Sketch Lab

The problem

Backend engineers have all heard of Bloom filters, Count-Min sketches and HyperLogLog. Few feel confident using one. These structures answer questions about huge data in a small, fixed amount of memory: a database skipping a disk read, a CDN counting unique visitors, a stream processor finding its most frequent items. The papers are dense, and most blog posts flatten the one thing that matters. The error is one-sided and can be calculated in advance. A Bloom filter's "no" is certain. Its "yes" is only likely.

Nobody adopts an approximate structure they can't check. I built Sketch Lab for two groups: engineers who need to size one for a production system, and engineers who need to see it work before they trust it.

What I built

The exact answer runs beside the approximate one

Every stream goes through the sketch and through an ordinary exact set or map at the same time. The page flags each false positive the moment it happens, naming the item and the slots it collided with. The main control is a memory slider. Drag memory down and two curves move together: the error the formula predicts, and the error measured from the stream.

Memory is measured in real bytes

The comparison with the exact structures only means something if the numbers are true. So every memory figure is the actual size of the allocated arrays, never a formula presented as a measurement. A calculator takes the number of items and a target error, and returns the sizing, the formula and the memory in real bytes. That screen is what makes it a tool someone can use at work.

I wrote the hash and the structures myself

All four structures come from their original papers: Bloom (1970), Count-Min (2005), HyperLogLog (2007) and a Cuckoo filter that supports deletion. I wrote the hash function too. Every error guarantee assumes a good hash, and a bad one breaks them silently. The hash tests run first in the build pipeline, and nothing else runs if they fail.

Guarantees and probabilities are tested apart

"A Bloom filter never says no to an item it holds" is a rule, and I test it with zero tolerance. "The error rate stays within the predicted bound" is a probability, tested with fixed seeds, many trials and a written-down tolerance. Keeping them apart stops anyone from adding slack to a rule that has none. Every run is seeded, so the same input gives the same result on any machine.

When a test failed, I fixed the model

One test failed at about 2,500 times the textbook error rate. I traced it to a structural effect of the hashing method when a filter has far more memory than it needs. I added the missing term to the model and left the tolerance alone. A later check showed this Cuckoo filter uses more real memory than a Bloom filter at every error rate. So the page offers it for deletion, and makes no claim about saving space.

Result

Sketch Lab is live, public, in English and Indonesian, and works offline after the first load. It covers four structures plus the exact reference, merge checks for combining sketches, and uniform, skewed, adversarial and text streams. Every setting can be shared as a link. Every milestone in the plan shipped, including the optional Cuckoo filter. I built it in one day, on 6 August 2026, with three runtime dependencies. It's a teaching instrument, not a production library, and that was the plan from the start.