Log-structured KV: readrandom at cache speed, write-amp at ~1x
You are looking at a 100M-key Bitcoin address & tx keyspace100M Bitcoin keys on disk in one Semurg substrate. Ask it to readrandom, scatter point GETs across the whole keyspace and it runs the
real query live on this box, on data you can
download and re-run yourself.
Click an example in the agent rail, or ask again to re-run readrandom, or re-pull the keyspace over Bitcoin P2P. Every run is measured end to end, on
this box, on the real dataset. Nothing is pre-computed.
The readrandom test: 100M Bitcoin keys (addresses + transactions) on disk, point GETs scattered across the whole keyspace. We serve each as a seqlock read off the RAM-ring fanned across every core, with no compaction-built tree in the read path. And because we are append-only, we write each byte about once instead of rewriting it through every compaction level.
the log-structured leader
~9.5x WRITE AMPLIFICATION. A leveled log-structured tree rewrites the same data through every compaction level, roughly 9.5 bytes written per byte ingested on a typical 7-level config, with the write-stalls and SSD wear that come with it.
semurg, one engine
~1x write-amp. Append-only, each byte written about once, no compaction tree to climb on the read path, and readrandom is a seqlock RAM-ring hit with every value durable on disk.
~9.5x less disk written. We do not out-compact compaction, we remove the need for it.
readrandom over 100M Bitcoin keys on disk. Measured live on this box. Re-pull the keyspace over Bitcoin P2P and re-run it yourself.
ask the agent to readrandom, the live throughput and write-amp render here
Two honest claims. READS win on raw speed: readrandom is a seqlock RAM-ring hit with the data durable underneath, and there is no compaction-built tree to descend. WRITES win by COLLAPSE, not raw speed: append-only writes each byte about once (~1x), where a leveled log-structured tree rewrites the same data through every level it compacts (~9.5x), so we remove the compaction tax rather than out-running it. Scope we decline: a sorted tree still wins ordered RANGE scans, we are a point store on this path. Durability we match: every key is on disk before ack, no weaker fsync mode assumed. Total time, no asterisk.