Under the surface of a database.

Learn the system by building its parts.

A database's public API hides considerable work. Writes need a path to persistent storage; reads must find the current value across several layers; old versions eventually need to be reclaimed.

This personal engineering project examines those mechanisms directly through an embedded key-value implementation.

Keep writes ordered, then merge.

Incoming changes enter a write-ahead log and an in-memory table. The storage path creates immutable sorted tables on disk, while the read path checks the relevant memory and disk structures.

Bloom filters help avoid unnecessary table lookups. Compaction brings older tables together and reconciles versions.

Memory, log, and immutable tables.

The implementation is written in C# on .NET 8. Its main mechanisms include batched WAL writes, memory tables, sorted-table storage, Bloom filters, and background compaction.

The diagram describes the source's storage flow. It is an explanation of the engine, rather than a product interface.

Work at the seams between layers.

The engineering interest lies in how the components coordinate, especially when a key has multiple versions across memory and disk.

Write-ahead log

Record changes in batches before they move through the storage layers.

Read filtering

Use Bloom filters to reduce unnecessary sorted-table reads.

Compaction

Merge stored tables and reconcile older values and deletion records.

A buffered write is a deliberate boundary.

The WAL batches writes by count and time. A call can return before its batch is flushed, so an acknowledged write is not a guarantee of crash-safe persistence.

This experiment makes the durability boundary explicit: buffering groups the log writes, while persistence depends on a completed flush.

An inspectable storage implementation.

The public repository contains the engine, examples, and checks for key-value operations. It offers a concrete way to study the data path and its tradeoffs, with the implementation available to read and run.