Programming Language &
Database Benchmark Test

Five databases, twelve core programming languages, and C# in three database rounds. The October 10, 2026 campaign tested concurrency levels from 1 through 1,000 in both limited- and full-resource configurations.

The goal is not to pick a favorite language first and design a test around it. We run the test, preserve the output, and let the measured results change our assumptions when they need to.

Current Test Status

Five-database benchmark campaign completed October 10, 2026

PostgreSQL, MariaDB, Cassandra, YugabyteDB and MongoDB each completed their scheduled language-by-language campaign sequence in both resource modes. The source file records 710 passing measurements, 144 failing measurements, and four preflight failures. The four preflight failures were Zig with PostgreSQL and Zig with YugabyteDB, in both modes.

Of 882 planned concurrency measurements, 854 were actually performed. The four preflight failures prevented 28 measurements from running. None of these results have been repaired, discarded, or replaced.

Limited resources: API 2 CPUs / 2 GB RAM; database 4 CPUs / 8 GB RAM. Full resources: API and database each 12 CPUs / 90 GB RAM. Each concurrency level was measured for approximately 30 seconds. APIs ran one at a time against the selected database.

These tables load the recorded values directly from the October 10 raw benchmark summary (TSV). Older comparison rounds remain available in the October 9 historical page.

Download Raw Benchmark Results

PROJECT DECISIONS · OCTOBER 7–10, 2026

Changes to the Benchmark Plan

The benchmark is intended to help choose the right technology for a particular job, not to force a preferred result. Earlier decisions and unsuccessful measurements are documented rather than concealed.

Database change: CockroachDB → YugabyteDB

An earlier CockroachDB v25.3.3 preliminary load test encountered a licensing-related limit on concurrent open transactions. Those preliminary runs were not comparable to isolated official runs.

No license installed. The maximum number of concurrently open transactions has been reached.

The team selected YugabyteDB as its replacement. YugabyteDB is now included in the October 10 campaign, with both limited- and full-resource tables below.

Language change: Haskell withdrawn from newer rounds

Haskell participated in earlier PostgreSQL and Cassandra benchmarks, which remain in the historical archive. Its development and dependency setup burden led to its removal from the newer campaign.

This is an assessment of the project's practical maintenance experience, not evidence that Haskell is inherently slow or unsuitable. C# was included for MariaDB, YugabyteDB and MongoDB in the new campaign, not for Cassandra or PostgreSQL.

Current database plan

  • PostgreSQL: October 10 limited and full rounds recorded.
  • Cassandra: October 10 limited and full rounds recorded.
  • MariaDB: October 10 limited and full rounds recorded.
  • YugabyteDB: October 10 limited and full rounds recorded.
  • MongoDB: October 10 limited and full rounds recorded.
  • CockroachDB: excluded from this five-database campaign following the earlier licensing limit.

How to read the results: Each table shows the October 10 campaign only. The older October 6–9 measurements are retained intact on the historical page, not mixed into the October 10 tables. A failed benchmark still shows its recorded RPS and errors; a preflight failure shows that the concurrency test did not run.

Benchmark Results

Each result cell shows recorded requests per second followed by the number of errors, using the original benchmark summary file. Failed levels remain visible. A preflight failure is not a zero-throughput result.

Click a concurrency column heading to sort the table by that concurrency level. Click it again to reverse the order. Select requests per second or errors to choose the sorting measure.

The tables are populated by JavaScript directly from summary.tsv on this server. No server-side programming language or database is required. Sorting changes only the order displayed in your browser.

Compare Results by Programming Language

Choose a language to see its results across all five databases and both resource configurations, from C1 through C1000. This uses the same October 10 raw benchmark data as the tables below.

Loading results...

Each cell shows requests per second, errors, and any FAIL status. Hover over a measurement for the request count. C# was tested with MariaDB, YugabyteDB and MongoDB only. Preflight failures are shown as not run; historical results are not mixed into this comparison. Click a column heading to sort.

Loading results...

Loading results...

Loading results...

Loading results...

Loading results...

Loading results...

Loading results...

Loading results...

Loading results...

Loading results...

What We Measure

Requests per second are only one part of the result. We also record errors, average response time, p50, p95, p99 and maximum response time in the detailed run logs. These summary tables display RPS and errors; the source TSV also includes the number of requests and pass/fail status.

Limited-resource and full-resource rounds are kept separate so that efficiency under constraints and the ability to use additional resources can be considered independently.

The saved campaign summary contains 858 rows: 710 PASS, 144 FAIL and four PREFLIGHT_FAILED entries. One failing benchmark recorded an unknown error count, and it is reported as unknown rather than changed to zero. The 28 skipped levels are explicitly marked as not measured.