Notices by yopp (alex@feed.yopp.me)

@GossiTheDog it would be great to add note that you don’t have to upgrade right now.

Disabling zlib for network compression in enough to mitigate

Either:

a) Restart mongod/mongos with option --networkMessageCompressors=snappy,zstd
(omit zstd on 3.6 and 4.0)

b) Disable in mongod.conf and restart

And then plan upgrade after holidays

@ryanc thank you for your interest. we appreciate time and effort you put in sending this email, but after careful consideration we decided not to move forward with delivery of your message at this time

@whitequark @fasterthanlime oh, nice! If it’s not a trade secret much they charge for ASN allocation and what’s their yearly fees?

@whitequark as people said you can get it as individual from RIPE. I’ve got mine from https://lagrange.cloud/products/lir

You can work around home address, but AFAIK your legal name is still going to be public, if it matters for you

And yeah, next you’ll want to get your v4 subnet. You can’t announce smaller than /28 and it cost fucking €10k on a second hand market :(

@whitequark Stanford be like “heck, can we spread compilation over virtually unlimited aws lambda calls” and make gg

http://stanford.edu/~sadjad/gg-paper.pdf

I’m not sure if there maintained tools based on this paper. llama seems dead https://github.com/nelhage/llama

@whitequark looks like minio support http tracing:

https://github.com/minio/minio-go/blob/master/api.go#L118

https://pkg.go.dev/net/http/httptrace

And it’s got all the goodies!

@whitequark til: native histograms in prometheus! (I’m in Victoria* camp)

I see you already measuring `git_pages_s3_get_object_duration_seconds`. This one should show a better picture latency wise, but still a function of data size.

If you are concerned about s3 storage performance connection time & ttfb for GetObject is most useful here I think. And maybe average object download speed (size / (time - ttfb))

Conn time more like “what’s going on with network”, ttfb more like “how fast tigris moves data to closest pop” and speed is speed :)

Anyways, if you need volunteer on observability front, I’m up for that! I love digging numbers to find meaningful metrics :)

@whitequark and as you deal by with static content, you might benefit from widening buckets a bit, something like:

.005, .01, .025, .05, .075, .1, .25, .5, .75, 1, 2, 5, 10, 30, 60, 120, 300

This gives a bit more coverage over small latencies and wider visibility over large latencies

@whitequark you can try adding another histogram here: https://codeberg.org/git-pages/git-pages/src/branch/main/src/observe.go#L209 to collect timings per function name so you can see how s3 actually behaves without client related noise

@whitequark aha, so there is little to no git-pages response buffering and those numbers include time spent sending bytes to client

Then this metrics are fine, for exactly stated reason: most users have shitty internet connection.

I’m not sure if caddy has something like nginx proxy buffering.

nginx can be configured to read full response from app sever into cache and so app can move to processing next request while nginx is finish sending response to slow clients

If caddy can’t do this, then it’s a bit tricky to setup observably.

Maybe tffb to s3 is more meaningful metrics in this case

@whitequark Did some quick poking and yeah, my bet is on “dynamic placement optimisation” being shit

I see consistent 2-10x difference between first and consequential queries. But not as terrible as on your screenshots, more like 40 vs 150ms

I wonder how latencies look like grouped by server. Is this git-pages or caddy metrics?

@whitequark hard to tell without info on how request time is measured, but overall doesn’t look good

Can you trace detailed timings of s3 http query: dns, connect, tls, transfer? On top of that try to poke bucket with s3cmd and measure latencies from single host. Try to fetch same object at least 3 times and watch if latency drops after first GET. If so, it’s most likely related to pull of object from storage to POP in your region. If so, then GET of same object in different region will have same pattern. If latency is not consistent over long run then most likely they don’t have enough capacity to do “free egress” reliably

@whitequark @ignaloidas around black friday there will be tons of cheap offers on lowendtalk.com.

Like 5$ per *year* for IPv6-only and something like 15$ for v4

But yeah, hosting is dirt cheap. I have 5 ancient xeon dedies that cost me about 60€/mo total

@whitequark it would be awesome if you’ll do a short write up on how you did that.

Also, who’s you are renting /32 from? Is transit expensive?

You: push to prod on Friday, ruining your weekend

Me: upgrade self-hosted stack on Friday, ruining my weekend

We are not the same

@whitequark @david_chisnall @janl and to make jsx/tsx components you have to setup said components and all the code around it is kinda boilerplate. Hardly any IDE have enough snippets to automate that.

But I totally agree that with LLM you trade speed for attention, especially if it’s non statically typed code so your IDE can’t catch bullshit on the spot.

I’m not defending LLMs: they are overhyped and they don’t live to their promise of the utility. For me it’s somewhat time saving in some scenarios, but overall it’s meh. I don’t believe that current architecture can achieve anything other than ruining the knowledge storages we had before

@david_chisnall @whitequark @janl depends on your definition of boilerplate I guess.

Say I have event driven system and I want isolate each even type in its own “container” (class, module, whatever). Consequently there will be a lot of structural similarity between “containers”, because they will expose same interface.

Some languages are verbose by design like HTML. If you are building a form you’ll have to repeat same things over and over again, and it’s… okay? Overall all UI/graphics code is very verbose because you have to setup a lot of things. And when you take component approach you’ll end up with case above.

LLM autocomplete predictions are just more context aware than plain autocomplete (intellisense or whatever) and can save you some time on typing, because it can just spit out pre-filled method call, with all arguments filled with values from the current context.

No wonder that thingie made to auto-complete based on a context does decent job as context-aware auto-complete!

@whitequark @janl in auto-complete mode, no surpries, it is somewhat usable, especially when it’s a lot of boilerplate. But as “coding agent” it’s, uh, /suboptimal/

@whitequark interesting! where is conceptual ground for you in this? I see this more like being given a tool that I can decide to use or not. akin to dressing the wound or knitting the sweater

@whitequark yeah, it’s my previous country as well. at that time there were like 4 psychiatrists in the whole country who treated adult adhd. I had my share of unuseful and even harmful solutions and trauma along the way

But do you see another way around?