the library
everything we measured.
How kiban is built, how it was tested, and what it lost along the way, in the engine's own reports. 20 documents, 55k words, at commit fd1b95c.
start here
What kiban is, how it is built, and what changed.
how it stays correct
The failures found, the repairs made, and how they were proved.
the repairs.
Three scale-validation failures, their root causes and the fixes.
9 min readthe audit.
Concrete findings from an engine audit, and how each was closed.
6 min readtesting the write path.
An independent QA pass on group commit, with one medium finding.
12 min readhow we measure.
Workloads, cache treatment, fault injection and how to reproduce them.
8 min read
head to head
Kiban against the engines people would otherwise use.
against the field.
Four engines at one commit: the campaign behind this site's numbers.
5 min readversus tuned RocksDB.
A fair fight: 425 commands, screened settings, and where RocksDB wins.
23 min readwhy not just use SQLite?
Both reasonable schemas, tuned, on an agent-memory workload.
23 min readHelix and LMDB.
In-process, with matched durability, at 10k and 1M records.
34 min read
how it got fast
One bottleneck at a time, including what was rejected.
one sync for many writers.
How group commit took durable writes from 718 to 5,379 per second.
4 min readwhere a read spends its time.
A ranked cycle profile of the read path, and what it pointed at.
24 min readreworking the read path.
Faster scans, and an optional range tier ported from HelixDB.
7 min readthe perfect hash.
The experiment that showed immutable tables could beat a B+tree.
11 min readmaking the index cheap.
PH3 cuts locator bytes by 61% and frees the path to a default.
11 min readlatest ten, faster.
A scan extension that beats LMDB in every measured cell.
3 min readreads on a small cache.
Point reads when the data is ten times larger than the cache.
39 min read
from the archive
Earlier work, kept because the context still matters.