250x faster iterator seek

[peakji]

Sorry for the scary title, but I'm not kidding! 🤓

Benchmarks first (on my MacBook @ 1.2 GHz Intel Core M):

SCAN: 5147.625ms
SKIP: 9171.527ms

SCAN_NEW: 5143.952ms
SKIP_NEW: 36.186ms

@kesla s amazing #96 brought super fast read streams, but I see some modules using something like:

function fuzzyGet (gt, lt){
	db.createReadStream({
		gt: gt,
		lt: lt,
		limit: 1
	}).on(...)
}

The buffering mechanism is wasting IO in this use case.

I want to keep the good parts of both worlds, so I've added a new bool landed to iterator, and let it ignore highWaterMark for once after seeking:

Now we can do really really fast skip scan, sorted set intersection and K-way merging, with multiple seeks from #323.

This change is transparent to levelup and users, and will not affect normal iterating without iterator.seek.

[peakji]

Heres the benchmark script, please ignore the nasty semicolons (the pride and joy of a C programmer) 😜

const leveldown = require("/Users/PeakJi/Gits/leveldown");
const DBPATH = __dirname + "/tempdb";

var targets = [];
var db = leveldown(DBPATH);
db.open(function(err) {
	var batch = db.batch();
	for (var i = 0; i < 0xFFFFF; i++) {
		var buf = Buffer.allocUnsafe(4);
		buf.writeUInt32BE(i, 0, true);
		if (i > 0 && i % 1000 == 0)
			targets.push(buf);
		batch.put(buf, buf);
	}
	batch.write(function(err) {
		console.time("SKIP");
		skip(function(err) {
			console.timeEnd("SKIP");
			db.close(function(err) {
				leveldown.destroy(DBPATH, function(err) {

				});
			});
		});
	});
});

function scan(end) {
	var ite = db.iterator();
	ite.next(function iterate(err, key, val) {
		if (key)
			ite.next(iterate);
		else {
			ite.end(function(err) {
				end();
			});
		}
	});
}

function skip(end) {
	var ite = db.iterator();
	var run = function(i) {
		if (i == targets.length) {
			ite.end(function(err) {
				end();
			});
			return;
		}
		ite.seek(targets[i]);
		ite.next(function(err, key, val) {
			run(i + 1);
		});
	};
	run(0);
}

[juliangruber]

impressive improvement!

[ralphtheninja]

Now we can do really really fast skip scan, sorted set intersection and K-way merging, with multiple seeks from #323.

What's K-way merging?

[peakji]

@ralphtheninja Its actually called K-Way Merge Algorithms or K-way merge sort, useful in external sorting and inverted index.

[dominictarr]

wow, great work @peakji

[ralphtheninja]

@peakji I'm just going through the code again. If I understand correctly, the first .next() you do after a .seek() will ignore highWaterMark and not buffer the results in order to optimize .seek(), .next(), .seek(), .next() etc, but once you do .seek(), .next(), .next(), .next() etc then you will buffer. Correct?

[peakji]

@ralphtheninja 💯👌