perf: dynamic dispatch overhead

[PSeitz]

This flamegraph is an term query aggregation on 10mio documents selected via a "*" query.
Only around 60% is spent in the aggregation, a large chunk seems to be calling overhead, e.g the dynamic dispatch on the callback in for_each_docset.

pub(crate) fn for_each_docset<T: DocSet + ?Sized>(docset: &mut T, callback: &mut dyn FnMut(DocId)) {
    let mut doc = docset.doc();
    while doc != TERMINATED {
        callback(doc);
        doc = docset.advance();
    }
}

[fulmicoton]

considering you do buffering in the aggregation layer, should we add some buffering solution in at the DocSet level maybe? can you experiment?

[PSeitz]

Yes, I think we have two options

• Buffering like in aggregations. It probably needs to be on the Collector, but could be combined with DocSet
• Remove the callback dynamic dispatch (maybe not feasible)

Buffering may have some free positive side effects on the generated code, but the (sometimes unused) score may complicate things a bit

[PSeitz]

#1937