raw mode?

[ronag]

In order to get key + value we use entries which are an array of [key,val]. However, allocating an array for each key value pair is slow. Would be nice if we could somehow return a flattened array with key = xs[idx * 2 + 0], val = xs[idx * 2 + 1] and avoid O(n) array allocations.

[vweevers]

That makes nextv() slower because that must then "unflatten" the array in JS land. The current approach makes nextv() fast and next() only slightly slower (compared to a flattened array which we used previously in leveldown).

[ronag]

That makes nextv() slower because that must then "unflatten" the array in JS land.

I was thinking that the flat array could go all the way to the user in js land, i.e. no need to unflatten. This would of course need a new api. Would help a bit in our use cases.

[vweevers]

Related: are you working with binary data? If so, I've been thinking about an entirely different approach, where we allocate one buffer and slice it up in JS. But that would not improve perf if you're working with string data (which includes the json encoding).

[ronag]

I'm currently working with json encoding. It's on my roadmap to eventually move to string and then Buffer but that adds significantly complexity so currently I'm trying to find simpler avenues of optimisation.

[ronag]

Ok, now I have 'string' encoding but the optimization is still just as relevant.

[ronag]

Another alternative would be to get keys and values in separate arrays? i.e. iterator.all((err, keys, values) => {})