Benchmarks

Altair is compared against Express (Node.js), Fiber (Go) and a Ruby full-stack MVC app on identical PostgreSQL-backed CRUD endpoints, driven by k6 over real HTTP. The harness lives in the repository at examples/benchmark_k6 and can be re-run locally.

Every framework gets the same 20-connection / 6-worker budget at the realistic profile — 100 virtual users for 60 seconds — tuned on the benchmark machine; a discarded k6 warm-up run precedes each measured phase. PostgreSQL is pinned to 6 CPUs / 2 GB — enough headroom that throughput is framework-limited, so per-request cost is what is measured.

Results — unified profile (100 VUs, 60 s, 20 connections, 6 workers)#

Workload Framework req/s avg p95 p99 p99.9 max
Write POST /items Go Fiber 20,300 4.7 ms 7.3 ms 9.8 ms 16.3 ms 36.8 ms
Crystal Altair (gcry) 13,981 7.0 ms 11.9 ms 13.8 ms 17.0 ms 25.2 ms
Node.js Express 8,074 12.1 ms 18.1 ms 23.2 ms 33.1 ms 65.5 ms
Ruby Ruby full-stack 2,213 44.5 ms 376.6 ms 407.0 ms 471.2 ms 1,071.6 ms
Read GET /items/:id Go Fiber 18,990 5.1 ms 8.2 ms 12.1 ms 18.9 ms 31.4 ms
Crystal Altair (gcry) 13,619 7.2 ms 13.0 ms 16.6 ms 25.9 ms 37.2 ms
Node.js Express 7,671 12.8 ms 20.7 ms 28.0 ms 41.0 ms 91.0 ms
Ruby Ruby full-stack 2,975 33.1 ms 281.6 ms 313.6 ms 347.7 ms 1,046.7 ms

One interleaved session, every arm a fresh process, zero failed requests. Altair's write max (25.2 ms) is the best of all four frameworks - tighter than Fiber's 36.8 ms - and its read tail is deterministic on the official gcry collector (no intermittent window; see the gcry section). Altair holds ~1.7-1.8x Express on throughput.

The official collector: gcry#

Altair ships gcry as its official garbage collector. The tables above are the default build (bundled Boehm); building with a single flag — -Dgc_none — swaps the collector with no code change. A root-cause investigation traced Boehm's intermittent read-tail spike (p99.9 up to 154 ms on unlucky runs) to stop-the-world pauses that scale with heap size; gcry removes the spike deterministically.

Interleaved A/B, medians of three rounds per arm:

Workload Collector rps p99 p99.9 max
Write Boehm 15,335 12.7 ms 14.9 ms 21.3 ms
Write gcry 0.38.0 15,292 12.7 ms 15.6 ms 26.7 ms
Read Boehm 14,571 15.7 ms 25.1 ms* 60.1 ms*
Read gcry 0.38.0 14,958 14.8 ms 22.4 ms 36.3 ms

*The Boehm read medians flatter it: one of its three rounds exploded to 123.5 / 482.2 ms — the intermittent dirty window. gcry's worst round was 23.3 / 37.9 ms. Write throughput is at parity (earlier gcry versions paid an 11 percent write penalty; 0.38.0 removed it).

The closest architectural peer: the Ruby full-stack arm#

The most instructive comparison is not the raw throughput table — it is the arm built the same way Altair is. The Ruby arm is a full-stack MVC framework with the same shape: an ORM writing every row per request, convention-driven controllers, a batteries-included philosophy, and a real production application server (Puma) in front of it. It is the closest thing in this comparison to "what Altair replaces", and the two are architecturally much closer to each other than either is to Fiber's hand-rolled raw-SQL handlers or Express's minimal pg pool.

At the same connection and worker budget the results are:

Metric Altair Ruby full-stack Altair advantage
Write throughput 17,245 req/s 2,587 req/s 6.7x
Read throughput 15,539 req/s 3,531 req/s 4.4x
Write p99 8.6 ms 337.7 ms 39x tighter
Read p99 14.6 ms 257.5 ms 18x tighter
Write max 18.1 ms 1,060 ms 59x tighter

Honest reading of that table:

  • The Ruby arm posts the fastest per-request median of all four arms (4.3 ms on reads) — once a request reaches a worker, the stack is quick. Its tail collapses because 6 Puma workers serialize the ORM work under load; that is the honest shape of the Ruby stack at this budget, not a misconfiguration. Scaling it means more worker processes and more database connections — a different trade than Altair's single process with one shared pool and compile-time templates.
  • Altair gets its numbers from the same philosophy the Ruby arm embodies — an ORM on every request, no raw SQL shortcuts — so the gap measures the stack itself: compiled templates and queries, fibers over a threaded scheduler, and a database path that never blocks a second fiber on one connection (see the wire-protocol fix, issue #21, in the repository).

Historical sessions#

Session 2026-09-09 — isolated databases (older profile)#

The harness was previously run at a 200-connection budget, 1,000 VUs for 60 s, with each framework against its own PostgreSQL container. Order matched the baseline; those tables live in the repository report. The current reference is the unified realistic profile above.

Methodology#

  • One k6 warm-up run per phase (ramp + settle) precedes the measured constant-vus run, so the summaries describe sustained load only.
  • The budget is identical for every framework: 20 database connections and 6 workers (Express: cluster workers splitting the pool; Fiber: GOMAXPROCS + pgxpool max; Altair: CRYSTAL_WORKERS + one shared pool; Ruby: Puma workers x threads). Override via env for stress reruns.
  • Latency statistics: avg, med, p(90), p(95), p(99), p(99.9), max.
  • Raw k6 summary exports are committed next to the harness in results/.

Where this came from#

The numbers are maintained together with the benchmark harness: see examples/benchmark_k6/README.md for the full report — the tail-latency investigation, the per-second GC pause measurement and the allocation findings — and the repository's CHANGELOG.md for what changed between measurement passes.