel Abadi, this is good stuff. The cross over between isolation levels and consistency levels adds a lot of complexity, which is a real problem when most developers don't deeply understand isolation levels already.
203.
by mjb 4y ago
Measured per gigabyte, which seems a little bit silly for a lot of applications (especially server applications). For example, a modern SSD can do as many random 4kB read IOs in a second as the average hard drive can do in a week! So if you
204.
by mjb 4y ago
Bonus all-time great paper review: "Jim Gray once told me that he had heard two different opinions of this paper: that it’s trivial and that it’s brilliant. I can’t argue with the former, and I am disinclined to argue with the latter.&
205.
by mjb 4y ago
Yes, that's exactly right. Lamport clocks provide a partial ordering of events, based on the idea of "happened before". Specifically, Lamport defines the happens before operator →, such that within any one thread/process
206.
by mjb 4y ago
The language I want is a specification language (a bit like Pluscal or P) which allows me to specify distributed system behavior, then ask both correctness (liveness/safety) questions and quantitative behavior questions. I wrote a bit
207.
by mjb 4y ago
R is my go-to for this, and one I recommend often. It makes a nice calculator language (with only a little weird syntax), has a very rich standard library of functions, and is reasonably fast. The great thing about R is that it scales well
208.
Amazon DynamoDB: A Scalable, Performant, and Managed NoSQL Database Service(usenix.org)
3 points by mjb 4y ago|0 comments
209.
by mjb 4y ago
> It's good to exercise the "everything is down/recover from zero" path now and then. Yes! The more often the better. Over the years I've been at AWS, we've focused more and more on building systems with re
210.
by mjb 4y ago
I think that's a reasonable criticism, but I'm more optimistic that there's value here for two reasons. One is that even a "basic proportional" level of control is under-used, and most practitioners don't even

More ›