papa kappa

Databases are tricky. There are no silver bullets, despite what marketing folks would like you to believe. ACID is an acronym where 2 letters don't matter. NOSQL was a thing, got over the hype, and is now just as boring as any other wave of technology.

Also if you learn enough about databases you'll shed a tear if you know it processes people's financial data. Everything can go wrong. Nothing works the way you would hope it would. At least not if you want enough throughput for it to be feasible. The present is grim dark.

We can't possibly start to cover all varieties of databases, and their tradeoffs. Instead what we'll do is pull a lever, fast forward into the future, and talk about a particular type of database (and architecture) that has us excited. Because yay, fringe development!

The Status Quo

In traditional databases you have a database process running on a machine somewhere. The database "speaks" a wire protocol, and servers process requests from clients and in turn persist them to the database.

The traditional architecture can at times be tricky. If you add more servers to talk to the database, the database might start having trouble handling requests. Migrating your database to a larger, and larger machine will only ever take you so far; so usually some form of horizontal scaling is applied.

Horizontal scaling in this sense means adding more machines to the problem, rather than upgrading machines to be more powerful. Adding machines can take many forms; and is often used in conjunction with partitioning the database, and/or adding caches.

Traditional databases come with their own fair share of other problems. Here's a fun few:


note: for brevity's sake we're going to gloss over replication modes, and the tradeoffs of multi-leader systems, concensus protocols and all that. Let's assume that that the things we're going to be discussing next can be replicated fairly easily.

The kappa architecture is a bit different. Instead of storing a computed value, we store the raw data in a log. And using specialized processors, we parse the data into an index (also known as materialized view).

You can think of a log as an Array, and an index as an Object (or key-value store). Each entry in the log contains some metadata — the time it was created, some form of sequence number. There's a field of studies on how to create non-conflicting entry types called CRDTs.

var log = []
var index = {} // Although this can be any value, key-value is the most common

You can think of using a log database as inverting the database. They come with a good set of benefits: data resolution will never get lost because you store the raw data. All that's needed to perform a database migration is to create a new index on top of the log (which can't really be considered a migration).

And finally: scaling is quite convenient. Because indexes use a log as the source of truth, they're much more reliable than traditional caches to scale data.

If you want to dig deeper into the topic, check out this video which goes in-depth into the kappa architecture.

Exercise 1

Alright, let's create an HTTP server that accepts a request, and stores data in an array.

var http = require('http')

http.createServer(function (req, res) {
  // persist data here to a log

Exercise 2

Cool, we should now have a bunch of numbers stored in a log. It's time to create a summary of all the data.

var http = require('http')

http.createServer(function (req, res) {
  if (req.url === '/summary') {
    // provide a summary
  } else if (req.url === '/add') {
    // add data

Exercise 3

Alright, neat! We're getting close to something useful. Now let's persist our log to disk on each write. Let's use .csv.

var path = require('path')
var fs = require('fs')

fs.write(path.join(__dirname, 'file.csv'), data, function (err) {
  // handle err

Wrapping Up

And that's it. You now have a working log database on your machine. There's lots of optimizations that would need to be done for it to work in production, not the least thinking about replication, preventing conflicts, persisting, using effient on-disk data structures, partial-replication and integrity checks. It's a lot; but they all operate on top of the same basics shown off here. We hope it was useful!

See Also

Edit on GitHub Status: offline