Name the transaction before comparing TPS
The official pgbench reference checked on 4 November 2026 distinguishes the default write-containing transaction, simple-update and select-only scripts. simple-update omits teller and branch updates from the default sequence; select-only issues only the SELECT. A higher number from a read-only run cannot be presented as equivalent write-transaction capacity.
Use the same script on each candidate and record whether it represents the application at all. A custom script counts one execution as a transaction, so even two custom-script TPS figures may describe very different amounts of work. Include query mix, transaction boundaries and durability settings in the report instead of naming only the tool.
Initialize only a disposable benchmark database
pgbench initialization creates its standard tables and can destroy existing tables with those names. Create a dedicated disposable database with no business data, confirm its identity and permissions, and decide the cleanup procedure before initializing. A benchmark tutorial is not a reason to run initialization against the application database.
The documented default scale factor of 1 creates 100,000 account rows; scale 100 creates 10,000,000. Dataset size affects what may fit in memory. Record scale, actual storage size and warm-up policy so that an in-cache run is not mistaken for evidence about a much larger working set. Keep benchmark data and write volume within the provider’s permitted resource use.
Distinguish maximum throughput from offered demand
By default, pgbench runs transactions as fast as possible. Its rate option, -R, instead targets a specified transaction arrival rate with a Poisson-distributed schedule. That asks a different question: can the system keep up with this offered workload rather than how much work can it complete when driven without that arrival limit?
Choose a bounded duration and a small set of increasing rates appropriate to your own application. Record clients, worker threads and where the load generator runs. A remote generator includes a network path; an undersized local generator competes for the same VPS resources. Neither should be silently compared with a run using a different layout.
Count the work that never reached the database
The latency-limit option, -L, reports transactions exceeding the specified milliseconds as late. With -R, transactions already too far behind schedule are not sent and are reported as skipped. These are distinct from database failures. Under throttling, reported latency includes schedule lag rather than starting only when the transaction actually begins.
A useful acceptance record therefore includes offered rate, completed work, late and skipped counts, failures and retries, not only average latency or TPS. An apparently healthy completed subset can coexist with substantial skipped demand. Do not change the limit between candidates and then rank their latency as if they met the same deadline.
Turn a test into a buying decision
Save versions, parameters, duration and output together, and compare repeated bounded runs with the same policy. If the application is slow while pgbench is healthy, inspect its actual queries, locks and connection behavior before concluding that more vCPUs solve the issue. This method supplements the PostgreSQL sizing guide; it does not supply measured results for catalog providers.
For a concrete resource reference, the catalog’s Dallas vhp-2c-4gb-intel variant lists 2 vCPUs, 4096 MB RAM and 100 GB disk, with an observed 4 November 2026 USD 24 one-month quote. Tax and renewal are unknown; optional backups and extras are excluded. Neither that tariff nor its disk label predicts transaction throughput. Use the configuration catalog to shortlist, then test only an authorized disposable environment.