Performance
Tetsu adds a fraction of a microsecond of processor time per request over a
handler written for Bun.serve directly. This page shows the numbers and
what they do not cover. The full method and every reading are in
bench/README.md.
Processor time per request
Section titled “Processor time per request”Four routes, the same in every target: GET /ping; GET /users/:id behind
a hook that adds user and an observer after the response; POST /items
with a validated JSON body; and a 404. Microseconds per request, lower is
better — Bun 1.4.2 on an Apple M5 Pro, medians of three interleaved rounds
(2026-09-27):
| ping | users + 2 hooks | POST, validated | 404 | |
|---|---|---|---|---|
| raw Bun | 4.78 | 4.81 | 5.85 | 4.82 |
| Tetsu | 4.95 | 5.09 | 6.60 | 5.13 |
| Elysia 1.4.30 | 4.89 | 5.13 | 6.31 | 4.91 |
| Elysia 2.0.0-beta.16 | 5.07 | 5.47 | 6.29 | 5.10 |
| Hono 4.13.8 | 5.16 | 5.87 | 6.88 | 5.41 |
Tetsu adds 0.17–0.75 µs a request over raw Bun and keeps 90–97% of its
throughput. It is ahead of Hono on every route. Against Elysia 1.4, which
compiles a function per route ahead of time, it is level on the GETs and
0.22–0.29 µs behind on the 404 and the validated POST.
Processor time is the number to compare, because it does not depend on whether the load generators kept up. Requests per second and latency are in the full report.
Startup and memory
Section titled “Startup and memory”| startup | slowest first request | idle RSS | peak RSS under load | |
|---|---|---|---|---|
| raw Bun | 3 ms | 0.28 ms | 14.3 MB | 38–53 MB |
| Tetsu | 5 ms | 0.51 ms | 21.7 MB | 47–59 MB |
| Hono | 6 ms | 1.17 ms | 22.1 MB | 51–60 MB |
| Elysia 1.4.30 | 24 ms | 2.15 ms | 38.3 MB | 59–72 MB |
| Elysia 2 beta | 13 ms | 2.24 ms | 41.2 MB | 65–77 MB |
The benchmark loads Tetsu from its sources. The published packages are bundled and idle at 14.1 MB. No target grows its heap after load.
The cost of the types
Section titled “The cost of the types”On every change, CI measures the compiler’s work for a generated application of 200 routes and fails if it goes over budget:
| routes | types | instantiations | memory | check |
|---|---|---|---|---|
| 2 | 10 648 | 27 218 | 56 MB | 0.03 s |
| 200 | 30 857 | 235 422 | 83 MB | 0.12 s |
| 800 | 92 057 | 866 322 | 166 MB | 0.45 s |
TypeScript 7.0.2. The cost grows linearly, about 800 instantiations a route.
The comparison with other frameworks uses an earlier version of these routes,
without an afterResponse observer and an onError mapper. Written as one
chain, the shape Elysia’s and Hono’s typed clients need, they cost 330–408
thousand instantiations at 200 routes and 2.75–3.04 million at 800, against
196 and 733 thousand in Tetsu.
Validation
Section titled “Validation”Validation goes through Standard Schema, so its cost is the library’s. From
bun run --cwd bench validators, in nanoseconds per check:
| Zod 4.6 | ArkType 2.2 | Valibot 1.5 | TypeBox via tb() |
|
|---|---|---|---|---|
| small body, valid | 23 | 24 | 21 | 6.7 |
| 20-item body, valid | 922 | 168 | 784 | 52 |
| 20-item body, one item invalid | 1,020 | 3,260 | 936 | 21,090 |
| memory to import | +21 MB | +57 MB | +3 MB | +36 MB |
TypeBox’s error path is the slowest of them: with issues: "summary" the
invalid body above takes 76 ns.
Database queries
Section titled “Database queries”What a query costs through @tetsujs/sql, next to
Bun.SQL alone and to Drizzle on the same client. Processor microseconds per
operation, with 32 operations in flight on a pool of 10 connections — Bun
1.4.3 and Postgres 18.6 in Docker on the same Apple M5 Pro, medians of five
interleaved rounds (2026-10-14):
| one row | 100 rows | insert | transaction | |
|---|---|---|---|---|
| Bun.SQL | 7.04 | 26.20 | 8.82 | 44.74 |
@tetsujs/sql, query files |
8.86 | 29.58 | 10.52 | 52.02 |
@tetsujs/sql, builder |
10.37 | 33.77 | 11.96 | 55.27 |
| Drizzle 0.45.3 | 15.71 | 62.57 | 15.22 | 59.73 |
| Drizzle, prepared | 8.16 | 47.84 | 9.21 | 59.47 |
Query files add 1.8 µs to a one-row read over Bun.SQL alone, and keep 93% of its throughput. The builder adds 3.3 µs, of which 1.5 is Kysely building and compiling the query, and costs a third less than Drizzle’s. A query runs on Bun’s pool, which sends the queries of many callers down one connection, so one-row reads reach 112,000 a second through query files on 10 connections, and 105,500 through the builder. One at a time, a one-row read takes 71–87 µs for every target: most of it is the round trip to the server.
Reproducing
Section titled “Reproducing”git clone https://github.com/tetsujs/tetsu && cd tetsubun installbun run --cwd bench httpEvery target and every load generator runs in its own process. ROUNDS=3
interleaves rounds and reports medians; ONLY= picks targets.
bun run --cwd bench types, validators and cost run the other
measurements, and DATABASE_URL=postgres://… bun run --cwd bench sql the
database’s, on a server where it may create a database of its own.
What these numbers do not say
Section titled “What these numbers do not say”- They come from one machine and one level of concurrency.
- The load generator counts responses but does not check their bodies.
- The routes are small. In an application that talks to a database, the framework’s share of each request is smaller still.
- Two runs back to back behave like two different machines. Compare targets within one run, with the baseline in it.