The Blog
Notes on databases, distributed systems, and the engineering habits that keep software dependable in production.
## Why Migrations Break A migration that runs in 200ms on your local machine takes 30 minutes on the production table with 50 million rows. During those 30 minutes, your `ALTER TABLE` holds a lock t...
## Why Caching Exists A single database query takes 5ms. A cache lookup takes 1ms. The difference seems small — but at 10,000 requests per second, that 4ms gap is 40 seconds of cumulative latency pe...
## 1. What TCP Actually Is, and Why It Exists Start with the network underneath TCP, because that's the whole reason TCP exists. The internet moves data in packets, and the protocol that handles tha...
## Where We Are In Part 1 we built a Write-Ahead Log. Every mutation — write or delete — is recorded to disk and fsynced before we acknowledge it to the caller. We have durability. What we do not h...
## Where We Are Post 1 built the WAL — every write is fsynced to disk before being acknowledged. Post 2 built the memtable — an in-memory skip list that keeps keys sorted and serves reads in O(log n...
## Implementing a Write-Ahead Log in Go A Write-Ahead Log is the mechanism behind the D in ACID. Before any mutation reaches your data files, a record describing that mutation is written and fsynced...
## Introduction You have three backend servers. Each one can handle 1,000 requests per second. Together they can handle 3,000 — but only if requests are spread evenly across all three. Something ha...
## Where We Left Off In Part 1 we built a reverse proxy that distributes requests across three backends using round robin. It works — but it is blind. Kill one of the backends and the load balancer ...
## Introduction Your API receives a request. You need to send a confirmation email, resize an image, call a third-party webhook, and update an analytics counter. You could do all of that synchronou...
## Introduction Your application writes a row. PostgreSQL says "committed." The server loses power one millisecond later. When it comes back, is your row there? The answer is yes — and the reason ...
## Introduction Your application has one user. Concurrency is not a problem. That user reads data, writes data, nothing conflicts with anything. Then you have ten thousand users. Two of them update...
## Introduction Your single database server handled your first 100,000 users just fine. Then you hit a million. Queries slowed down. Disk filled up. You threw more RAM at it. It helped — for a while...
## Why B+ Trees Exist Your database has 50 million rows. You query by user ID. Without an index, the database reads every row until it finds yours. That's a full table scan. It gets slower with ever...
## Why Single-Database Transactions Don't Scale A transaction in a single database is simple. You start a transaction, make changes, commit. Either everything commits or nothing does. The database h...
## Why Replication Exists Your database will crash. Not maybe. It will. Hard drives fail, memory corrupts, someone runs the wrong command, power goes out. When that happens, you need another copy of...
## Introduction Most developers understand SQL. Fewer understand what happens after the query planner finishes — how rows are physically laid out on disk, why certain writes cost more than others, w...
## Introduction At some point you ship something that works perfectly in staging. The code is clean. The tests pass. You deploy it and feel good about it. Six months later you're debugging a produc...
## Why This Matters in Production Every backend engineer has a story about a production incident that made ACID properties real. Maybe it was a race condition in a payment system that took hours to ...
## The Real Cost - What happens when you don't have idempotency Every backend engineer has experienced this: a user clicks a button, the request times out, they panic and click it again. Now somethi...
## Introduction Your database has 50 million users. Someone searches by email. Without an index, the database reads every single row until it finds a match. That's a full table scan — and it gets sl...