I had an interesting time investigating a bug related to SQLite performance today, so I thought I'd share a thread.

A PicoShare user reported that it took 14.57 seconds to load their list of files.

The first thing I noticed was that they had uploaded a 1.24 GB file. The files I use with PicoShare tend to be images and documents in the 5-50 MB range, so 1.24 GB is larger than I've tested with.

One of the wacky architectural choices I made with PicoShare was to store all file data in SQLite as BLOB data. This is an unusual choice, as most apps store file uploads on the filesystem. But putting it in SQLite means I get cloud replication "for free" with @litestreamio

Storing file data in SQLite leads to odd behavior like this, so I suspected the bottleneck was somewhere in my SQLite logic.

Here's the code to retrieve file metadata. I keep one table for file metadata and another table for file data, broken into 32 KB chunks.

I suspected that the problem was on the SUM(LENGTH(chunk)) line because it's the only one that's touching the file data.

I was able to reproduce the issue locally by uploading a 1.1 GB file. The page load time jumped to 9.3 seconds.

I tried removing the LENGTH() function and just hardcoding the size calculation to 1.

It worked! Page load time dropped to 8ms.

I was confident that the SUM(LENGTH(chunk)) line was causing the latency.

Does calling LENGTH cause SQLite to read the full BLOB data into memory? That would certainly explain why it took 9 seconds to calculate the length of a 1.1 GB file.

I checked the SQLite docs. They didn't explicitly say that LENGTH reads the full blob data, but it suggested for strings, it calculated length on the fly by looking for the first null byte. I'm assuming for BLOB types, SQLite iterates through the full contents of the column

I had specifically avoided storing the file size in my SQLite DB because it was redundant. The raw data is there, so we can always derive the size. It shouldn't be something we store independently.