I quickly ruled out using database/sql drivers as the indirection through interface types added a bunch of overhead and stymied my attempts for reasonable memory layout. For my use-case, I found the crawshaw driver performed the best, but I ended up forking it as well as the Golang standard library CSV parser as I found defensive copying & allocation was the largest bottleneck. I ended up cycling several very large arenas among a CSV parser thread that filled the arena with column bytes and several threads writing to different temporary sqlite databases. Then at the end I ATTACHED them together and copied them into one big file (idk exactly why this is faster, but my profiles showed most cpu time spent in sqlite doing query binding things so MOAR CORES).

One notable optimization was exposing a way to bind borrowed bytes to query parameters without inducing a copy in either Golang caller code, or SQLite library code. The crawshaw driver upstream only exposes sqlite_bind_blob with SQLITE_TRANSIENT mode, which tells SQLite to copy the input to a private allocation before returning from the sqlite_bind* call. I added a version that passes SQLITE_STATIC, which means "trust me, I won't touch these bytes until the query is done, and I'll free them afterwards". This is safe in Rust who's "borrow" and "lifetime" concept models this perfectly, but I guess in Golang its dicey enough to not expose in your public package.

Here's the relevant commit in my fork: https://github.com/crawshaw/sqlite/commit/82ad4f03528e8fdc6a...

I'm curious how OP's https://github.com/cvilsmeier/sqinn would fare, I'm somewhat sus about copying 200GB to stdin but the benchmark results are pretty good so ¯\_(ツ)_/¯