And all those careers and all that capex was profit-taking along the way. Not a LOT of profit. But enough to keep reinvesting surplus US dollars into additional capex in China.

There are too many layers of permission needed to innovate, including groups whose interests run counter to innovation.

Construction projects are undertaken within a legal and regulatory system that presents persistent, costly obstacles, while projects are being overseen by agencies who lack the resources and in some cases even the expertise to manage them.

You can’t connect information if you don’t know it.

Imagine you're an engineering manager. Who would you rather hire: the person who knows exactly what features are available in PHP 7 and which are only available in PHP 8, or the one who will figure it out by trial-and-error of while writing each application and seeing what fails? Of course, the second engineer certainly may produce quality work. But the first one unquestionably has an comprehensively organized framework of the tools he has at his disposal.

Mathematical software providers would have preferred to standardize high-level language facilities, but there was no chance of getting language standards interested until there were some hardware implementations and some compilers and libraries proving the concepts.

GLUT framework

20,000 clicks/second and 1,000,000 queries/second.

In practice, we’d want to add more processes to increase parallelism—to handle accumulated backlog after any downtime or traffic spikes.

We’ve just described an example of the well-known consensus problem in distributed systems engineering.

We need to evolve our design to use more than one datacenter.

all its work must be redone

keep implementations consistent

we know from our one-machine iteration that the individual small writes are too frequent to store on a hard drive

there are potentially three ad_ids that could be clicked on for each search query

ignore ad_id and search_term because they are a small linear factor

We can delete data that’s too old to be of value.

introducing a new distributed system component

For simplicity’s sake

we can round our 86.4 TB/day up to 100 TB of space required

we can use scientific notation to limit errors caused by arithmetic on inconsistent units

because the average CTR is 2% (10,000 clicks / 500,000 queries), the click log will have 2% as many records as the query log.

scanning through the query log and the click log to generate the dashboard will be very inefficient

Achieving this SLO requires that the speed of calculating a CTR remains constant as the system handles large amounts of clicks and queries.

Click logs are derived from URLs, which have inherent size limitations, making the separate query log a more scalable solution.

scale up our basic design

invent a design that works in principle

This analysis either feeds into the next iteration or indicates when the design is good enough to recommend.

CTR is the ratio of times the ad is clicked versus the number of times the ad is shown.

Ultimately, we arrive at a system that defends against many failure modes and satisfies both the initial requirements and additional details that emerged as we iterated

By following an iterative style of system design and implementation, we arrive at robust and scalable designs with low operational costs.

problem

a coordination problem

I think this viewpoint was developed and espoused (in the form of funding preferences and memos) by inaugural IPTO director at ARPA (1962-1964) J. C. R. Licklider.

M. Mitchell Waldrop’s biography

Protocols are just agreed ways to communicate. A protocol embodies an architecture of participation.

These are both ecosystems that provide infrastructure by harnessing market forces. Get it right, and the incentives align towards getting cheaper, better, and more accessible.

until you realize that it’s still just incredibly tiny compared to real-world code bases

Our expectation is that you will build on these techniques, improving the tools and models.

No, I think the big problem will be that legal system only works for humans.

People readily find humanity in the unlikeliest of places

a tree

By the early 1900s, Cleveland, Ohio was instrumental in pushing forward the frontier on electricity and steel.

Next time you hear someone complaining about a baby crying on an airplane, tell them they’re being a jerk.

While immigration undoubtedly helps with an aging population, it also has a destabilising effect on democracies, which leads to bad electoral outcomes and policies.

Malthusian trap

The real differentiator is the collection, curation and use of data to improve user experience of end users, whether for the development of foundation models or in applications.

Coming from game development

it is easier to modify, build on, and maintain

Instead of building MVPs, we should be building SLCs. Something Simple, Loveable, and Complete.

One night, in desperation, I made a long list of concepts that I wanted to reflect in V, moving from one to another with a rapid free- association that would make any good psychiatrist reach for the emergency cord. The list was something as follows;Orwell. Huxley. Thomas Disch. Judge Dredd. Harlan Ellison’s “'Repent, Harlequin!’ Said the Ticktockman.” “Catman” and “Prowler in the City at the Edge of the World” by the same author. Vincent Price’s Dr. Phibes and Theatre Of Blood. David Bowie. The Shadow. Nightraven. Batman. Fahrenheit 451. The writings of the New Worlds school of science -fiction. Max Ernst’s painting “Europe After The Rains,” Thomas Pynchon, The atmosphere of British Second World War films. The Prisoner. Robin Hood. DickTurpin…

high coherence of process information, high process complexity and high process controllability (whether manual or by adequate automatics) were all associated with low levels of stress and workload and good health, and the inverse, while fast process dynamics and a high frequency of actions which cannot be made directly on the interface were associated with high stress and workload and poor health. High process controllability, good interface ergonomics and a rich pattern of activities were all associated with high feeling of achievement. Many studies show that high levels of stress lead to errors, whitc poor health and low job satisfaction lead to the high indirect costs of absenteeism, etc. (e.g. Mobley and colleagues, 1979i.

The level of skill that a worker has is also a major aspect of his status, both within and outside the working community.

The second problem is that if the decisions can be fully specified then a computer can make them more quickly, taking into account more dimensions and using more accurately specified criteria than a human operator can. There is therefore no way in which the human operator can check in real-time that the computer is following its rules correctly.

We know from many 'vigilance" studies (Mackworth, 1950) that it is impossible for even a highly motivated human being to maintain effective visual attention towards a source of information on which very little happens, for more than about half an hour.

Manual operators may come into the control room quarter to half an hour before they are due to take over control, so they can get this feel for what the process is doing.

There is some concern that the present generation of automated systems, which are monitored by former manual operators, are riding on their skills, which later generations of operators cannot be expected to have.

One is that efficient retrieval of knowledge from long- term memory depends on frequency of use (consider any subject which you passed an examination in at school and have not thought about since).

This means that a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one.

the operator can be left with an arbitrary collection of tasks, and little thought may have been given to providing support for them.

The difficulty remains that they are less effective when under time pressure

"even highly automated systems, such as electric power networks, need human beings for supervision, adjustment, main.tenance, expansion and improvement. Therefore one can draw the paradoxical conclusion that automated systems still are man-machine systems, for which both technical and human factors are important."

he only operation that can’t be completed is thecreation of a newfile for which the inode is never written;in this case the directory entry will be removed

Sprite LFS currently uses acheckpoint interval of thirty seconds, which is probablymuch too short.

In order to handle a crash during acheckpoint operation there are actually two checkpointregions, and checkpoint operations alternate between them.

The key to achieving high performance at low cost ina log-structuredfile system is to force the disk into a bimo-dal segment distribution where most of the segments arenearly full, a few are empty or nearly empty, and thecleaner can almost always work with the empty segments.

Figure 3 suggeststhat the segments cleaned must have a utilization of lessthan .8 in order for a log-structuredfile system to outper-form the current Unix FFS; the utilization must be less than.5 to outperform an improved Unix FFS.

we callthis approachage sort

The version number combined with the inode number forman unique identifier (uid) for the contents of thefile.

If it does, then the block is live; if itdoesn’t, then the block is dead

segment summary block

This allowswhole-segment operations to run at nearly the fullbandwidth of the disk, regardless of the order in which seg-ments are accessed

Once afile’s inode has beenfound, the number of disk I/Os required to read thefile isidentical in Sprite LFS and Unix FFS.

The fundamental idea of a log-structuredfile systemis to improve write performance by buffering a sequence offile system changes in thefile cache and then writing all thechanges to disk sequentially in a single disk write opera-tion.

In designing a log-structuredfile system we decided tofocus on the efficiency of small-file accesses, and leave itto hardware designers to improve bandwidth for large-fileaccesse

several studies have measured meanfilesizes of only a few kilobytes

Buffering

for applicationsthat require better crash recovery, non-volatile RAM maybe used for the write buffer.

Disk transfer bandwidth

echnology, which provides a set of basic buildingblocks, and workload, which determines a set of operationsthat must be carried out efficiently.

System

Why Aren’t Operating Sys-tems Getting Faster AsFast as Hardware?

collect large amounts of new data in afile cache in main memory, then write the data to disk in asingle large I/O that can use all of the disk’s bandwidth

nix systemscan only utilize 5-10% of a disk’s raw bandwidth for writ-ing new data; the rest of the time is spent seeking

t segregates older, more slowly changing datafrom young rapidly-changing data and treats them dif-ferently during cleaning.

February 1992

With chaining, increased collisions means an increased number of items on each chain.

Go has a culture of rejecting large dependency trees, and of preferring a bit of copying to adding a new dependency.

Right now, roughly 95% of the hydrogen produced in the U.S. comes from natural gas. It’s produced through a process called steam methane reforming and emits roughly 830 million tonnes of carbon dioxide per year(link is external)

Published: 27 Dec 2013

What we should do about this is a big topic, in the time we have left, one thing we can do instead of writing to files is to use databases.

Life is inherently full of unavoidable and heartbreaking suffering. At some point, you or everyone you love will die. It’s unavoidable. So, in between bouts of unavoidable suffering, we benefit from seeking out the beauty life entails.

We need to rediscover our ambition to improve public health as we did in the 18th and 19th centuries.

6: int mid = low + ((high - low) / 2);

'm going to make another note about the close()/dup() combination since it's pretty weird. close(1) frees up file descriptor 1 (standard output). dup(pfds[1]) makes a copy of the write-end of the pipe in the first available file descriptor, which is "1", since we just closed that. In this way, anything that ls writes to standard output (file descriptor 1) will instead go to pfds[1] (the write end of the pipe).

read from the other end in the order it came in

pipes will fill up after you write about 10K to them without reading anything out.

File descriptors are simply ints that are analogous to FILE*'s in stdio.h.

Since all source programswere always available and easily modified on-line, we werewilling to revise and rewrite the system and its softwarewhen new ideas were invented, discovered, or suggested byothers.

Second there have always been fairly severe size con-straints on the system and its software

First, since we are programmers, we naturally designedthe system to make it easy to write, test, and run programs.

Our goals through-out the effort, when articulated at all, have always con-cerned themselves with building a comfortable relationshipwith the machine and with exploring ideas and inventionsin operating systems.

dissatisfied

floating point hardware

There is also a quit signal which is used to force a coreimage to be produced. Thus programs which loop unex-pectedly may be halted and the core image examined with-out prearrangement.

Such faults cause the processor to trapto a system routine.

Thus when editing system userslog in, they are inside the editor and can begin work imme-diately;

Thus a user may log out simply by typing the end-of-file sequence in place of a command to the Shell.

This is easy because, byagreement, the smallest unused file descriptor is assignedwhen a new file is opened (or created); it is only necessaryto close file 0 (or 1) and open the named file.

The

When this happens, the Shellknows the command is finished, so it types its prompt andreads the typewriter to obtain another command.

The Shell also returns immedi-ately for another request.

Some filters which we have found useful

Actually it would be surprising, and in fact unwisefor efficiency reasons, to expect authors of commands suchas ls to provide such a wide variety of output options.

in fact it is interpretedcompletely by the Shell and is not passed to the commandat all.

the command need merelyuse the standard file descriptors 0 and 1 where appropriate.

ile descriptors 0 and 1

If file command cannot be found, the Shell prefixes thestring /bin/ to command and attempts again to find the file.Directory /bin contains all the commands intended to begenerally used.

Wa i t may also presentstatus from a grandchild or more distant ancestor;

Ordinarily, arg1 should be the same string as file,so that the program may determine the name by which itwas invoked.

the pipe must be set up by a commonancestor of the processes involved

flies

using the same system read and write calls that are used forfile system I/O

fork (label)

the size of which may be extended by a systemcall.

We will notattempt any interpretation of these figures nor any compari-son with other systems, but merely note that we are gener-ally satisfied with the overall performance of the system.

Thecurrent version of UNIX avoids the issue by not chargingany fees at all.

To the user, both reading and writing of files appear tobe synchronous and unbuffered.

These files may be as large as 8⋅256⋅512, or l,048,576 (220)bytes.

If the link-count drops to 0, any disk blocksin the file are freed and the i-node is deallocated.

This pointer is an integer called the i-number (forindex number) of the file. When the file is accessed, its i-number is used as an index into a system table (the i-list)stored in a known part of the device on which the directoryresides. The entry thereby found (the file’s i-node) containsthe description of the file as follows.

It is possible to generate an end-of-filefrom a typewriter by use of an escape sequence whichdepends on the device used

he byte array

truncates it to zero length if it does exist

Since theactual user ID of the invoker of any program is alwaysavailable

There is a threefold advantage in treating I/O devicesthis way: file and device I/O are as similar as possible; fileand device names have the same syntax and meaning, sothat a program expecting a file name as a parameter can bepassed a device name; finally, special files are subject to thesame protection mechanism as regular files.

byconvention

UNIX differs from other systems in which link-ing is permitted in that all links to a file have equal status.That is, a file does not exist within a particular directory;the directory entry for a file consists merely of its name anda pointer to the information actually describing the file.

Files are named by sequences of 14 or fewer charac-ters.

All files in the sys-tem can be found by tracing a path through a chain of direc-tories until the desired file is reached.

symbolic or binary (object) programs

However, the structure of files is controlled by theprograms which use them, not by the system

The PDP-11 has a 1M byte fixed-head disk, used for filesystem storage and swapping, four moving-head disk driveswhich each provide 2.5M bytes on removable disk car-tridges, and a single moving-head disk drive which usesremovable 40M byte disk packs

All of these programs were writtenlocally. It is worth noting that the system is totally self-sup-porting.

preparationand formatting of patent applications

(1) a hierarchical file system incorpo-rating demountable volumes; (2) compatible file, device, and inter-process I/O; (3) the ability to initiate asynchro-nous processes; (4) system command language select-able on a per-user basis; and (5) over 100 subsystems including a dozen languages.

The list of arguments must be terminated by a null pointer, and, since these are variadic functions, this pointer must be cast (char *) NULL.

The word meritocracy has been around since the late 1950s, when a British sociologist named Michael Young published The Rise of the Meritocracy. He meant this new word as a warning

consumers often come to expect a certain level of performance from its implementation.

Brinch Hansen’s excellenthistory

which would skipped over by JavaScript engines

But if those trends we mentioned above continue, compiling away your types might be the only step between writing your TypeScript and running it, and we don’t want to be the ones standing in the way of a good developer experience!

evergreen browsers

You don’t send a message unless you need to teach someone else a lesson. The one you’re making an example of isn’t the one you’re trying to teach a lesson. It’s the ones who are next in line. Putin is telling us he won’t stop here. We are the intended recipients of this message he is sending by making an example of Ukraine, in blood and death.

It helps to jot down the keypoints, or to make comments in the margins, as you read.

read their related worksections

that is,making the same assumptions as the authors, re-create thework.

To fully understand a paper, particularly if you are re-viewer, requires a third pass

Common mistakes like these will separaterushed, shoddy work from the truly excellent.

you can expect mostreviewers (and readers) to make only one pass over it

Thefirst pass is adequate for papers that aren’t in your researcharea, but may someday prove relevant.

A typi-cal researcher will likely spend hundreds of hours every yearreading papers

However, this skill is rarely taught, leading to muchwasted effort

10x speed

I highly recommend recording your screen.

Recovery

we could shard the set of files into multiple directories.

Assuming a disk IO rate of 100MB/s (ballpark range for modern drives)

level

a special young level

These merges have the effect of gradually migrating new updates from the young level to the largest level using only bulk reads and writes (i.e., minimizing expensive seeks).

This copy is consulted on every read so that read operations reflect all logged updates.

You’ve written a note. Mark the page with a sticky flag or dog-ear it for access later.

With ALPN, the client gives the server a list of protocols in its order of preference and the server picks the one it wants, while with NPN the client makes the final choice.

ALPN is being promoted for use by http2, while SPDY clients and servers still use NPN.

Build an MVP” is the mental image that comes to mind, but “Get One User” is a better answer.

Learning more about software systems is a compounding skill. The more systems you’ve seen, the more patterns you have available to match future systems against, and the more skills and tricks and techniques you develop to apply to future problems.

It’s nearly always worth trying the easier approach first (upgrading a dependency, reaching for the debugger, a few passes of trial-and-error cargo-culting from working examples), and only reach for the big guns if those tools fail you.

However, that wasn’t my goal; I just wanted something that worked, within a time budget. And so, I instead just ended up without anything that worked, and without much of a better understanding of anything, either.

They apparently had an explicit goal of a 100% rate of root-causing kernel crashes based on a single crash report. In order to strive for this goal they had built a lot of elaborate crash-reporting and debugging technology — it wasn’t just raw thinking hard about bugs — but at root I think this goal comes from the deep belief that their system, while complex, is understandable and mostly deterministic, and that they have the ability to reason about it.

These bugs practically require finding someone on your team who is comfortable moving between multiple layers of the stack, in order to track down

checkout of that library’s source code on my laptop

spend time learning more about the systems they depend on, and how they work and how they are implemented.

Because of this compounding effect, the most important step toward becoming a blub master is to kickstart your “blub flywheel”—the virtuous cycle of blub accumulation—however you can.

“the destiny of computing is to become interactive intellectual amplifiers for all humanity pervasively networked worldwide”.

The reason it pays to put off even those errands is that real work needs two things errands don't: big chunks of time, and the right mood. If you get inspired by some project, it can be a net win to blow off everything you were supposed to do for the next few days to work on it. Yes, those errands may cost you more time when you finally get around to them. But if you get a lot done during those few days, you will be net more productive.

Look with suspicion upon every comfortable moment.

“No one has the right to be an amateur in the matter of mental training. It is a shame for a person to grow old without seeing the beauty and strength of which the mind is capable.”

What does the output tell you?

In the late nineteenth century, the wealthy taxpayers of Boston were convinced to build out water and sewage systems by a straightforward logic: every person in the city, rich or poor, needed clean water to drink every day. Without it, they would be at risk of contracting water-borne diseases like cholera. And with so many people in proximity, wealth alone couldn’t provide protection from contagious disease.

The life of a tech journalist can be soul-sapping. I talk to so many founders — capable, smart, well-connected, experienced founders — who are working on problems that genuinely just don’t matter. They might have successfully raised money. Perhaps they even rounded up a team of smart people like them. They are solving problems that even if they execute everything A++ and they change the world exactly in the way they envision, even if they make their board and VCs happy, it’s ultimately so fantastically futile. I don’t want to name any names. Not to inspire existential angst, but if you’re reading this, and you’re feeling a twinge of “god what am I doing with my life,” well, sorry, friend, I am talking about you.

I need to be more trusting and slightly less hardcore.

I’ve introduced a new benchmark, BenchmarkRand, to measure the overhead of the RNG. We can substract the resulting overhead from the other benchmarks for an approximation of their actual runtime.

And of course, nearly all of the internet would exist without them.

VR headsets aren’t the Metaverse any more than smartphones are the mobile internet

like saying the mobile internet is an app

In other words, we will constantly be ‘within’ the internet, rather than have access to it, and within the billions of interconnected computers around us, rather than occasionally reach for them, and alongside all other users and real-time.

Once a few plants began this transformation, the entire market was forced to catch up, thereby spurring more investment and innovation in electricity-based infrastructure, equipment, and processes.

When plants first adopted electrical power, it was typically used for lighting and/or to replace a plant’s on-premises source of power (usually steam). These plants did not, however, rethink or replace the legacy infrastructure which would carry this power throughout the factory and put it to work.

int is still 32 bits