When to use something else

YAML, TOML, CSV, MessagePack and Protobuf each win somewhere JSON does not - and each has a cost worth naming before you adopt it.

Configuration: YAML and TOML

# YAML: comments, anchors, and significant indentation
defaults: &defaults
  retries: 3
  timeout_seconds: 30

services:
  api:
    <<: *defaults
    port: 8080
    hosts:
      - api.example.com
      - api.internal

# TOML: explicit, typed, and unambiguous about nesting
[service]
name = "api"
port = 8080

[service.limits]
retries = 3
timeout_seconds = 30
enabled = true
released = 2026-09-18          # a real date type, not a string

Tabular data: CSV

id,email,plan,created_at
1042,[email protected],annual,2026-01-04
1043,[email protected],monthly,2026-02-11

• CSV is smaller, streams line by line, and opens in every spreadsheet - which is often the actual requirement.
• There are no types: everything is a string, and the reader decides whether 007 is a number or an identifier.
• There is no nesting. A list of line items needs either a second file or an encoded column.
• Quoting rules are only mostly standard; use a real CSV writer rather than joining with commas.

# converting between the two, with the shape made explicit
jq -r '.users[] | [.id, .email, .plan, .created_at] | @csv' users.json > users.csv

# and back, treating every value as text
python - <<'PY'
import csv, json
with open("users.csv") as f:
    rows = list(csv.DictReader(f))
print(json.dumps(rows, indent=2))
PY

Binary formats: size and speed

import { encode, decode } from "@msgpack/msgpack";

const bytes = encode({ id: "ord_9f2c", total_minor: 3998 });
byteLength(bytes);                 // 40, against 46 for the JSON text

// the real saving is in a large array of similar records
const compact = encode(orders);    // repeated keys are the compression win

1. Measure before switching. Compressed JSON is frequently as small as a binary format, and it needs no new tooling.
2. Binary formats trade debuggability for size and speed: a payload is unreadable in a log, a proxy or a terminal, which is a real operational cost.
3. Protobuf and Avro earn their complexity when a schema registry and code generation remove more work than they add, and when the same data crosses several services.
4. Never use a binary format for data a human will read: configuration, fixtures and API responses that people debug by hand.
5. Keep JSON at the edges. Services can speak binary internally while the public interface stays JSON - most of the size saving is between services, not at the boundary.

# the comparison that matters, on your own data
ls -l payload.json
gzip -9 -c payload.json | wc -c
node -e 'const m=require("@msgpack/msgpack"),fs=require("fs");
console.log(m.encode(JSON.parse(fs.readFileSync("payload.json"))).length)'

FAQ

Related

Reading and inspecting JSON from the command line JSON in databases and storage

Last refreshed 2026-09-18.