Anyone tempted to write a custom database should ask if they're prepared to be in the database maintenance business.
Because that's where they're heading, for a really long time.
this post was submitted on 05 Apr 2024
88 points (95.8% liked)
Programming
17408 readers
68 users here now
Welcome to the main community in programming.dev! Feel free to post anything relating to programming here!
Cross posting is strongly encouraged in the instance. If you feel your post or another person's post makes sense in another community cross post into it.
Hope you enjoy the instance!
Rules
Rules
- Follow the programming.dev instance rules
- Keep content related to programming in some way
- If you're posting long videos try to add in some form of tldr for those who don't want to watch videos
Wormhole
Follow the wormhole through a path of communities !webdev@programming.dev
founded 1 year ago
MODERATORS
we’ve replaced our $10k/month Aurora instances with a $200/month Elastic Block Storage (EBS) volume.
Lol, i’m sure just refactoring it and using a different service could have lowered that price significantly.
The description of their data says “relational database” to me, but don’t murder me for using such an unpopular phrase!
I had the same thought. Like, I think Aurora is one of the most expensive ways to do this in AWS. But, since this particular set of data is so well-defined, and unlikely to change, roll your own is maybe not crazy. The transactions per second and size don't seem that huge to me, so as things grow I imagine they can revisit this.
But, since this particular set of data is so well-defined, and unlikely to change, roll your own is maybe not crazy.
I think that's the trick here. A relational database lets you do a whole bunch of complex operations on all sorts of data. That flexibility doesn't come for free - financially nor performance-wise! Given:
- engineering chops
- a firm idea of the type of data
- a firm idea of the possible operations you may want to do with that data
then there's a whole range of different approaches to take. The "just use Postgresql" guideline makes sense for most CRUD web systems out there. And there are architecture astronauts who will push stuff because they can, not because they should.
Every now and then it's nice to think about what exactly is needed and be able to build that. That's engineering after all!
They never would have been able to get the same performance from any solution that incorporates a general purpose database.
Their requirements/explicitly-not-required-ments include that it's fine to drop 1s of data. That would be an insane proposition for any other database. Also their read/write rates and latency requirements are unusual to say the least.
It's the same thing as tiger beetle. Ridiculously narrow domains allow for ridiculous performance improvements compared to of-the-shelf solutions.
What is tiger beetle?
A new database specifically designed for financial transactions.
I'm not an expert on finance software, so I can't critically assert how good they really are. But they claim much much higher throughput than traditional databases, higher fault tolerance, self healing networks if several replicas are running, etc.
From a purely technical standpoint it's interesting for being written in zig. Because the database scope is so narrow they know exactly how much memory they will need on startup and just allocate all required memory on startup and never allocate more, nor free the aquired memory.
I'm really excited to see what forks of tiger Beatle for other domains look like. They, supposedly, built it to be able to modify the state machine to other data schemes, but that code mostly just made realize I had no idea what I was looking at.
As soon as someone makes a KV on it, I'm tried to have be my ETCD database
Aurora is a relational database
Cool
For anyone seeking to write their own database, I have one recommendation: the Tokyo Cabinet Library.
Tokyo Cabinet abstracts away all the needs of writing your own serializers and deserilizers for binary formats. You can have hashtable databases , B+ trees and everything else all prepared for you under one roof.
Of course that is if you have brain, and don't use a text storage format like JSON. If you use shit like JSON and YAML
and add potentially hundreds of millisceonds of parsing time just to serialize data from text into machine-readable binary, then please submit your name and address to my council so we can get rid of you when we own the world.
Cleansing of the undesirables aside (seriously, give me ONE good thing about text storage formats! They are EXCHANGE formats, not STORAGE formats!), Tokyo Cabinet is written in C so you can easily bind it with SWIG. But there's probably bindings around if you look.
Never heard of this. Thanks for mentioning it!
Comparing cost to AWS Aurora is unfair. Give us the self host price, and compare to that.
Also, they should have tried Scylla or Cassandra. It's very scalable and handles a lot of writes.
Why does everyone on the planet hate sql nowadays?
If your not dealing with billions of transactions why not?
Maybe it reminds them of Oracle.
What has changed though, is that we’ve replaced our $10k/month Aurora instances with a $200/month Elastic Block Storage (EBS) volume.
Holy shit, I hope whoever wrote this gets a fat bonus at the end of the year. That's a truly astounding savings.
What they don’t say is how much in developer time they’ve spent rolling their own database. Then there’s also maintenance and new features.
It's called job security, bro.
I'm sure their custom database will be easy to find people to support and maintain
It seems like features would be mainly additional key/values like temperature, humidity, etc. This wouldn't really change the underlying infrastructure greatly but still give good enhancements to certain customers.
Aren't time series databases like Prometheus pretty good at storing this kind of data?
I don't believe Prometheus supports geospatial data. Two minutes of googling though, so I could be wrong.
They're just storing doubles in their own format too. I'm not sure if they even need any spatial lookups on the data, they didn't mention anything about that in the article. Maybe they do that in-memory?
I imagine a delta encoding scheme similar to what the time series DBs use would work well for the geo points. Maybe even a delta-of-delta encoding for things like ships which move very consistently.
It's probably not worth it given how small they've already go their data. But it is fun.
If you’re using Java you can use eclipsestore. Seems like a good project