lun-4 (imagine gpt-3 but gayer)

a robot girl does writing and shit. reach at email [email protected]

A bit of history

In the current LavaTech infrastructure, we maintain a single application server that runs Proxmox VE, and from there we slice it up containers for most of our applications, with some (like the ejabberd instance powering having their own dedicated servers.

Proxmox VE provides both LXC containers and KVM/QEMU virtual machines so we can use specific operating systems. One of the VMs provides FreeBSD CI infrastructure for the Zig project. The architecture for it is composed of two VMs: one that runs the Sourcehut instance on Alpine Linux, and another running FreeBSD for experimentation─such as compiling LLVM if needed─and also to update the Sourcehut instance's FreeBSD CI image.

That was the first proper contact with FreeBSD I had, lol.


Since then, I have been experimenting about with NetBSD and FreeBSD, and one of the latest experiments was installing FreeBSD over a serial line.

There wasn't a production need to do it, as we don't have a project that needs thousands of FreeBSD VMs, but the serial line idea helps us because we wouldn't need to go the whole way through Proxmox VE's web admin panel to get a VNC session. In theory we should be able to just SSH to the host and manage that VM directly, without needing SSH inside the actual VM.

The experiment was succesful and here are my findings.


Proxmox provides a way to attach to a VM that has a serial line, all documented here, so the bit that I had to dig through was FreeBSD's own interaction with the serial line.

I found a random blog post that helped in explaining the general idea of FreeBSD-install-in-a-serial-line, but it isn't directly applicable because this process uses USB installation media, and that's something that I did not want to research, because our mental model to creating VMs describes creating them with ISO files/CD/DVD, and Proxmox doesn't quite provide an obvious button for that.

From that finding, we know that we would need to create our own ISO file containing the serial line settings, then I found the installation manuals for FreeBSD 7.4 that talked about using the serial line for installation, all here, strangely I couldn't locate that same documentation on 12.1-RELEASE.

Actual installation

After creating a modified ISO file with the process from the FreeBSD manuals (and from that blogpost, as I don't know if comconsole_speed="115200" is a default, I assumed not), it was straightforward to add it to the Proxmox VM image library, and boot up the VM. After doing qm terminal ... I got the bootloader messages, the bootloader screen, and was able to have it process my input to boot (it would automatically boot either way, but pressing a button and seeing things happen is a good feedback mechanism).

FreeBSD bootloader screen on serial line

While it was booting though, we have found an anomaly.

I typed cd9660:/dev/cd0, hoping it would work, because I didn't understand the issue that was happening (Ave cited that maybe something related to ISO repacking was doing something bad here, but I didn't investigate it further), sure enough, it worked, and I got to a selection of which terminal am I on so that the installer can draw its boxes!

I recommend you to use xterm when possible, because ansi and vt100 can be hard to use and lead you to type wrong things in the wrong places.

Drawbacks compared to doing VGA

Serial line can get clunky, especially if you try to just tmux(1) inside of it (I was a reboot away from being unable to boot due to a broken rc.conf, as vi drew things on the terminal in the wrong ways), but it is enough to curl my SSH keys to a user, and use it inside a proper terminal.

I'm glad you asked (you did, I promise) because I can infodump for an hour.

hatsune miku votes too!


Every 2 years, Brazil runs elections. One for the local elections, where someone votes for the city Mayor and the city's representative, and other where someone votes for the President, State Senator, State governor, and State deputies.

Local elections happened in November 15th 2020, so what better day to start writing a blogpost describing how it works than today.

This is going to be a technical explanation of how things work. So I'm skipping many pre-election and post-election processes. I'm going from candidate loading until vote counting.

The courts

In Brazil, we have the TSE, Tribunal Superior Eleitoral, it translates directly to the Supreme Electoral Court. It is the practical creator and maintainer of the voting systems used around the country. After TSE comes the TREs, Tribunal Regional Eleitoral, translates to Regional Electoral Court, and are the courts that actually collect the data from the voting machines back to TSE (more about that in the future of this post).

Pre-election things

Every Brazillian citizen between 18-70 years of age must vote. Citizens above 16 can vote. Not voting when you must vote means that you must pay a fine to TSE (if the fine is not paid, there are other things as well, such as being unable to emit a Passport, being unable to be admitted to a public university, etc). To be able to vote, a citizen must create their Voter Card (my translation to Título Eleitoral), and as of recent years, the Voter Card creation process involves taking a picture of the voter and a record of their fingerprints (more on that later).

In the creation process, the citizen is asked which place would they want to have their voting take place. I don't know the criteria, but places can become unavailable as they might be full of voters. After your selections, you are assigned an Electoral Zone and an Electoral Session. The hierarchy goes as follows: State –> Location (city, suburb) –> Zone (number) –> Session (number). Schools are the most common place to be Electoral Zones.

After candidates for the respective year's positions are selected, they are loaded in blank voting machines. The loading process is done via a specific kind of USB flash drive with a specific plastic shell to guide it into the back of the machine (preventing people from trying to jam it in, then switch sides, then attempt again, the usual USB problem). After that is done, another USB flash drive (of same structure) is loaded in, and the spot it's on is locked, as that flash drive will contain votes.

Electoral Session pre-election-day ceremonies

The election days are selected by TSE, composed of a 1st and 2nd round. 2nd rounds happen when a candidate for a major position (such as mayor, president, or governor) does not reach a majority (51%+) in a location with more than 200 thousand voters. 2nd rounds just have the 2 leading candidates from the 1st round.

On election day, the mesários (which are voluntary workers for TSE/TRE to manage the election) assigned for a specific Electoral Session, with relevant managers, do the starting ceremonies for the voting machine the respective TRE gave them. That ceremony happens before the actual hours of the election day (election day happens from 07:00 (7AM) until 17:00 (5PM)), so between 6-7AM.

The setup ceremonies involve testing the machine's hardware and making it print a report that it does not contain any votes inside. That report is called the Zerésima. You can see examples of them on the TSE website.

After those ceremonies, the session is ready to accept voters.

Incoming voter process

When a new voter comes in the voting session, they must have some form of identification of thsemselves and their “voter selves”. For the former, that can be done either via TSE's digital voter card app (recommended during the pandemic), the E-Título (only works for voters that registered a picture of themselves) or any government-issued ID document (passport, national ID card, driver's license). For the latter, either E-Título or their voter card works.

Upon showing their identification, one of the Mesários checks the identification provided on the session's book. That is a physical paper notebook possibly contains a picture of a voter, their name, and their voter card number. When a match is found, the voter number is told to the Mesário operating the Terminal. Depending if the voter had their fingerprint registered or not, the Terminal requires the voter to put that in its reader. (This bit may be disabled entirely, considering COVID. As an example, I was not asked to put my fingerprint in for 2020's elections, even though I did in 2018's)

The Terminal controls the voting machine in terms of “starting a voting session”, “ending the session”, “stopping a vote if the voter didn't finish their vote”, “requesting accessibility options”, etc.

After voter identification, the voting machine is now ready to receive the vote.

Voting machine architecture (physical)

The voting machines are all manufactured by TSE and TSE contractors and sent out to TREs around the country in the months leading to the election.

Here's a picture of it:

brazillian voting machine

You can use a little simulation TSE made using highly advanced JavaScript (note: fully portuguese)

Voting machines do not have any networking equipment (Wi-Fi, Ethernet, Bluetooth, etc), its only form of communication is with the Mesário Terminal, and a power cable. Voting machines have batteries as well, to handle remote locations.

Voting machine software architecture (high-level)

The voting machine runs Linux with a highly customized userland made by TSE.

Inside the voting machine there is the Registro Digital do Voto (RDV, Digital Vote Registry), that file can be thought of as a spreadsheet where the columns are the political positions, each row is a voter, and each cell is a vote for that political position. The vote, when casted by the voter, is randomized amongst the other votes, emulating a physical ballot box being shuffled around. Portuguese high-level explanation by TSE.

After a single vote

After the voter has cast their vote, they're asked to sign the notebook and the Mesário gives them a piece of paper, taken from the notebook itself, which is the proof that voter actually voted. The voter can now leave the session, and the next one comes in. This is the cycle that happens until 5PM.

Electoral session post-election-day ceremonies

At 17:00 (5PM), all voting sessions are due to stop soon. If there are any remaining voters in the queue, they must be attended (which means the session might stop after 5PM), but no new voters will be attended (so the actual place closes its doors, but the sessions inside are still open).

After all voters went through, the Mesários issue the command to make the voting machine stop. When that happens, it prints out the total amount of votes that were casted on that session, as in, “candidate X got N votes”, “candidate Y got M votes”, etc. This report is called “Boletim de Urna”, and I'll keep its translation as Vote Report.

The Vote Report is generated from the RDV file cited earlier, it is printed by the voting machine up to 11 times. The first copy is a check to see if the printer of the machine is operational, then 5 required copies, then 5 optional copies that may be requested by political party representatives, the media, or the Ministério Público. Other reports printed are the Justificative Report, and the Mesário Identification Report.

All 5 required copies of the Vote Report and the Justificative Report are signed by the President of the Voting Session, the Mesários, and other Inspectors. Mesários must sign the Mesário Identification Report.

The Vote Report MUST be publically available, either through physical means, where one of the copies is put on the voting session itself, or through digital means, on the TSE's website. The Vote Report contains a QR code which means you can cross-reference the tally made by the voting machine with what the TSE received.

Another copy is given to one of the Inspectors in the session, or the media, or a representative from Ministério Público. Remaining reports are sent back to TREs.

After all of those are finished, the voting machine displays that the session is over, and can now be turned off by the President of the voting session.

Vote Transfer

The Vote Report, while it is sent to TREs, is not the way votes are processed. The voting machine writes all of the information in its Vote Report on the locked-up USB flash drive. The seal and lock are undone, with the President for the session's supervision, and all flash drives for the Zone are sent physically back to TRE.

On the TRE, those flash drives are loaded for final counting, their data is sent back to TSE via a separate network (network is provided by Embratel), and from there, results are distributed via online/TV to everyone in the country. Here's the website for the 2020 Local Elections.

Before actually starting this I want to say that programming language discourse is 99% of the time based on subjective experience. The writing here is my subjective experience dealing with Python. Python may be good for you, or bad for you, or whatever. Replacements beware.

Time changes, people change

I've started learning Python after attempting to learn C but failing back in approximately 2013. But I would only consider writing Python for internet-facing production in 2017.

This rant has hindsight bias, for sure. But a good point on Python is that I didn't quite knew how to program things before learning it. Sure, I did write massive if chains with C but nothing more complex because as soon as I started reading the pointers chapter I would get confused.

Nowadays I'm not that confused with C pointers, or the difference between the stack and the heap (thanks Zig), and I'm a different person, with different goals and needs on what I want my software to be, or, what its principles are. But, even though I'm a different person, I'm still locked in to the younger me with Python, and now that I can put “5 years of experience writing Python” on my CV, I don't think this will go away any time soon.

A non-exhaustive list of things that keep me sad while writing Python

I could go on and enumerate things I'm not happy with and that I ultimately don't have power to fix, so here we go:

  • Typing is sad, it's both a mix of giving static type checking but without actually breaking the language. I mean, yes, typing is good, but in the end you can declare a function that takes an int and give it a string and you'll only know something is wrong at execution time if you don't have mypy, and so you go to install mypy, and then it doesn't find some library, or a library doesn't have typehints, or it decides to never want typehints, etc, etc, etc.
    • A subproblem of typing is using things like typing.Protocol. Which I tried to use, once. It type passed afterwards, but I feel dirty after writing something like that. I wish I could elaborate further on why, other than that general description of “something's wrong” while looking at it.
  • Packaging. God. Packaging. After going through requirements.txt,, poetry, dephell, pipenv, I just gave up and settled on since it's the only guaranteed thing that might work.
  • Not being happy with the “large” frameworks AKA Django, meaning I have to go to things like Flask, but then there's asyncio, so I'm going to Sanic, then sanic is it's own hellish thing, and I'm now comfortable on Quart. And by using a niche of a niche of a niche web framework, I'm having to create my own patterns on, for example, how to hold a database connection by just putting app.db = ... on @app.before_serving, you can be sure mypy does not catch anything about that at all. Love it.
  • A Bug in a completely separate part of my software was caused by a bump in dependencies but not bumping my python version on CI. I kept banging my head for hours until I gave up and updated from 3.7 to 3.9.
  • Endless language constructs seem to be the norm now. I don't like that we now have 4 different ways to merge dictionaries together, because the existing one didn't have “a good syntax”. That feels like a really low bar for “new” language constructs.
    • This is new on 3.9, so its likely the two last entries are more of rants than critiques, idk.

The future?

I don't know. None of those things look like will ever be fixed. I'm seeing if I can write webshit with Zig, but that's all experimental. Might take much, much longer to “ship” but I feel that by not having the ecosystem or an interpreter slowing me down I'd have a better development experience.

I can't deny that I'm now with years-long experience with Python and I have to put something in my curriculum, though.

The tech industry makes me sad.

Or, “Monolith First, Complicated Thing Later, Also Plan Before Switching, And How Those Things Have Been Known For Ages, HELP”

I wan't to start this with Gall's law, from General Systemantics:

A complex system that works is invariably found to have evolved from a simple system that worked. The inverse proposition also appears to be true: A complex system designed from scratch never works and cannot be made to work. You have to start over, beginning with a working simple system.

When I see large scale distributed systems being written with Kubernetes or whatever, I have the gut feeling Gall's law wasn't followed, though my sample size is very small: 1.5 systems. 0.5 comes in because the other system I'm managing isn't Kubernetes (yet, I'm sure it'll become that after a while).

I consider that jumping from a monolith line of thought to a microservice line of thought, for any team or system, is a very large jump that must be scrutinized before actually doing it. Microservices are no silver bullet, mostly because developers are taught on writing monoliths for their whole life, and then you push up microservices, which come with a lot of specific tooling, and as years pass, even more tooling is created, to the point enumerating them all can be absolute sensory overload. Really.

My Bad Kubernetes Experience

In one of the systems I was developing on, I wasn't actually messing with Kubernetes directly, but, regardless, I had a very poor experience integrating with it. The monolith app I developed was inside a VPS, and had to push data to MongoDB and Redis nodes which were living inside the Kubernetes cluster. The person that managed it told us to use kubectl port-forward, and sure, we tested it out, and it worked, at least at the start, but we all know how it goes from here.

While the system was operating at spike load, the cluster restarted itself, and kubectl port-forward lost connectivity with it. In my mental model, it would reconnect to the cluster: it has authentication, it doesn't need to hold any state. Yet it didn't, and our app kept unable to connect to Redis (wish a very cryptic ECONNRESET in the app logs) for a long time while we attempted to debug it, a restart on the systemd unit that did the port-forward “fixed” it, after we talked with the person operating the cluster how redis was, and they told it was up.

After talking to a friend about that situation (a long time later), they told me port-forward is “a broken ssh tunnel that might not even work”, and “it's not meant to be used for anything other than testing”. So I could have used some other Kubernetes solution which wouldn't have made me write three paragraphs.

My Bad Microservices Experience

I think I can say this one is caused by bad management wanting to put “novel” things on teams that are very new, don't know anything about what a container actually is, or don't know what distributed system design actually entails to the failure modes of such system.

When you jump into microserivces, and design with them in mind, you have to know that you're designing a distributed system, and while the units composing that system are small and simple and understandable, the whole can be much, much more complicated (see: natural systems in the whole world, inculding human-made ones!). As well as you must know all the “fallacies of distributed systems”, which can become somewhat cliché, but are actually quite true, and at this rate, if I'm developing a distributed system, I should print out the list and put it next to me and every other developer, I've seen them happen, even though the fallacies have been enumerated since... 1997? It's weird how we're still having issues that come from those fallacies in 2020. We should do better.

I think this boils down to “people should learn a little bit before jumping in the microservice hype”, and in most cases, Gall's law still applies. I believe that things should be written first as a monolith, to have the developers learn the business logic in an environment that function calls aren't remote everywhere. And if scaling is required, then proceed to investigate a distributed architecture, if you can actually scale there, or do something else.

Offloading to the database

Another data point from overall system design is that you shouldn't offload everything to the database, as in, make every microservice need it for any operation, because as things like autoscaling jump in and now you have 10 times the amount of microservices, your database will suffer 10 times the load. That might catch fire.

Thouh if the database does catch on fire AND is a managed database solution, you can blame the cloud company providing it instead of yourself (hrrrrm Discord loved to do this on their postmortems, but nowadays they don't even give any postmortem. It's sad).

Disclaimer: I do not work for Discord, or with Discool. If i wanted to look cool I would call myself an “independent security researcher”, but in the end I just stare at Chromium's DevTools while Discord is open for an hour.

What are Intents?

Discord has unveiled Gateway Intents a while ago. A PR on the Discord API Docs repository was created to help library developers start adding support for Intents, and provides some answers to questions developers raised up.

The general idea of Intents is that instead of having all events being thrown away at your gateway connection, only the events you actually want are sent, this gives huge wins to general network usage. Most bots get a ton of typing events while being unable to do anything about that, for example. There was a stop-gap measure for this via the “guild_subscriptions” field, but the general idea is that Intents is the way to go, as it's more generic.

From all intents, two of them are considered “Privileged”: members and presence, providing the member list and user status (online/offline/what game they're playing) respectively.

I'm writing this on October 3rd, 2020 (with many drafts afterward). I did have a lot of time in my hands to study about Intents but I'm only doing it right now and sharing the results since I had to port my personal bot to it.

What's the actual deal/reasoning with Privileged Intents?

From the original blogpost:

We believe that whitelisting access to certain information at scale, as well as requiring verification to reach that scale, will be a big positive step towards combating bad actors and continuing to uphold the privacy and safety of Discord users.

To use Privileged Intents, your bot: – must not be in 100 guilds, or – if you are in more than 100 guilds, you must be a “verified developer” with a “verified bot”. If you attempt to bring your bot down to less than 100 guilds after that, you can't have the intents back

The verification process involves filling a form describing your bot's functionality, and give a government-issued identification document via Stripe Identity. Depending on your country, your passport may be the only available option. You can check that information on the Stripe Identity documentation.

There is a grace period before bots are forced to be verified: October 7th, 2020. Bots beyond 100 guilds past that date will not be able to join any more guilds.

From the blogpost, and the surrounding timing on it, I can safely say the creation of Privileged Intents is to provide an answer to what I dub the “discool discourse”. They haven't said that publicly, and I can only base this on feeling alone.

Discool was a service that stored a lot of Discord user metadata. You could put a User ID in, and get the list of guilds they were in (of course, not all guilds, but a lot of public guilds had their data scraped).

The way they did it was by creating burner user accounts able to join those guilds, scraping all the data already available to users (the Discord client needs the member list after all), and store them in a database. Give enough time and effort, and you have a pretty big privacy scandal regarding user data and Discord.

From my perspective, Privileged Intents were made to make developers that create a large-scale privacy scandal legally binded to Discord, and go to court, and etc, etc, etc.

Are the Privileged Intents a wrong way to protect data?

I think so, or at least, very flawed. As I said, Discool uses user accounts, they never used bots in the first place, any Intent business would never get to those bad actors. Of course, Discord has been implementing various kinds of user extra-verification with “machine learning”, but the point is that anyone with a browser can start up Selenium, point it to Discord, and start scraping. Hell, given enough time and effort someone could gcore the browser process to scrape data while still being classified as having “human sentient behavior”

Even in the case of an actual privacy scandal on Discord, the damage for it, even though you're limited to 100 guilds, even though it's technically less users it's the same level of data, you can still fuck up the lives of some people, but I think Discord only cares about having less of the numbers.

When giving data out to anyone, you should assume the worst, because you don't control their intent (pun unintended). Sure, requiring ID and so providing the existing legal system as a protection against misuse of user data works with more than 0% efficacy, but I don't think that is a good solution.

In regards to giving out my own data, I'm not sure I can trust Discord in handling of that. I lost my trust on that after they explained that their data collection (the famous /api/science route, that was formerly /api/tracking, but got blocked by extensions to the point they had to rename it), if turned off by the client, would still continue (you can prove this with devtools), and in their FAQ post, just said that it's dropped on the server.

The nitty gritty: when you turn the flag off, the events are sent, but we tell our servers to not store these events. They're dropped immediately — they're not stored or processed at all. The reason that we chose to do it this way is so that when you turn it off on your desktop app it also turns off automatically on your phone – and vice-versa. This allows us to keep things the same across all of our apps and clients, across upgrades.

Believe me when I say, since they already sync user configuration across clients, it's easy to make the app disable tracking upon seeing the already existing configuration option to disable tracking.

Learning to bypass limitations imposed by Privileged Intents

For context: in the library that I'm using for my bot,, there's a full framework, discord.ext.commands, in it you can declare a command like such:

from discord.ext import commands

async def some_command(ctx, argument1, argument2):

One specific feature of the framework is the ability to convert from raw strings to useful objects, for example, if I wanted to receive a user, I could say

async def some_command(ctx, user: discord.User):

and will convert either a mention, or an ID, or the username back to the User object. This is possible because Discord gives all member data to the bot upon the bot's startup. It is fully dependent on the members privileged intent.

Without the members intent, the member lists are basically empty, and the feature doesn't work anymore.

Bringing the feature back

Discord's message events contain full user objects (which have ID, name, avatar, discriminator, etc) and member objects, (which contain roles, nickname, etc) since the client must be able to draw out the UI containing that user.

Even though you don't have full member lists on startup, you, using that fact, can start passively collecting data to re-create the feature. It works like this: on any message, store the user in persistent storage, when needed, look up from that storage, and maybe put it in cache for performance reasons.

Considering message events also contain the guild's ID, you can safely say that with any message that contains an author and a guild ID, you can say that author is a member of that guild, and still do the same thing Discool thing. It is technically less data than originally, since you're only collecting data about the people that are actually sending messages in the guild.

Keep in mind that messages being sent also include the system messages created upon a user's join (they contain the same user object). They're messages like any other, and will reach any bot in the guild. You can have the same effect with bots that create welcome messages to new users, as they usually put a mention to the user (“welcome @user to $guild!”), and since mentions internally are represented by putting the user ID on the message, bots can still scrape it up.

You can also use that same method to extract relationships with typing events as well, as they contain guild ID and user ID

Future work

New Discord features

Recently I've seen work-in-progress inside the Discord client on a way for bots to declare their own commands (you can also see the Figma prototypes on the “The Future of Bots” blogpost), removing the need for bots to have every single message being sent to them (it currently works with bots checking in if the message starts with a given prefix, etc).

It is defnitely a better idea than the current state of things, but I'm not sure if the boundary of “hey, your user data is now in the hands of the bot developer” is very well defined. Many users would still use the bot, even once, and have their user info stored on a database. I'm not using it for malicious intent, but if we're talking about the technical purpose of Privileged Intents, we must assume every bot is malicious, and design around that. using Request Guild Members for the User and Member converters

See this issue. It would sidestep the problem of having to extract info out of messages, which would help my usecase immensely.

I don't know if it'll be implemented, and I'm too anxious to open an issue on to use it, but it is a nice idea.

Closing thoughts

this music is fucking awesome

I tried, as an experiment/curious endeavour, to run osu!lazer (2020.717.0) on a NetBSD 9.0 QEMU/KVM guest, on my machine. Here's the results:

  • Follow the Linux Emulation NetBSD guides, installing the linux compatibility packages (suse_*) was good enough to continue.
  • FUSE is not included by default, but thankfully, the AppImage gives instructions to extracting itself into a folder.
  • An AppRun file exists in the folder. Trying to execute it would give “Failed to initialize CoreCLR, HRESULT: 0x80070008”. It is related to ulimit -n, the default is way lower than 2048
  • Raise the file descriptor hard limit using sysctl kern.maxfiles, then bring up the hard limit (ulimit -Hn), then bring up the soft limit (ulimit -Sn). If I recall correctly, I bumped it up to 8000.
  • Some other failure happened while running AppRun, upon further inspection via ktrace/kdump, a sched_setaffinity syscall failed because of permissions.
  • Running it as root fixed that.
  • A traceback appeared regarding ICU libraries not being found, even though I have them installed.
  • You can edit osu!.runtimeconfig.json to disable that requirement. More here
  • osuTK kept talking about unsupported platforms. Upon further investigation, it was because no valid GPUs were found. A very cryptic error, for sure.
  • In the end and some more discussion, we blamed the fact I was running NetBSD as a KVM guest as the culprit. virt-manager only provides QXL, VGA, and virtio, and virtio gpu drivers aren't on NetBSD 9.0.

Maybe I'll run NetBSD as a host on one of my machines and keep this experiment running. Until then, that's what I got.

To be continued..?


beep boop

welcome to writing moment

  • lunlunlunlunluna