In large part because 99% (+/-) of their traffic is read only. While Facebook and Google have to do heavy workloads for every click and action taken on their services, Wikimedia can cache basically everything. Allowing them to operate on a tiny fraction of the number of machines (and infrastructure) that the rest of the players do.
It hasn’t been this way since around 2013 but again I am fuzzy on how. I think that’s when most such data was switched to TAO, which has local read what you wrote consistency. As long as users landed in the same cluster (and thus TAO cluster) what they wrote was visible to them, even if the DB write hadn’t yet replicated to their region.
FlightTracker postdates my time at FB (ended 2018ish) so I’m not sure how that is used. These systems evolved a lot over time as requirements changed.
I don’t remember anything about writes being batched in memcached and merged in on page load.
What can be interesting, I think, is that you have a completely open infrastructure that has to solve problems on a global traffic scale.
If people are interested in knowing more, I suggest you also take a peek at the wikimedia techblog, specifically to the SRE category https://techblog.wikimedia.org/category/site-reliability-eng... and the performance one https://techblog.wikimedia.org/category/performance/
Do you have a source for this?
I'd be willing to bet that the ONLY reason why 15% of their daily queries "haven't been seen before" is because they add un-needed complexity like fingerprinting. You're making it seem like they've never seen a query for "cute animals" before when obviously they have. They choose to do a lot of extra leg work because of who you are.
So your claim that 15% of their queries have "never been seen before" is probably inaccurate. I'd be willing to bet that "15% of their queries are unique because of the user, location, or other external factor separate from the query itself."
They've seen your query before. They've just never seen you make this query from this device on this side of town before.
15% of the queries themselves are unique. https://blog.google/products/search/our-latest-quality-impro...
https://www.google.com/search/howsearchworks/responses/
I work for Google (and used to work on Search).
Internet users who came online later, from GenZ to many boomers, will often just write conversational sentences and questions.
I know many folks IRL who work at big tech who have no interest in posting here because the community comes off as very unwelcoming. That’s a shame, because they have insight that would be great to hear. Regardless of anyone’s opinion of their employer.
Apologies in advance if your intent was purely about the topic. I just thought I read something in your tone that might hinder discourse rather than encourage it. I wanted to point it out, in case it was unintentional.
On the other hand, if I'm not responding, it's not because I find HN too abrasive — it's because I am afraid of leaking non-public information. That's why whenever I talk about Google, I try to cite a Google blog post or other authoritative source, or talk about my own personal experience; hence, "I rarely search for the same query twice."
I didn't dismiss his argument; I said that he was correct right after he posted: https://news.ycombinator.com/item?id=26073488
"That's not the point" can be interpreted as respectful, but it also can be interpreted as argumentative. I chose to assume good intentions, but I offered a different phrasing that would have a higher chance of not being misinterpreted: i.e. using "yes and" instead of "no but": https://www.theheretic.org/2017/yes-and-vs-no-but/
Having said that, I do think this is clearly a sensitive issues, not a purely technical one. I can appreciate the nuance of working for Google and doing excellent work while seeing the company criticized left and right for its business model. I think given the community, while there is opposition to how Google may at certain points conduct itself as a corporation, there is no lack of respect for any individual working there. I certainly view my comment and the discussion of privacy as having 50% to do with Google’s strategy and 50% to do with the technical aspects of whether you can build a search engine that holds user privacy as a core priority rather than trying to launch an ad hominem on you or anyone. And I saw your other comment that agreed with me and the GP comment so I think my first sentence aside, we are on the same page :)
I doubt you store the history of all searches ever? People don't need a google account to query the engine, others disable history, etc.
Are you saying you still have all searches ever made ever? Because you would need this to say a query hasn't been made before wouldn't you?
"There are trillions of searches on Google every year. In fact, 15 percent of searches we see every day are new"
Does it mean literally the text string typed into the box by the user is new?
Or does it mean the text string combined with a bunch of other inferred parameters we don’t know about is new?
I guess that could be the case. Many could be related to things that are on the news. Like, 'the cw powerpuff girls' for the new show that was announced. No one was searching for that until the announcement, probably
Alternatively, if we assume that google has already recorded 20 Trillion unique search queries (~ 1 Trillion new ones per year for 20 years), the odds that a query composed of 3 correctly spelled english words that are not names has been seen before is 1.6%. Even if we restrict queries to those using the most common 1000 words, there's a 50/50 chance of a query composed of 4 words being unique.
Of course people do not just type random words into the search bar and some terms will be searched many thousands or even millions of times, but still if anything the fact that 85% of searches aren't unique seems surprising.
They briefly mention the statistic in the last paragraph.
Also, speaking of people behaviour, it would not make sense to search everyday for "cute animals", but the volume of searches done for new things people discover as they get older would make more sense. I mean just look at search trends for things like "hydroxychloroquine" for example (and that's not to mention people who get it wrong, i.e. other factors for differing search queries too)
Also, other languages can change the queries depending on how you phrase the sentence too. Add to that the people using other ways to search instead of just visiting google.com and I think you can get pretty close to 10%.
If fingerprinting is the reason, 15% would be a figure too low I surmise. Would that be the case I think that would make probably 20-25% of searches rather than 15%.
It could very well be that they do classify fingerprinted search differently only in some countries and not others? That would/might explain the 15% figure.
I might be wrong and under-estimated fingerprinting techniques for Google. If they have really good fingerprinting techniques, that would reduce the estimate I have in mind to a better number (close to 15, maybe?)
Nobody has ever searched for hydroxychloroquine before today. Today is the day the word is hypothetically invented. Today 2 million people will search for hydroxychloroquine. But only one of them was the first to do it.
What I know about pop-culture and viral internet culture is telling me that 15% of 1 trillion searches being unique is shady math.
So I am not fully convinced that the 15% claim is completely transparent.
The folks who use keyword-based searches are largely those who got on the Internet before ~2007. Tech-savvy, relatively well-off, usually Millenial or Gen-X, plugged into trends. This happens to be the demographic dominant at Hacker News. But there's a much larger demographic who just types in whatever they're thinking of, in natural language, and expects to get answers.
Come to think of it, this is also the demographic that doesn't use tabbed browsing, and uses whichever browser ships with their OEM, and often doesn't realize that there's a separate program called a "browser" running when they click on the "Internet", and issues a Google Search for [google] (#3 query in 2010) when they want to get to Google even though they're on Google already but don't realize it, and doesn't know what a URL is. When a big-tech company makes a brain-dead usability decision you don't like, first consider how that usability choice might appear to your grandmother and it might not seem so brain-dead.
I'm not sure, on my productive days maybe >50% of my Google searches are not very cachable. (for example, I just googled "htop namespace", "htop novel bytes", "htop pss", "htop nightly build ubuntu 14.04")
They need it to be a certain way in order to operate. The limitations and advantages of how software gets made. Why it gets made. The way the software works. How and why product decisions were made over the last 2 decades. What resources they have/had available. It's all a totally different game. Not surprising that different soil and a different climate grow different plants.
One of Google's early coup d'etats, when they were a strategic step ahead of the boomers, was bankrolling gmail, youtube and such. Gmail offered free giant inboxes. They got all the customers. This cost billions (maybe 100s of millions), but storage costs go down every year while the value of ads/data/lock-in and such go up every year. Similar logic for youtube. (1) Buy a leading video-sharing site; (2)bankroll HD streaming because you have the deepest pockets (3) Own online free TV entirely.
That's who Google is, good or bad. How funding works. What products get built. What infrastructure is necessary, possible, affordable. All interlinked. Wikipedia & Google were founded at the same time. Within 5 years (circa 2006) Google was buying charters and fiefdoms. Wikimedia, meanwhile, was starting to take flak for raising 3 or 4 million in donations.
It's kinda crazy that Wikipedia is comparable in scale to FAANGs when you consider these disparities.
The contempt on here is crazy sometimes.
Nobody in this thread is saying that. Parent to you said:
> they could just stop being absurd instead [of building more DCs]
implying FB could build fewer DCs by scaling down some of their per-page complexity/"absurdity". Basically saying their needs are artificial or borne of requirements that aren't.
> conglomerate entities are fundamentally opposed to my right to privacy
That's a common view, but it's not on topic to this thread. This thread is mostly about the tech itself and how WikiMedia scales versus how the bigger techs scale. It has an interesting diversion into some of the reasons why their scaling needs are different.
You could instead continue the thread stating that they could save a lot of money and complexity while also tearing down some of their reputation for being slow and privacy-hostile by removing some of the very features these DCs support (perhaps) without ruining the net bottom line.
This continues the thread and allows the conversation to continue to what the ROI actually is on the sort of complexity that benefits the company but not the user.
Let’s suppose the Facebook cluster spends the equivalent of 1 full second of 1 full CPU core per request. That’s a lot of processing power and for most small scale architectures likely adding wildly unacceptable latency per page load. Further, as small scale traffic is very spiky even low traffic sites would be expensive to host making it a ludicrous amount of processing power.
However, Google has enough traffic to smooth things out, it’s splitting that across multiple of computers and much of it is after the request so latency isn’t an issue, and it isn’t paying retail so processing power is little more than just hardware costs and electricity. Estimate the rough order of magnitude their paying for 1 second of 1 core per request and it’s cheap enough to be a rounding error.
So it wouldn't necessarily be contradictory if most of their core functionality could be replicated very simply, yet the actual product is immensely complicated. I forget where I first read this point, but probably on HN.
Edit: I don’t know what I’m talking about. Happy Monday!
Also, there's way more than just the web tier out there.
Everyone, please listen to Rachel and never ever me.
I recall reading HackerNews used to have that problem, unsure if it still does.