Question

A single-node OLTP engine's commit path is fsync-bound: every committing transaction calls fsync on the WAL to be durable, and under 50k tx/sec the disk can't sustain one fsync per commit — throughput collapses and commit latency spikes. You must raise commit throughput without weakening the durability guarantee (a transaction reported committed must survive a power loss). Design the WAL write/commit path: how you batch durability work across concurrent committers, how you keep latency bounded, and the knobs that trade latency for throughput.