Oh yeah, it's a long story!

My core idea is to build a task manager that will stay with me for the rest of my life. Because of that, my main requirement is for it to be fast even with a massive database. If I have 10k+ tasks saved over the years, it should still load and feel instant.

Another requirement is that it must be offline-first. I live in a country where the internet goes down pretty often, and I need my tasks to be available regardless of the server status.

Finally, I wanted a clean API so I could connect things like an MCP server or create tasks via Telegram.

I couldn't find an existing app that met all these needs, here is a table where I compared all self-hosted apps that I found(in awesome self-hosted github):

Here is how my journey went:

1. First approach: I tried using Redux and MobX. Performance sucked. The main reason is that MobX/Redux don't support B-tree indexing. Local-first apps rely heavily on "intervals"(cause you use fractional-indexing-jittered to order things and to support LWW per column), and that’s exactly where B-trees shine. Without them, it just got slower as the data grew.
2. Second approach: I tried a reactive UI based on persistent SQLite queries (using wa-sqlite). It worked well for small lists, but it felt sluggish once the dataset got large.
3. Third approach (Current): I decided to build my own DB solution, which you can see here: HyperDB. The trick is that you write the logic once, but on the frontend, it runs against an in-memory B-tree index (which is incredibly fast) and data still stored at persistent wa-sqlite layer, while on the backend, it runs against persistent SQLite (which keeps memory usage low).

This third approach finally solved all three requirements. It’s hard to find an open-source app that does this because this specific architecture is difficult to "cook" correctly. On top of the database speed, I also had to solve the sync problem, that should also somehow resolve conflicts while clients are offline, which I handled by building own sync layer with LWW CRDT per column.