Before signing a hardware PO, run a standardized benchmark that mirrors your transaction mix. This guide uses sysbench oltp_read_write as a stand-in for TPC-C style OLTP.
Cluster prep
- 3 OBServer nodes, 64 vCPU, 256 GB RAM each.
- OBProxy on separate VMs (2× for HA).
- Disable unrelated tenants; single benchmark tenant with 32 log streams.
sysbench command
sysbench oltp_read_write \
--mysql-host=obproxy.bench.internal \
--mysql-port=2883 \
--mysql-user=root@bench \
--tables=32 --table-size=5000000 \
--threads=128 --time=600 \
--report-interval=10 run
Record
| Metric | Target |
|---|---|
| TPS | baseline for regression |
| p95 latency | < 40ms |
| CPU per node | < 75% |
Scale threads until TPS plateaus. That is your per-cluster ceiling before adding nodes.