Stephan Garland

@nateberkopec I would slow all queries more and more every day, until I was given the green light to move the DB to bare metal as Codd intended.

Imagine that one of the DBs for your app suddenly had 100ms added to every call. You need to access this DB currently 1 to 30 times per transaction. What would you do to compensate for this added latency?

@AdamRackis @tylermac @bravelyjake Sure – so they can boost executive compensation and shareholder profits.

@tylermac @bravelyjake I’m just dumbfounded at the arrogance at cavalierly saying “well these multi-national, global corporations should ‘just’ improve their margins” - like … do you not think they already relentlessly chase better margins whenever they can

Do these people ever get “radicalized” into learning about profit margin?

@mipsytipsy With all due respect, you absolutely did (ISBN-13: 978-1491925942).

I did *not* write a book on "igniting your purpose". 😬 That would be the *other* charity majors, pastor's wife and life coach and my personal nemesis.

@NielsHoven I had this experience at Samsung. Sitting in on meetings where PhD chemists were discussing chip yields, pulling up various charts at dizzying rates, absorbing them, and moving on before I had read the axes. I consider myself decently intelligent, but that was on another level.

Grad school was where I first had the experience of maxing out my cognitive horsepower. I was one of the dumbest EE majors. There was a very clear hierarchy. [Generalizing heavily] the EE majors weren’t as smart as the stat majors, who weren’t as smart as the math majors. There was no value judgement about it. It didn't make anybody "better" than anybody else, it was just a blindingly obvious fact of life while we were there, like noticing that some people were taller than other people. I took the intro level stats course that every 1st-year stats grad student took, and I barely understood half of it. I didn't take it for a grade, and it's a good thing because I probably wouldn't have passed. The next year, I tried to take the intro level math class that every 1st-year math grad student took, and I had to drop it. It wasn't an issue of not studying enough. The concepts were just too abstract and no matter how much time I put in, I couldn't make them make sense in my head. On the other hand, for some of my friends (who also happened to be IMO gold medal winners), the class was totally intuitive. It just made sense to them, but as much as they tried, they couldn't break it down for me in a way that I could understand. It's a humbling experience to experience at a deeply visceral level, exactly HOW much smarter than you other people are. More people should have that experience. Maybe we all will with AI. Long tails can be very long.

@FranckPachot The fact that ClusterVolumeIOPs is actually “IO/300 seconds” will never fail to infuriate me. If I wanted a 5 minute average, I’d ask for that.

Surprising Write IOPS in Amazon Aurora dev.to/franckpachot/1…

@fasterthanlime Wait until he learns about utf8mb4.

@jeremycole Do you have any specific gripes against it? Genuinely curious; I know it changed, but I have no idea about any cons it brings.

@JamesRLandrum @DhravyaShah Sure, but you said “few ms.” 54 is far beyond any definition of “few” I’ve ever heard. Agreed that sites don’t necessarily have to have multiple round trips, but unless you can fit everything into initcwnd – 14.6 KB by default – there are at least two, plus handshake.

@StephanGarland @DhravyaShah 54ms is nothing. Slightly north of 3 frames of a 60fps display. Half to 1/3rd the time of the average blink. The issue is making serving a multiple trip process when it doesn’t have to be. To assert that every file or packet also has a 54ms delay between each is flat out wrong.

people don't talk about the latency issues with Hetzner enough. For just a PING, IT TAKES LIKE 100-200ms i think that's just crazy, but maybe also because i'm in US and their servers are in Germany

@JamesRLandrum @DhravyaShah Even if data traveled at the speed of light (it doesn’t), and there was 0 overhead from processing (there isn’t), a round-trip 5000 mile distance would take 54 msec. Very interested to see your math.

@DhravyaShah Latency != speed. You can talk to servers 5,000 miles away fast enough to deliver every single bit of content they need in a few ms. 300ms is still faster than most websites out there. The tech for this is 10+ years old and nobody uses it because CDNs are “easier”

@vvsevolodovich @rmoff There is nothing wrong with this stack in the slightest. I would take Django over literally every other ORM. MySQL is more performant than Postgres if you know how to work with it, rather than against it (also, it takes effectively zero ongoing maintenance, unlike PG).

@rmoff Wow, yelp gives an impression of is 2011: MySQL, ampq, django

How Yelp migrated with zero downtime and no data loss from PostgreSQL to MySQL engineeringblog.yelp.com/2024/10/migrat…

@joshmanders @samlambert And network: people always forget that EC2s have network bandwidth limits which are confusing as hell, and it absolutely is possible to saturate the smaller – medium ones. That’s not something really anyone would expect to happen, and yet. Surprise!

@joshmanders @samlambert K8s on cloud (which is most) brings other challenges that people might not be aware of: network and disk performance limits. If you use the default StorageClass, last I checked it’s gp3. Maybe that’s good enough, but maybe for bigger DBs it isn’t (at least, not with defaults). /2

we run petabytes of state on kubernetes. i guess we should talk about it more?

@m1ru1 @ChShersh @_atlj I don’t know how to say this other than your average Haskell dev is ridiculously more skilled than your average JS dev. I am neither, to be clear. Haskell is such a niche language that if you learn it, you must be innately curious and interested in computers. That pays dividends

@ChShersh @_atlj But why? Haskell developers are easier to get? Existing staffs easier to be trained?

I worked at startup to rewrite a Node.js backend to Haskell. Main reasons for choosing Haskell were: 1. CTO really liked Haskell. 2. Huge companies offered higher salaries, so all Node.js developers preferred to work there, and that startup had difficulties with hiring.

@blbraner @dhh Do you think people are running their own DCs? There is a very small handful of companies that do that. Rent rack space in a colo. Done.

@dhh I’d be curious to see a post about how you build in power and internet redundancy. It’s one thing to admin your own servers, it’s another to have fault tolerance with utilities you don’t control, which the cloud provides.

There's this weird aversion in some startup circles to saving money. Like you're admitting weakness if you take a few days to setup some servers to save big money because OPPORTUNITY COST. An internalized performance show to prove unicorn-rocket-ship potential to investors.

@CoderciseYT @kenwheeler

@kenwheeler "Oh no we don't do our own auth, sessions, database, servers, queues, cron or uploads. We focus on the _business logic_ and telling you that 1+1 = 2"

backend devs entire job is recommending 3rd party saas, and then they say front end isn’t real programming

I simultaneously want to and do not want to find out why large loop performance (and apparently other things) has been steadily declining in Python since 3.11. I also have a sinking feeling that it's not going to be fixed, and I'll be stuck at 3.11. github.com/python/cpython…

@yacineMTB False. I read it offline to better absorb the information, and then go try out some of the concepts I want to explore more, referencing the book as necessary. Also, black-on-white is flaring, and inverse looks odd when there’s just a huge expanse of space. Paper reads better.

if you were actually someone who read this book you wouldn't own it, you'd have a PDF because reading technical information on paperback and not a screen is dumb

@ClassicGamerTWR @thdxr I've seen a mid-sized MySQL instance happily running with 1000 connections. I can't say the same about Postgres. Don't take this as a slam on Postgres; I like it plenty, it's just worse at connection handling. Which is why a pooler is practically required.

@ClassicGamerTWR @thdxr No RDBMS like it, sure, but again my only assertion is that Postgres is uniquely bad at it. There's a reason that the RDS default for `max_connections` parameter works out to 12/ram_size_mib for MySQL, and 9/ram_size_mib per Postgres. /1

i do not understand how postgres is so popular when it is using the connection per process model in 2024