uniqpath

Stop overwriting your files.
Three functions. Zero dependencies. Race-free when you need it.

---

uniqpath gives you a unique filename or directory path in Python, so writing output.txt twice never overwrites the first one. It can also reserve the path atomically, which is what you need when parallel jobs write into the same folder.

---

We have all shipped this function:

# every codebase, somewhere
i = 1
while os.path.exists(f"{name}_{i}.txt"):
    i += 1

And we have all watched a parallel job quietly overwrite three hours of results, because two workers ran that loop at the same time.

pip install uniqpath

from uniqpath import unique_path

unique_path("output.txt")   # → output.txt      nothing there yet
unique_path("output.txt")   # → output_1.txt    now there is
unique_path("output.txt")   # → output_2.txt    you get the idea

That is the whole library. Then there is the part that actually matters.

The race nobody thinks about

unique_path() tells you a name that is free right now. Between that answer and your open(), another process can grab it. In a single script, who cares. In 16 parallel workers, that is your evening.

reserve_path() closes the gap: it creates the path in the same syscall that tests it — O_CREAT | O_EXCL for a file, mkdir for a directory. The loser of a race gets FileExistsError and quietly moves to the next candidate.

from uniqpath import reserve_path

run_dir = reserve_path("experiments/run", is_dir=True)   # yours, guaranteed

# 16 workers, 16 distinct directories, zero coordination
with ThreadPoolExecutor(max_workers=16) as pool:
    dirs = pool.map(lambda _: reserve_path("out/run", is_dir=True), range(16))

assert len({str(d) for d in dirs}) == 16   # run, run_1, run_2, ... run_15

And when you were going to write to the file anyway, skip the middleman:

from uniqpath import uniq_open

with uniq_open("results.csv") as f:
    f.write("...")
    print("wrote", f.name)      # results_3.csv

uniq_open() reserves the path, opens it, and closes it on the way out. f.name is the path it actually used.

	tells you a free name	creates it	opens it	safe under concurrency
unique_path()	✅	❌	❌	❌
reserve_path()	✅	✅	❌	✅
uniq_open()	✅	✅	✅	✅

Install

pip install uniqpath          # library + CLI
pipx install uniqpath         # just the CLI, isolated
uv tool install uniqpath      # same, with uv
uvx uniqpath output.txt       # no install at all

Homebrew, conda, Arch

brew install julienrabault/tap/uniqpath
conda install -c conda-forge uniqpath
yay -S python-uniqpath

Recipes and their status live in packaging/.

Python 3.9+. Linux, macOS, Windows. No runtime dependencies, ever.

Name your files however you like

The suffix is a format string. Mix and match:

unique_path("run.log",   suffix_format="_{date:%Y-%m-%d}_{num:03d}")
# → run_2026-09-22_001.log

unique_path("shard.parquet", suffix_format="_{pid}_{uuid:8}")
# → shard_48213_5ba950c1.parquet

unique_path("backup.tar.gz", suffix_format=".{timestamp}")
# → backup.tar.1790080200.gz

Placeholder	What you get	Example
{num}	attempt counter, from 1	_7
{num:03d}	…zero-padded, any format spec	_007
{date}	current datetime, full strftime	{date:%Y-%m-%d} → _2026-09-22
{timestamp}	UNIX seconds	_1790080200
{rand}	random alphanumerics, 6 by default	{rand:4} → _a7Zq
{uuid}	UUID4 hex, 32 by default	{uuid:8} → _5ba950c1
{pid}	current process id	_48213

Typo a placeholder and you get told immediately, with the list of valid ones — not a KeyError five minutes into a job.

From the shell

$ uniqpath output.txt
output.txt

$ touch output.txt && uniqpath output.txt
output_1.txt

$ uniqpath results --dir --reserve     # creates it, then prints it
results_1

Or skip the variable entirely — --exec reserves the path and drops it into your command wherever you put {}:

uniqpath results/run --dir --exec -- python train.py --out {}

uniqpath exits with your command's exit code, and leaves stdout alone so pipes keep working.

Which makes job scripts boring, in the good way:

#!/bin/bash
#SBATCH --array=0-63

uniqpath "results/$SLURM_JOB_NAME" --dir --exec -- python train.py --out {}

64 array tasks, 64 directories, no $SLURM_ARRAY_TASK_ID arithmetic, no collisions.

Completion

eval "$(uniqpath --completion bash)"                               # bash
uniqpath --completion zsh  > "${fpath[1]}/_uniqpath"               # zsh
uniqpath --completion fish > ~/.config/fish/completions/uniqpath.fish

It completes paths, options, and the suffix formats above — so you stop looking them up.

Files vs directories

By default the kind is guessed: a path with an extension that is not already a directory is a file, and the suffix goes before the extension.

unique_path("archive.tar.gz")   # → archive.tar_1.gz
unique_path("results")          # → results_1

Guessing has limits. Say so explicitly when it matters:

unique_path("release.v1.0", is_dir=True)    # → release.v1.0_1
unique_path("release.v1.0", is_dir=False)   # → release.v1_1.0

Full API

unique_path(
    path,                        # str | os.PathLike
    suffix_format="_{num}",
    if_exists_only=True,         # return path untouched when it is already free
    return_str=False,            # return str instead of Path
    max_num=50_000,              # attempts before giving up
    verbose=False,               # log each attempt on the "uniqpath" logger
    is_dir=None,                 # None = guess, True = directory, False = file
) -> Path | str

reserve_path(
    ...,                         # everything above, plus:
    parents=True,                # create missing parent directories
) -> Path | str

uniq_open(                       # context manager
    path,
    mode="w",                    # any writing mode: w, a, x, +b variants
    *,                           # plus suffix_format, if_exists_only, max_num,
    ...,                         # parents, buffering, encoding, errors, newline
) -> IO                          # f.name is the path that was used

Errors

from uniqpath import UniqPathError, MaxAttemptsError, InvalidFormatError

• MaxAttemptsError — no free path within max_num attempts. Also a RuntimeError, so pre-0.2 handlers keep working.
• InvalidFormatError — unknown or malformed placeholder. Also a KeyError, for the same reason.

A suffix with no varying part ("_backup", "_{timestamp}") has exactly one possible candidate. If it is taken, you get MaxAttemptsError straight away instead of 50 000 pointless stat calls.

Gotchas worth knowing

• unique_path() does one stat per attempt. A directory holding thousands of name_N siblings costs thousands of syscalls — use {rand} or {uuid} there, they land on the first try.
• reserve_path() creates an empty file. Opening it afterwards in "w" is the normal flow, and it never truncates something that already existed.
• Reservations are not garbage-collected. A path you reserve and never use stays on disk.
• Atomicity rests on O_EXCL and mkdir. On NFS without working locking, O_EXCL guarantees are weaker — a filesystem limitation, not a uniqpath one.

Common questions

How do I avoid overwriting a file in Python? Ask for a free name before you write: unique_path("output.txt") returns output_1.txt once output.txt exists. If another process might be writing to the same folder, use reserve_path() or uniq_open() instead — they hold the path for you.

How do I increment a filename if it already exists? That is the default: _{num} counts from 1. Use suffix_format="_{num:03d}" if you want _001, _002, and so on.

How do I get a unique filename for every run of a script? unique_path("run.log", suffix_format="_{date:%Y-%m-%d}_{num:03d}") for something readable, or _{uuid:8} when you only care that it never collides.

How do I create a new output directory per job, without collisions? reserve_path("results/run", is_dir=True), or from the shell: uniqpath results/run --dir --exec -- python train.py --out {}.

Contributing

Issues and PRs welcome.

git clone https://github.com/JulienRabault/uniqpath
cd uniqpath
pip install -e ".[dev]"

pytest -q --cov=uniqpath              # 190 tests
ruff check . && ruff format --check .
mypy                                  # strict

CI runs the suite on Linux, macOS and Windows across Python 3.9 → 3.13.

License

MIT — Julien Rabault.

_{If this saved you a results directory, a ⭐ is appreciated.}