I chatted with Richard Newman about this a few months ago, and he thinks the current implementation (an async call from JS to Rust that retrieves the pairs eagerly on a background thread, then synchronous iteration in JS of the retrieved pairs) is actually likely to be the optimal one, as holding onto an LMDB transaction across many async iterations will be harder to reason about, may cause memory and disk growth, and could end up being less performant/responsive overall anyway.

So I'll close this bug. But we can reopen (or, more likely, file anew) if we discover issues with the current implementation that warrant revisiting the idea.