GOFORJ / BENCHMARKS

Performance, backed by evidence.

Benchmarking is a first-class part of how we engineer for performance and stability. We aim for 100% code coverage, test integrations against real services, containers, and emulators, and maintain performance benchmarks alongside our libraries.

Compare latency, throughput, and allocation costs before choosing what runs in your app. Published workloads, source data, and environment details let you inspect the conditions behind each result, reproduce the measurements, and test your own workloads.

HTTP7 routers · 3 workloads↓Cache9 driver results↓Queue9 driver results↓Events7 driver results↓Storage8 driver results↓Mail8 driver results↓Database3 driver results↓

Compare HTTP routers.#

Same endpoint workload, different request pipelines. httprouter focuses on routing; GoForj Web includes Echo's context and response handling plus a bridge for Web middleware.

HTTP: Read an ID from the URL#

Return the ID in a JSON response

7 routers

Source ↗Reproduce ↗

Router	Time / opLower is faster 100 ns1 µs	Operations / secHigher is faster 1M10M	Allocs / opLower uses less 015	Bytes / opLower uses less 0 B1,500 B
httprouter	Time / op236 ns	Operations / sec4.24M	Allocs / op4	Bytes / op80 B
GoForj Web	Time / op305.3 ns	Operations / sec3.28M	Allocs / op3	Bytes / op48 B
Gin	Time / op307.2 ns	Operations / sec3.25M	Allocs / op4	Bytes / op64 B
Echo	Time / op314.3 ns	Operations / sec3.18M	Allocs / op3	Bytes / op48 B
net/http	Time / op328.9 ns	Operations / sec3.04M	Allocs / op4	Bytes / op64 B
Chi	Time / op548.6 ns	Operations / sec1.82M	Allocs / op7	Bytes / op752 B
Gorilla Mux	Time / op762.4 ns	Operations / sec1.31M	Allocs / op11	Bytes / op1,200 B

The same in-process router workload, without a socket, network, or application business logic. Time and throughput use logarithmic axes (each interval is 10×). Allocations and bytes use linear axes from zero. Rates describe serial handler dispatch, not a concurrent HTTP load test. Time is the median of seven run averages; whiskers show run-to-run ranges, not request percentiles.

Snapshot 41f644d · Capture date not recorded · go1.27.2 · linux/arm64 · arm64 (CPU model unavailable) · GOMAXPROCS=1 · 7 runs × 1s

HTTP: Return a text response#

Return fixed text without middleware

7 routers

Source ↗Reproduce ↗

Router	Time / opLower is faster 10 ns1 µs	Operations / secHigher is faster 1M100M	Allocs / opLower uses less 08	Bytes / opLower uses less 0 B900 B
httprouter	Time / op65.13 ns	Operations / sec15.35M	Allocs / op1	Bytes / op16 B
net/http	Time / op125.4 ns	Operations / sec7.98M	Allocs / op1	Bytes / op16 B
GoForj Web	Time / op135.6 ns	Operations / sec7.37M	Allocs / op1	Bytes / op16 B
Echo	Time / op138 ns	Operations / sec7.25M	Allocs / op1	Bytes / op16 B
Gin	Time / op147.8 ns	Operations / sec6.76M	Allocs / op1	Bytes / op48 B
Chi	Time / op249.6 ns	Operations / sec4.01M	Allocs / op3	Bytes / op384 B
Gorilla Mux	Time / op468.7 ns	Operations / sec2.13M	Allocs / op8	Bytes / op864 B

The same in-process router workload, without a socket, network, or application business logic. Time and throughput use logarithmic axes (each interval is 10×). Allocations and bytes use linear axes from zero. Rates describe serial handler dispatch, not a concurrent HTTP load test. Time is the median of seven run averages; whiskers show run-to-run ranges, not request percentiles.

Snapshot 41f644d · Capture date not recorded · go1.27.2 · linux/arm64 · arm64 (CPU model unavailable) · GOMAXPROCS=1 · 7 runs × 1s

HTTP: Run request middleware#

Set request state, add a header, validate the request

7 routers

Source ↗Reproduce ↗

Router	Time / opLower is faster 100 ns1 µs	Operations / secHigher is faster 1M10M	Allocs / opLower uses less 015	Bytes / opLower uses less 0 B1,500 B
httprouter	Time / op237.6 ns	Operations / sec4.21M	Allocs / op4	Bytes / op400 B
Echo	Time / op238.4 ns	Operations / sec4.2M	Allocs / op2	Bytes / op32 B
net/http	Time / op309 ns	Operations / sec3.24M	Allocs / op4	Bytes / op400 B
Gin	Time / op362.4 ns	Operations / sec2.76M	Allocs / op4	Bytes / op400 B
GoForj Web	Time / op369.7 ns	Operations / sec2.71M	Allocs / op3	Bytes / op48 B
Chi	Time / op421.1 ns	Operations / sec2.37M	Allocs / op6	Bytes / op768 B
Gorilla Mux	Time / op723.7 ns	Operations / sec1.38M	Allocs / op14	Bytes / op1,320 B

The same in-process router workload, without a socket, network, or application business logic. Time and throughput use logarithmic axes (each interval is 10×). Allocations and bytes use linear axes from zero. Rates describe serial handler dispatch, not a concurrent HTTP load test. Time is the median of seven run averages; whiskers show run-to-run ranges, not request percentiles.

Snapshot 41f644d · Capture date not recorded · go1.27.2 · linux/arm64 · arm64 (CPU model unavailable) · GOMAXPROCS=1 · 7 runs × 1s

Cache#

GetBytes

9 results

Source ↗Reproduce ↗

Driver	Time / opLower is faster 10 ns1 ms	Operations / secHigher is faster 1K100M	Allocs / opLower uses less 040	Bytes / opLower uses less 0 B3,000 B
Memory	Time / op70 ns	Operations / sec14.29M	Allocs / op1	Bytes / op1 B
SQLite	Time / op7.49 µs	Operations / sec133.44K	Allocs / op39	Bytes / op1,432 B
Memcached	Time / op70.36 µs	Operations / sec14.21K	Allocs / op9	Bytes / op192 B
Redis	Time / op73.17 µs	Operations / sec13.67K	Allocs / op8	Bytes / op200 B
MySQL	Time / op81.96 µs	Operations / sec12.2K	Allocs / op26	Bytes / op744 B
NATS KV + TTL	Time / op82.87 µs	Operations / sec12.07K	Allocs / op40	Bytes / op2,556 B
PostgreSQL	Time / op86.66 µs	Operations / sec11.54K	Allocs / op32	Bytes / op1,064 B
NATS	Time / op96.45 µs	Operations / sec10.37K	Allocs / op40	Bytes / op2,573 B
File	Time / op100.98 µs	Operations / sec9.9K	Allocs / op9	Bytes / op1,352 B

Seeded byte-value reads across in-process, filesystem, and local service fixtures. NATS KV + TTL is a second configuration of the NATS driver. Time and throughput use logarithmic axes (each interval is 10×). Allocations and bytes use linear axes from zero. Rates are calculated from time per operation, not a concurrent load test. Recorded per-operation averages; repeated-run samples are unavailable.

Snapshot 5e5d0f2 · Capture date, environment, and repeated-run details were not recorded.

Not in this snapshot

Null

No-op driver; not included in the seeded-read benchmark.

DynamoDB

Intentionally omitted: DynamoDB Local timings do not represent real AWS latency.

Queue#

Dispatch return

9 results

Source ↗Reproduce ↗

Driver	Time / opLower is faster 100 ns10 ms	Operations / secHigher is faster 10010M	Allocs / opLower uses less 01,500	Bytes / opLower uses less 0 B95,000 B
Sync	Time / op282.5 ns	Operations / sec3.54M	Allocs / op6	Bytes / op408 B
Worker pool	Time / op650 ns	Operations / sec1.54M	Allocs / op7	Bytes / op456 B
NATS	Time / op774.1 ns	Operations / sec1.29M	Allocs / op13	Bytes / op1,258 B
Redis	Time / op95.3 µs	Operations / sec10.49K	Allocs / op33	Bytes / op2,113 B
RabbitMQ	Time / op165.78 µs	Operations / sec6.03K	Allocs / op57	Bytes / op1,882 B
SQLite	Time / op202.38 µs	Operations / sec4.94K	Allocs / op47	Bytes / op1,931 B
PostgreSQL	Time / op1.06 ms	Operations / sec946.31	Allocs / op78	Bytes / op3,809 B
SQS	Time / op1.87 ms	Operations / sec533.64	Allocs / op1,082	Bytes / op94,784 B
MySQL	Time / op2.29 ms	Operations / sec437.37	Allocs / op62	Bytes / op3,303 B

DISCARD-ONLY REFERENCE · Jobs are dropped

NullTime / op37.49 nsOperations / sec26.67MAllocs / op1Bytes / op128 B

Producer-side dispatch return across local fixtures, not job completion. Core NATS publish return is not durable acknowledgement. Time and throughput use logarithmic axes (each interval is 10×). Allocations and bytes use linear axes from zero. Rates are calculated from time per operation, not a concurrent load test. Recorded per-operation averages; repeated-run samples are unavailable.

Snapshot 439da9b · Capture date, environment, and repeated-run details were not recorded.

Events#

Publish + handler delivery

7 results

Source ↗Reproduce ↗

Driver	Time / opLower is faster 100 ns100 ms	Operations / secHigher is faster 1010M	Allocs / opLower uses less 01,500	Bytes / opLower uses less 0 B150,000 B
Synchronous	Time / op371.8 ns	Operations / sec2.69M	Allocs / op9	Bytes / op320 B
NATS	Time / op73.32 µs	Operations / sec13.64K	Allocs / op18	Bytes / op681 B
Redis	Time / op95.15 µs	Operations / sec10.51K	Allocs / op31	Bytes / op914 B
NATS JetStream	Time / op127.17 µs	Operations / sec7.86K	Allocs / op60	Bytes / op3,609 B
Kafka	Time / op499.09 µs	Operations / sec2K	Allocs / op67	Bytes / op5,040 B
SNS	Time / op2.55 ms	Operations / sec391.68	Allocs / op1,271	Bytes / op104,563 B
Google Pub/Sub	Time / op52.85 ms	Operations / sec18.92	Allocs / op369	Bytes / op21,971 B

Publish plus observed handler delivery across in-process and local service fixtures, including the service round trip. Time and throughput use logarithmic axes (each interval is 10×). Allocations and bytes use linear axes from zero. Rates are calculated from time per operation, not a concurrent load test. Recorded per-operation averages; repeated-run samples are unavailable.

Snapshot 21c6e50 · Capture date, environment, and repeated-run details were not recorded.

Not in this snapshot

Null

No-op driver; not included in the publish-and-deliver benchmark.

Storage#

Get small object

8 results

Source ↗Reproduce ↗

Driver	Time / opLower is faster 100 ns1 ms	Operations / secHigher is faster 1K10M	Allocs / opLower uses less 07,000	Bytes / opLower uses less 0 B250,000 B
Memory	Time / op752.91 ns	Operations / sec1.33M	Allocs / op9	Bytes / op496.08 B
Local	Time / op13.81 µs	Operations / sec72.4K	Allocs / op14	Bytes / op1,536.24 B
rclone local	Time / op17.96 µs	Operations / sec55.69K	Allocs / op21	Bytes / op2,168.03 B
Redis	Time / op79.51 µs	Operations / sec12.58K	Allocs / op24	Bytes / op936.03 B
GCS	Time / op210.78 µs	Operations / sec4.74K	Allocs / op6,831.4	Bytes / op201,477.99 B
FTP	Time / op243.28 µs	Operations / sec4.11K	Allocs / op99.03	Bytes / op5,429.79 B
S3	Time / op382.58 µs	Operations / sec2.61K	Allocs / op630.14	Bytes / op45,419.27 B
SFTP	Time / op429.23 µs	Operations / sec2.33K	Allocs / op47.01	Bytes / op2,523.78 B

Read a previously written object across in-process, filesystem, and local service fixtures. rclone uses a local filesystem remote. Time and throughput use logarithmic axes (each interval is 10×). Allocations and bytes use linear axes from zero. Rates are calculated from time per operation, not a concurrent load test. Recorded per-operation averages; repeated-run samples are unavailable.

Snapshot 1ba3e4c · Capture date, environment, and repeated-run details were not recorded.

Not in this snapshot

Dropbox

No local emulator or container fixture is configured for Dropbox in this benchmark suite.

Mail#

Client send return · one recipient + 4 KiB attachment

8 results

Source ↗Reproduce ↗

Driver	Time / opLower is faster 1 µs10 ms	Operations / secHigher is faster 1001M	Allocs / opLower uses less 0650	Bytes / opLower uses less 0 B150,000 B
Fake	Time / op2.69 µs	Operations / sec372.3K	Allocs / op30	Bytes / op9,450 B
Log	Time / op3.88 µs	Operations / sec257.4K	Allocs / op34	Bytes / op10,192 B
Mailgun	Time / op48.05 µs	Operations / sec20.81K	Allocs / op223	Bytes / op27,414 B
Resend	Time / op52.99 µs	Operations / sec18.87K	Allocs / op113	Bytes / op33,298 B
Postmark	Time / op53.52 µs	Operations / sec18.68K	Allocs / op116	Bytes / op33,426 B
SendGrid	Time / op53.6 µs	Operations / sec18.66K	Allocs / op103	Bytes / op32,911 B
SES	Time / op103.72 µs	Operations / sec9.64K	Allocs / op610	Bytes / op106,222 B
SMTP	Time / op3.71 ms	Operations / sec269.18	Allocs / op316	Bytes / op58,850 B

Local client costs, not real provider latency or inbox delivery. API drivers use HTTP acceptance fixtures in separate processes without TLS; SMTP uses Mailpit. Log writes to io.Discard; Fake resets every 128 sends. Time and throughput use logarithmic axes (each interval is 10×). Allocations and bytes use linear axes from zero. Rates are calculated from time per operation, not a concurrent load test. Time is the median of seven run averages; whiskers show run-to-run ranges, not request percentiles.

Snapshot 3e74fd8 · 2026-10-17 · go1.27.2 darwin/arm64 · darwin/arm64 · Apple M4 Max · GOMAXPROCS=1 · 7 runs × 1s

Database#

Read one row by primary key · GORM

3 results

Source ↗Reproduce ↗

Driver	Time / opLower is faster 10 µs1 ms	Operations / secHigher is faster 1K100K	Allocs / opLower uses less 070	Bytes / opLower uses less 0 B4,500 B
SQLite	Time / op11.05 µs	Operations / sec90.52K	Allocs / op66	Bytes / op4,330 B
PostgreSQL	Time / op117.75 µs	Operations / sec8.49K	Allocs / op65	Bytes / op4,324 B
MySQL	Time / op188.1 µs	Operations / sec5.32K	Allocs / op68	Bytes / op4,236 B

Warm lookups across 1,000 rows with one pooled connection and GoForj database defaults. MySQL and PostgreSQL use local containers; SQLite uses a temporary file. Measures the GORM operation, without generated-app instrumentation. Time and throughput use logarithmic axes (each interval is 10×). Allocations and bytes use linear axes from zero. Rates are calculated from time per operation, not a concurrent load test. Time is the median of seven run averages; whiskers show run-to-run ranges, not request percentiles.

Snapshot 0bdbb84 · 2026-10-17 · go1.27.2 darwin/arm64 · darwin/arm64 · Apple M4 Max · GOMAXPROCS=1 · 7 runs × 1s

These results describe library harnesses and local service fixtures. Allocations measure the benchmarked Go process, not backend memory usage. Measure your own payloads, network, concurrency, and delivery requirements before making capacity decisions.

Benchmark your app in Lighthouse ↗Driver capabilities ↗