Varaq 0.1.0 · Benchmarks
Every number, with its method.
Measured on a 4-vCPU cloud VM, small on purpose: a machine that size keeps the engine’s own costs in view. Runs on large Linux machines with dedicated NVMe drives come next.
How these were measured
- Machine
- A Firecracker VM: Linux 6.18, 4 vCPUs, 15 GiB of memory, one virtio block device, on a shared host. Every run was made in a Claude Code cloud session.
- Storage
- Every store is a buffered file: no
O_DIRECTand noio_uring. The log is written but not forced, so these numbers do not include making commits durable. Every store fits in its pool. - crud-bench
- 1,000,000 rows, 4 clients, a 4 GiB data pool and 4 GiB of frames for old versions, three rounds, every run successful. Each column is the median over the rounds; the dots on a bar are the rounds.
- TPC-C
- Varaq’s own driver, a fresh load for every row, three rounds a row; every round passed TPC-C 3.3.2’s consistency checks. No keying or think time, so these are not comparable to audited results. An aborted transaction lost a first-writer-wins race and is not retried.
- Lookup
- One thread, the tree alone without a transaction, 20,000 keys with 1,300-byte values in memory. Three runs of nine batches of two million lookups; the number is the median of the three runs’ medians. The record holds the program and its full output.
The announcement’s Measurements section gives each number its source and its limits.