Repository files navigation

This package implements the "FastCDC" content defined chunking algorithm inPython with optional cython support. To learn more about contentdefined chunking and its applications, see the reference material linked below.

• Python Version 3.7 and later. Tested on Linux, Mac andWindows

$ pip install fastcdc

To enable add additional support for the hash algorithms(xxhash andblake3) use

$ pip install fastcdc[hashes]

$ fastcdc tests/SekienAkashita.jpghash=103159aa68bb1ea98f64248c647b8fe9a303365d80cb63974a73bba8bc3167d7 offset=0 size=22366hash=3f2b58dc77982e763e75db76c4205aaab4e18ff8929e298ca5c58500fee5530d offset=22366 size=10491hash=fcfb2f49ccb2640887a74fad1fb8a32368b5461a9dccc28f29ddb896b489b913 offset=32857 size=14094hash=bd1198535cdb87c5571378db08b6e886daf810873f5d77000a54795409464138 offset=46951 size=18696hash=d6347a2e5bf586d42f2d80559d4f4a2bf160dce8f77eede023ad2314856f3086 offset=65647 size=43819

$ fastcdc -mi 16384 -s 32768 -ma 65536 -hf sha256 tests/SekienAkashita.jpghash=5a80871bad4588c7278d39707fe68b8b174b1aa54c59169d3c2c72f1e16ef46d offset=0 size=32857hash=13f6a4c6d42df2b76c138c13e86e1379c203445055c2b5f043a5f6c291fa520d offset=32857 size=16408hash=0fe7305ba21a5a5ca9f89962c5a6f3e29cd3e2b36f00e565858e0012e5f8df36 offset=49265 size=60201

$ fastcdc scan~/Downloads[####################################]  100%Files:          1,332Chunk Sizes:    min 4096 - avg 16384 - max 131072Unique Chunks:  506,077Total Data:     9.3 GBDupe Data:      873.8 MBDeDupe Ratio:   9.36 %Throughput:     135.2 MB/s

$ fastcdcUsage: fastcdc [OPTIONS] COMMAND [ARGS]...Options:  --version  Show the version and exit.  --help     Show this message and exit.Commands:  chunkify*  Find variable sized chunksfor FILE and compute hashes.  benchmark  Benchmark chunking performance.  scan       Scan filesin directory and report duplication.

The tests also have some short examples of using the chunker, of which thiscode snippet is an example:

fromfastcdcimportfastcdcresults=list(fastcdc("tests/SekienAkashita.jpg",16384,32768,65536))assertlen(results)==3assertresults[0].offset==0assertresults[0].length==32857assertresults[1].offset==32857assertresults[1].length==16408assertresults[2].offset==49265assertresults[2].length==60201

The algorithm is as described in "FastCDC: a Fast and Efficient Content-DefinedChunking Approach for Data Deduplication"; see thepaper,andpresentationfor details. There are some minor differences, as described below.

The explanation below is copied fromronomon/deduplication since thiscodebase is little more than a translation of that implementation:

The following optimizations and variations on FastCDC are involved in the chunking algorithm:

• 31 bit integers to avoid 64-bit integers for the sake of the Javascript reference implementation.
• A right shift instead of a left shift to remove the need for an additional modulus operator, which would otherwise have been necessary to prevent overflow.
• Masks are no longer zero-padded since a right shift is used instead of a left shift.
• A more adaptive threshold based on a combination of average and minimum chunk size (rather than just average chunk size) to decide the pivot point at which to switch masks. A larger minimum chunk size now switches from the strict mask to the eager mask earlier.
• Masks use 1 bit of chunk size normalization instead of 2 bits of chunk size normalization.

The primary objective of this codebase was to have a Python implementation with apermissive license, which could be used for new projects, without concern fordata parity with existing implementations.

This package started as Python port of the implementation by Nathan Fiedler (see thenlfiedler link below).

• nlfiedler/fastcdc-rs
  • Rust implementation on which this code is based.
• ronomon/deduplication
  • C++ and JavaScript implementation on which the rust implementation is based.
• rdedup_cdc at docs.rs
  • An alternative implementation of FastCDC to the one in this crate.
• jrobhoward/quickcdc
  • Similar but slightly earlier algorithm by some of the same researchers.

• Performance improvement@dw
• Fixed issue with inputs smaller than min_size@grote

• added python 3.12 support
• removed python 3.7 support
• updated dependencies

• added python 3.10/3.11 support
• removed python 3.6 support
• update dependencies

• add binary releases to PyPI (Xie Yanbo)
• update dependencies

• fix issue with fat option in cython version
• updated dependencies

• add support for multiple path with scan command
• fix issue with building cython extension
• fix issue with fat option
• fix zero-devision error

• add newscan command to calculate deduplication ratio for directories

• faster optional cython implementation
• benchmark command

• high-level API
• support for streams
• support for custom hash functions

• Initial release (port ofnlfiedler/fastcdc-rs).