YADF — Yet Another Dupes Finder

It's fast on my machine.

Installation

Prebuilt Packages

Executable binaries for some platforms are available in the releases section.

Building from source

1. Install Rust Toolchain
2. Run cargo install yadf

Usage

yadf defaults:

• search current working directory $PWD
• output format is the same as the "standard" fdupes, newline separated groups
• descends automatically into subdirectories
• search includes every files (including empty files)

bash yadf # find duplicate files in current directory yadf ~/Documents ~/Pictures # find duplicate files in two directories yadf --depth 0 file1 file2 # compare two files yadf --depth 1 # find duplicates in current directory without descending fd --type d a | yadf --depth 1 # find directories with an "a" and search them for duplicates without descending fd --type f a | yadf # find files with an "a" and check them for duplicates

Filtering

bash yadf --min 100M # find duplicate files of at least 100 MB yadf --max 100M # find duplicate files below 100 MB yadf --pattern '*.jpg' # find duplicate jpg yadf --regex '^g' # find duplicate starting with 'g' yadf --rfactor over:10 # find files with more than 10 copies yadf --rfactor under:10 # find files with less than 10 copies yadf --rfactor equal:1 # find unique files

Formatting

Look up the help for a list of output formats yadf -h.

bash yadf -f json yadf -f fdupes yadf -f csv yadf -f ldjson

Notes on the algorithm

Most¹ dupe finders follow a 3 steps algorithm:

1. group files by their size
2. group files by their first few bytes
3. group files by their entire content

yadf skips the first step, and only does the steps 2 and 3, preferring hashing rather than byte comparison. In my tests having the first step on a SSD actually slowed down the program. yadf makes heavy use of the standard library BTreeMap, it uses a cache aware implementation avoiding too many cache misses. yadf uses the parallel walker provided by ignore (disabling its ignore features) and rayon's parallel iterators to do each of these 2 steps in parallel.

¹: some need a different algorithm to support different features or different performance trade-offs

Design goals

I sought out to build a high performing artefact by assembling together libraries doing the actual work, nothing here is custom made, it's all "off-the-shelf" software.

Benchmarks

The performance of yadf is heavily tied to the hardware, specifically the NVMe SSD. I recommend fclones as it has more hardware heuristics. and in general more features.

My home directory contains upwards of 700k paths and 39 GB of data, and is probably a pathological case of file duplication with all the node_modules, python virtual environments, rust target, etc. Arguably, the most important column here is the mean time when the filesystem cache is cold.

| Program | Version | Warm Mean time (s) | Cold Mean time (s) | | :-------------- | ------: | -----------------: | -----------------: | | yadf | 0.13.1 | 2.812 | 21.554 | | fclones | 0.8.0 | 4.111 | 19.452 | | jdupes | 1.14.0 | 11.815 | 129.132 | | ddh | 0.11.3 | 10.424 | 27.241 | | fddf | 1.7.0 | 5.595 | 32.661 | | rmlint | 2.9.0 | 17.516 | 67.580 | | dupe-krill | 1.4.4 | 8.791 | 127.860 |

fdupes is excluded from this benchmark because it's really slow.

The script used to benchmark can be read here.