1,331 karma · joined October 18, 2010
https://en.wikipedia.org/wiki/Henry_Spencer https://en.wikipedia.org/wiki/Tom_Lane_%28computer_scientist...
Even a revised heuristic that only spots large, individual allocations is not going to do the job.
Oom score adjust also doesn’t do the job: because the only interesting workload is Postgres, if a backend does a page fault that needs memory, who dies? Another sibling Postgres, almost certainly. Then postmaster does crash recovery, which most would rather avoid. High performance databases with distant checkpoints can take a while to come back up.
This is all to underscore the author's point: NAT may necessitate stateful tracking, but firewalls without translation has been deployed at massive scale for one of the most numerous types of device in existence.
The other area I'd like to see some software engineering thinking that's more open ended is on regression testing: ways of storing or referencing old versions of texts to see if the agent can complete old transformations properly even with a context change that patches up a weakness in a transformation that is desirable. This is tricky as it interacts with something essential in software engineering, the ability to run test suites and responding to the outcome. I don't think we know yet when to apply what fidelity of testing, e.g. one-shot on snippets versus a more realistic test based on git worktrees.
This is not something you'd want for every context, but a lot of my effort is spent building up prompt fragments to normalize and clean up the code coming out of a model that did some ad-hoc work that meets the test coverage bar, which constrains it decently into having achieved "something." Kind of like a prototype. But often, a lot of ungratifying massaging is required to even cover the annoying but not dangerous tics of the LLM, to bring clarity to where it wrote, well, very bad and unprincipled code...as it does sometimes.
Steve has gone "a bit" loopy, in a (so far) self aware manner, but he has some kind of insight into the software engineering process, I think. Yet, I predict beads will break under the weight of no-supervision eventually if he keeps churning it, but some others will pick up where he left off, with more modest goals. He did, to his credit, kill off several generations of project before this one in a similar category.
Because there's a lot one could write about each of: equities, real estate, gold, silver, platinum (which have very different industrial exposures), and bitcoin, which have many price drivers.
So let's try something more parsimonious: what do you make of people, institutions, etc that bid on short and even long-dated sovereign debt around the globe, and come up the collective discovered price of, say...3.5%, annualized, for maturity in a month? https://www.treasurydirect.gov/auctions/announcements-data-r...
Not to suggest CPI is redundant, there's a reason why central bankers read it after all. For one, it's the most timely data they have. But it's impossible to nudge it year after year -- accumulative error -- without it become obviously decoupled from other data, including the long-term bond market data. It just so happens commodities are the wrong yardstick.
I like to think about the inherent contradictions of goldbugs going long on central bank portfolio policy: they both tend to distrust the central bank but in a way the central bank activities partially endorse their habits, and are the source of recent appreciation and thus accusations of "hidden" inflation. But central banks operate in an anarchic world system where they need something even independent of reserves held in other sovereign currencies, I presume most gold bugs are holding ETFs in an existing financial system (which is non-orthogonal: if you assume a financial system, why not avail yourself of the superior alternatives?) or have it in a safe in their house which has some other obvious problems.
I hold no gold, if I want hydraulic and non-volatile inflation compensation, it's quite simple: short-dated sovereign debt, aka the humble money market fund, which can be seen as the lower-fee version of the checking account. Nobody likes being a sucker, holding debt for below the time value of money, including changes in nominal value. It has immense price discovery pressure, and it finds its level nicely. If I were to hold gold, I would need some viable theory about how much I should hold to be de-correlated from other assets to be worthwhile. Maybe if I was exposed to jewelry costs and wanted to hedge them.
See https://www.jpmorgan.com/insights/markets-and-economy/market..., https://www.ecb.europa.eu/press/other-publications/ire/focus...
No chance they're going to take risks to share that hardware with anyone given what it does.
The scaled down version of El Capitan is used for non-classified workloads, some of which are proprietary, like drug simulation. It is called Tuolumne. Not long ago, it was nevertheless still a top ten supercomputer.
Like OP, I also don't see why a government supercomputer does it better than hyperscalers, coreweave, neoclouds, et al, who have put in a ton of capital as even compared to government. For loads where institutional continuity is extremely important, like weather -- and maybe one day, a public LLM model or three -- maybe. But we're not there yet, and there's so much competition in LLM infrastructure that it's quite likely some of these entrants will be bag holders, not a world of juicy margins at all...rather, playing chicken with negative gross margins.
I'll give you a real cursed Postgres one: prepared statement names are silently truncated to NAMEDATALEN-1. NAMEDATALEN is 64. This goes back to 2001...or rather, that's when NAMEDATALEN was increased in size from 32. The truncation behavior itself is older still. It's something ORMs need to know about it -- few humans are preparing statement names of sixty-plus characters.
For whatever reason, remuneration seems more concentrated than fundamentals. I don't begrudge those involved their good luck, though: I've had more than my fair share of good luck in my life, it wouldn't be me with the standing to complain.
The Electron Application is somewhere between tolerated and reviled by consumers, often on grounds of performance, but it's probably the single innovation that made using my Linux laptop in the workplace tractable. And it is genuinely useful to, for example, drop into a MS Teams meeting without installing.
So, everyone laments that nothing is as tightly coded as Winamp anymore, without remembering the first three characters.
Building and owning an institution that finances, racks, services, networks, and disposes of servers, both takes time and increases the commitment level. Hetzner is month to month, with a fixed overhead for fresh leasing of servers: the set-up fee.
This is a lot to administer when also building a software institution, and a business. It was not certain at the outset, for example, that the GitHub Actions Runner product would be as popular as it became. In its earliest form, it was partially an engineering test for our virtual machines, and we went around asking friendly contacts that we knew would report abnormalities to use it. There's another universe where it only went as far as an engineering test, and our utilization and revenue pattern (that is, utility to other people) is different.
> The inspection system as currently used by the Skunk Works, which has been approved by both the Air Force and the Navy, meets the intent of existing military requirements and should be used on new projects. Push more basic inspection responsibility back to the subcontractors and vendors. Don't duplicate so much inspection.
But this will be the only and last time Ubicloud does not burn in a new model, or even tranches of purchases (I also work there...and am a founder).
That said, a lot of posts here don't seem to reckon with the fact that a slim majority of www.google.com connections in the United States are via IPv6, and a super-majority from India, Germany, and France. Comcast, T-Mobile, Verizon, as far as I have experienced, these all default to IPv6. While dropping IPv4 support is both a worthy, distant goal and sometimes used in goal-post moving rhetoric, it's not like nobody uses IPv6...rather, mobile broadband networks have depended on it for over a decade (see T-Mobile's deployment of 464XLAT)
A fair number of the dependencies we have also have 100% branch coverage, because I copied the practice, starting about ten years ago, from Jeremy Evans, who maintains a huge number of libraries under that principle. That includes "Sequel," the ORM that I've used for many years and originally copied the practice from, around 2015. You can see the libraries he maintains in this way: http://code.jeremyevans.net/ruby.html. He has joined Ubicloud somewhat recently, so I look forward to getting a sense of how he completes the rest of his rather singular & extraordinary maintenance regime.
To have Ubicloud rest at this standard is my objective. My tendentious claim is as follows: this is higher than any other constellation of libraries I have seen in any programming language. If anyone knows of any constellation of libraries that is more capably and rigorously maintained in any language, let me know. The bar as roughly as follows they need to release every month, or something like that (yes really: https://rubygems.org/gems/sequel/versions/), have a wide interface with your program, and break it no more than once every five years.
It's also not a Rails program, and I have never written or maintained a Rails program in any seriousness, which makes me an odd duck among longtime Ruby programmers.
upgrade_check_ssh = ->(vmh) do
p [vmh.ubid, vmh.created_at, vmh.sshable.host]
vmh.sshable.cmd(<<BASH)
set -xeuo pipefail
sudo apt-get update -qq && sudo apt -qq -y satisfy 'openssh-server (>= 1:8.9p1-3ubuntu0.10)' && sudo systemctl restart ssh.service
BASH
vmh.sshable.cmd(<<VERIFY)
set -xeuo pipefail
dpkg-query --showformat='${Version}\n' --show openssh-server
ssh_pid="$(systemctl show -p MainPID ssh.service | cut -d= -f2)"
(set +e && sudo grep -F deleted "/proc/$ssh_pid/maps" ; [ $? -eq 1 ])
VERIFY
end
cohort_draining = VmHost.where(allocation_state: 'draining').order_by(:created_at)
cohort_draining.map { upgrade_check_ssh.call(_1).tap { sleep 3 } }
I then wrote many minor variants, to change the cohort (eventually targeting all servers), as well as a verification pass. The methodology and output is recorded, along with the time, in Slack. That's how I'm able to roll the tape for you now with precision, almost two months later, in late August. This kind of precision in recall and methodology is important for efficient operations...especially when things go wrong. A common thing we do, upon seeing, say, a broken VM Host, is paste its identifier into slack, to see if it's something of a troublemaker. From people's other code-and-output pastes, we can see what they ascertained, and how, and what was done.
I would not consider a language without a robust REPL for this kind of work. It is connected with an integrated develop-operate model, where the people writing the programs in these symbols every day are also assaying the problems. This unification is key.
And, somewhat related to that, I have not seen JVM nor BEAM libraries as high quality as Sequel, Roda, and Rodauth in their respective functions, and roughly in that order of importance, descending. These dependencies are invasive to how my code is written: above, you see some Sequel. We rely on other libraries being high quality (e.g. the pg driver gem, or net-ssh), but they are less invasive in this crucial way.
I did, at various points, consider applying this methodology to Python (the grammer's whitespace sensitivity is a serious problem, consider "cpaste"), TypeScript, Elixir, Scala, Julia, and even Swift. Although these rather conspicuously have REPLs, none have a Sequel.
I think people could make other REPL-enabled choices that work for them. But in my evaluation, some of the features of these runtimes did not overcome the consideration of a handful of key libraries.
In that common formulation, it would compress consumption by the entire tax+benefit base, that is, everyone would move towards median consumption by some amount, keyed to the magnitude of the UBI, if funded by any kind of proportional taxation (including a nominally regressive proportional tax, like consumption tax/VAT).
Politically, it has tough problems: 18% of the population [over age 65] already has a "MeBI" in the form of Social Security that they can vote to increase, and 22% of the population is below the age of 18, and can't vote. So that's 40% right there. Of the remaining 60% in their working years that produce the output split among themselves and that 40%, quite a few would rather not be compressed towards median consumption: the voting population is shifted higher in the consumption deciles, and people are not often so disposed to think they might find themselves luckless in the future. There's a thicket of "tax expenditures" that can form a "MeBI" for the electorate at the upper-half, like the mortgage interest deduction.
If we look at the difficulty in gaining electoral support in splitting consumption to the benefit of minors (thus, future labor) to even things out a bit, in the form of the semi-recently expired expanded child tax credit, we see the magnitude of the political problem.
Personally, I prefer to see UBI as tax reform to avoid crazy wiggling in effective marginal tax rate. But there are many reasons why it's unlikely that the electorate would see it that way, or approve of it even if they did.
To give a sense how much benefits code and tax code have in common, see this worksheet for SNAP eligibility, which resembles a second tax return: https://www.fns.usda.gov/snap/recipient/eligibility. You get to do something similar, again(!), for Medicaid.
The American benefits code is a patchwork of conflicting sensibilities of the electorate: the smallest possible tax, paternalism and suspicion against the poor, plus a few policy analysis trying to obtain the maximum poverty reduction within those constraints. The result is a thicket of means tested programs with extremely steep phase-outs and a lot of paperwork. The all-in EMTR for an American with income between 0-40K a year is chaotic beyond reason as a result as they roll up the income spectrum.
This person who gave the presentation is indeed in one of the worst cases for the code: a single parent with multiple children.
It's probably about time to swap the default, but some people's once-working pg_upgrade programs that they haven't looked at in a while might break. Probably okay; those things need to happen...once in a while. I suppose some people that resent the overhead of Postgres checksumming atop their ZFS/btrfs/dm-integrity/whatever stacks, but they are somewhat rarer.
How I interpret it: it is a more powerful version of stemming and synonym expansion of information retrieval classics when generating the queries it feeds into traditional information systems (such as the Bing search engine via API, or other index).
After retrieval, it's a selector and summarizer of repetition seen in the results to give you something of a blended outcome, pertinent to the prompt you gave it. Like any other tool, you get a feel for when it has is having problems, and some of those problems can be assessed by at least glancing at the sources it consulted. You get all sorts of weird stuff when your sources don't include relevant results or biased results
The first problem happens when the documents you are searching for do not exist, or something about your prompt -- it's usually obvious what it is -- is not sourcing documents you know to exist.
The second, bias, I've seen when researching something like the design conceits of Infiniband. While it has its genuine virtues, almost nobody talks about it...and many of those things that discuss it are Infiniband marketing materials that are both a bit too fluffy and sometimes stretch the truth, as marketing materials are wont to do. But you can spot this in the sources panel immediately.
I never found "disembodied" LLMs very useful.