tarfile vs zipfile Create+Extract: Localhost Lab

Intro — what this post promises

Pack and unpack the same tree with tarfile (uncompressed) vs zipfile (ZIP_STORED and ZIP_DEFLATED). This lab reports create/extract MB/s (payload ÷ wall) on Linux localhost.

It is not a remake of the zstd vs gzip vs lz4 localhost lab (stream codecs). Here the unit is an archive format — many small files, one create, one extract.

Related links:

• 01shoppercove.comzstd vs gzip lz4 localhost lab
• 02shoppercove.comscandir vs listdir localhost lab
• 03shoppercove.combytesio vs spooled tempfile localhost lab
• 04shoppercove.comshutil copyfile vs manual localhost lab
• 05shoppercove.commmap vs read scan localhost lab
• 06shoppercove.comglob vs rglob vs walk localhost lab
• 07shoppercove.comsecrets vs urandom localhost lab
• 08shoppercove.comhashlib md5 vs blake2b localhost lab

Lab honesty (1 Oct 2026 IST): Python 3.13.5. Fixture ~7.81 MiB payload (40×50 × 4 KiB files + README). Affiliates: 0. No Docker. Compression stated per arm.

Verdict up front: create — zip stored ~64.3 MB/s vs tar ~52.3 vs zip deflated ~44.3. Extract — zip stored ~69.4 MB/s vs deflated ~63.9 vs tar ~29.3. Deflated archive size ~873 KiB vs stored ~8270 KiB vs tar ~11070 KiB.

Arms

Lab topology

fixture: 40 dirs × 50 files × 4096 B ≈ 7.81 MiB
create + extract · 5 rounds · p50
metric: MB/s = fixture_MiB / p50_s

Script: lab-evidence/91-tarfile-vs-zipfile/results/run_lab.py.

Lead table — create (p50)

Lead table — extract (p50)

Reading it

• ZIP_STORED won raw create/extract speed here — archive format overhead beat uncompressed tar on this many-small-files tree.
• ZIP_DEFLATED traded CPU for a much smaller archive (~9.47× smaller than stored on this compressible payload).
• Uncompressed tar was largest on disk (block padding) and slowest to extract in this run.
• Pick format for interop + tooling first; measure when installers/CI spend wall time on pack/unpack.

Compression clarity

This lab’s deflate is zipfile’s zlib, not standalone gzip/zstd stream APIs. For codec bake-offs see the zstd/gzip/lz4 post. For “ship a folder,” compare archive containers as done here.

Security note

tarfile.extractall used filter='data' (3.12+). Always constrain extracts from untrusted archives (path traversal). Throughput numbers do not make untrusted zips safe.

Many small files

Archive microbenches with one huge blob hide per-member overhead (headers, central directory, ustar padding). This fixture is intentionally many small files — closer to source trees and build artifacts. That is why uncompressed tar’s padded size and extract time looked worse than zip stored despite “no compression.”

When to deflate

Deflate paid ~45% create-time vs stored here but cut archive size dramatically on repetitive 4 KiB payloads. For already-compressed inputs (jpg, mp4, .whl), deflate often wastes CPU — prefer stored or an outer zstd layer designed for that job.

Pitfalls

• Comparing deflated zip speed to uncompressed tar without stating compression.
• Using tar.gz in one arm and forgetting CPU belongs to gzip, not tar.
• Ignoring archive size when “faster” stored zip is 10× larger on the wire.
• Extracting untrusted archives without filters.

Reproduce

python3 lab-evidence/91-tarfile-vs-zipfile/results/run_lab.py

Evidence: summary.json, fixture_tree/, archives/.

Limits

One Linux box. Synthetic compressible payloads. Not tar.gz/tar.zst, not ZIP_LZMA/BZIP2. Python stdlib only.

Takeaway

On this ~7.81 MiB small-file tree, zip stored create ~64.3 MB/s and extract ~69.4 MB/s led; deflate shrank the archive to ~873 KiB at ~44.3 MB/s create. Uncompressed tar lagged extract (~29.3 MB/s). Choose zip vs tar for ecosystem needs; measure when CI packs trees every build.