Kenji Murakami spent eight years at a Tokyo-based payment processor, where the core transaction ledger ran on a relational database that was never designed for the write volumes the business eventually reached. By 2017, the team was patching around consistency gaps with application-level locking, retry loops, and a reconciliation job that ran every four hours to catch the anomalies that slipped through. The patches worked, mostly. But every quarter brought a new edge case, and the engineering cost of maintaining the workarounds was growing faster than the product itself.
01 Strict serializability enforced at the storage layer, not configurable away
In the spring of 2019, Kenji left to build the engine he had wanted to buy but could not find. The design started with two non-negotiable properties: strict serializability and synchronous replication. Everything else, including throughput, was a constraint to be optimized within those bounds. The first internal benchmark, run on a four-node cluster in a rented rack in Osaka, showed 12,000 transactions per second with a median commit latency of 1.1 ms. That was slower than the target, but the consistency guarantees held under every failure scenario the team could construct. Over the following two years, the storage engine was rewritten twice and the replication protocol was revised after a careful reading of the Viewstamped Replication Revisited paper. The current version sustains 80,000 transactions per second at 0.4 ms median commit latency.
02 Synchronous replication to two regions before any commit acknowledgement
Grit Core Lab incorporated formally in Okayama in 2019 and opened its first managed cloud region in Osaka in early 2021, followed by Tokyo later that year. The team is small by design: seven engineers, two of whom focus exclusively on correctness testing and fault injection. The test suite runs 14,000 scenarios nightly, including network partition simulations, disk-failure injections, and clock-skew tests. If a scenario causes a consistency violation, the build does not ship. That constraint has held since the first public release.
03 14,000 nightly fault-injection scenarios; a consistency failure…
04 WAL is fsync-confirmed before the client receives a response
Kenji Murakami is the founder and lead engineer of Grit Core Lab. He spent eight years at a Tokyo payment processor, where he maintained the transaction ledger and eventually led the database infrastructure team. Before that, he contributed to the open-source Paxos implementation used in a distributed key-value store maintained by a university research group in Kyoto. He holds a degree in information engineering from Osaka University. Outside of work, he runs a small reading group in Okayama that works through distributed systems papers one chapter at a time.