json.dumps Compact vs Indent Lab

Intro — what this post promises

json.dumps formatting: how much do separators=(',', ':'), default spaced separators, and indent=2 cost in time and bytes? This lab focuses on dumps options (not a rematch of the json-vs-orjson codec war). orjson is measured only because it is installed on this box.

Related links:

• json vs orjson vs msgpack localhost lab
• shutil copyfile vs manual localhost lab
• str translate vs replace localhost lab
• bytes vs bytearray localhost lab
• setdefault vs defaultdict localhost lab
• string concat vs join localhost lab
• statistics vs manual mean localhost lab
• perf_counter vs time localhost lab

Lab honesty (1 Oct 2026 IST): Python 3.13.5. orjson installed: True (ujson: no). Affiliates: 0. Wire JSON should be compact; indent is for humans.

Verdict up front (medium ~200 rows): compact ~2497 dumps/s (34,035 B) vs default ~2527 (39,837 B, speed ~0.99×, size compact/default ~0.85×). indent=2 ~2155 dumps/s (86,647 B, ~2.55× compact size; compact ~1.16× faster). orjson compact ~7.4× stdlib compact.

Arms

Payloads: small (20 rows), medium (200), large (2000).

Lab topology

nested dict/list records; trials scale with size
metric: dumps/s + output UTF-8 bytes

Script: lab-evidence/75-json-dumps-compact-vs-indent/results/run_lab.py.

Lead table — medium payload (p50)

Size tradeoffs (medium)

Pretty JSON more than doubled payload size here — bandwidth and disk tax, not just CPU.

Reading it

• Compact separators are nearly free in time vs default (~0.99×) but save ~15% bytes — always use for APIs/logs on the wire.
• indent=2 costs CPU (compact ~1.16×) and ~2.5× size — debug/pretty only.
• sort_keys adds measurable overhead (compact ~1.23×) — use for deterministic diffs, not hot paths.
• orjson still dominates encode when installed (~7.4×); indent remains costly even there (~1.30×).

Relation to the orjson lab

Lab 36 compared codecs end-to-end. This lab answers: within json.dumps, which kwargs blow up size/time? Use both posts together — pick a library, then pick formatting.

Indent throughput illusion

Indent arms can show high “MB/s out” because they emit more bytes per dump — that is not a win. Compare dumps/s and bytes per document separately. For APIs, optimize for compact dumps/s and small out_bytes, not pretty MB/s.

Pitfalls

1. Shipping indent=2 in production APIs — multiplies bytes.
2. Forgetting separators=(',', ':') — default inserts spaces.
3. Sorting keys on every response — pay only when you need stable output.
4. Inventing orjson numbers when it is not installed — we only report it because import orjson succeeded here.

When to pick what

Reproduce

python3 lab-evidence/75-json-dumps-compact-vs-indent/results/run_lab.py

Evidence: /workspace/lab-evidence/75-json-dumps-compact-vs-indent/results/.

Closing

Compact on the wire; indent for eyeballs. On this box medium compact matched default speed (~0.99×) at ~15% fewer bytes, while indent=2 grew size ~2.5× and slowed dumps ~1.16×. orjson added another ~7× on compact encode when available.