A tool for splitting a large dataset into graph slices to make deals in the Filecoin Network
When storing a large dataset, we need to split it into smaller pieces to fit the sector's size, which could generally be 32GiB or 64GiB.
If we make these data into a large tarball, chunk it into small pieces, and then make storage deals with miners with these pieces, on the side of storage, it will be pretty efficient and allow us to store hundreds of TiB data in a month. However, this way will also bring difficulties for data retrieval. Even if we only needed to retrieve a small file, we would first have to retrieve and download all the pieces of this tarball, decompress it, and find the specific file we needed.
Graphsplit can solve this problem. It takes advantage of IPLD protocol, follows the Unixfs format data structures, and regards the dataset or its sub-directory as a big graph, then cuts it into small graphs. Each small graph will keep its file system structure as possible as it used to be. After that, we only need to organize these small graphs into a car file. If one data piece has a complete file and we need to retrieve it, we only need to use payload CID to retrieve it through the lotus client, fetch it back, and get the file. Besides, Graphsplit will create a manifest.csv to save the mapping with graph slice name, payload CID, Piece CID, and the inner file structure.
Another advantage of Graphsplit is it can perfectly match IPFS. Like if you build an IPFS website as your Deal UI website, the inner file structure of each data piece can be shown on it, and it is easier for users to retrieve and download the data they stored.
git clone https://github.com/filedrive-team/go-graphsplit.git
cd go-graphsplit
# get submodules
git submodule update --init --recursive
# build filecoin-ffi
make ffi
make
See the work flow of graphsplit
Splitting dataset:
./graphsplit chunk \
# car-dir: folder for splitted smaller pieces, in form of .car
--car-dir=path/to/car-dir \
# slice-size: size for each pieces
--slice-size=17179869184 \
# parallel: number goroutines run when building ipld nodes
--parallel=2 \
# graph-name: it will use graph-name for prefix of smaller pieces
--graph-name=gs-test \
# calc-commp: calculation of pieceCID, default value is false. Be careful, a lot of cpu, memory and time would be consumed if slice size is very large.
--calc-commp=false \
# set true if want padding the car file to fit piece size
--add-padding=false \
# set true if want using piececid to name the chunk file
--rename=false \
# parent-path: usually just be the same as /path/to/dataset, it's just a method to figure out relative path when building IPLD graph
--parent-path=/path/to/dataset \
# whether add padding file to car file if car file size is smaller than slice-size
--car-padding=false \
# whether write car file descriptions to json file after car file generated, only valid when '--calc-commp=true'
--car-json=false \
# whether calculate car file md5, only valid when '--car-json=true'
--car-md5=false \
# support one or more target paths
/path/to/dataset /path/to/dataset2
Notes: A manifest.csv will created to save the mapping with graph slice name, the payload cid and slice inner structure. As following:
cat /path/to/car-dir/manifest.csv
payload_cid,filename,detail
ba...,graph-slice-name.car,inner-structure-json
If set --calc-commp=true, two another fields would be add to manifest.csv
cat /path/to/car-dir/manifest.csv
payload_cid,filename,piece_cid,piece_size,detail
ba...,graph-slice-name.car,baga...,16646144,inner-structure-json
Import car file to IPFS:
ipfs dag import /path/to/car-dir/car-file
Restore files:
# car-path: directory or file, in form of .car
# output-dir: usually just be the same as /path/to/output-dir
# parallel: number goroutines run when restoring
./graphsplit restore \
--car-path=/path/to/car-path \
--output-dir=/path/to/output-dir \
--parallel=2 \
# car-padding: if true, padding file will be removed after file restored
--car-padding=true
PieceCID Calculation for a single car file:
# Calculate pieceCID for a single car file
#
./graphsplit commP /path/to/carfile
PRs are welcome!
MIT