#graph #graph-node #graph-traversal #sha-1 #webgraph #software-heritage

bin+lib swh-graph

Compressed in-memory representation of the Software Heritage archive graph

6 stable releases

6.2.0 Oct 29, 2024
6.1.0 Oct 14, 2024
6.0.0 Sep 10, 2024
5.1.1 Aug 23, 2024

#74 in Compression

Download history 62/week @ 2024-08-03 147/week @ 2024-08-10 105/week @ 2024-08-17 50/week @ 2024-08-24 5/week @ 2024-08-31 138/week @ 2024-09-07 43/week @ 2024-09-14 13/week @ 2024-09-21 20/week @ 2024-09-28 9/week @ 2024-10-05 176/week @ 2024-10-12 20/week @ 2024-10-19 120/week @ 2024-10-26 19/week @ 2024-11-02 11/week @ 2024-11-09 16/week @ 2024-11-16

166 downloads per month
Used in swh-graph-grpc-server

GPL-3.0-or-later

545KB
11K SLoC

Software Heritage - graph service

Overview

swh_graph is a library and server allowing fast traversal queries through the Software Heritage archive, by using an in-memory compressed representation of the Software Heritage DAG based on webgraph.

The first version of this library and server was written in Java; and this documentation describes the Rust rewrite, which implements the same concepts and algorithms, but streamlined and faster.

Every node is identified by an integer from 0 to n-1 where n is the total number of nodes in a graph

The provided API is:

  • given a SWHID v1.1 return a node id, and vice versa
  • likewise for origins, encoded as swh:1:ori:$hash, where $hash is the SHA1 hash of the origin's URL
  • given a node id, list all its successors, such as all the entries in a directory, or all parents of a revision/commit (note that a "parent revision" is a successor here when seen as a directed arc in the graph). Some arcs may have labels, such as the name of a directory entry or a branch.
  • given a node id, list all its predecessors, such as all the directories containing a given directory entries in a directory, or all children of revision/commit. Some arcs may have labels too.
  • given a node id, return properties of this node, such as:
    • its type
    • a message or date if the node is a revision or a release, or
    • a length if the node is a content

swh_graph also provides an implementation of a handful of recursive traversal queries on top of these building blocks, as a gRPC API

See the Tutorial, or the Crash Course if you are already familiar with the Java API.

Building a distribution

Dependencies

The code needs stable Rust to be >= 1.79 because we depend webgraph uses associated type bounds.

We also need protoc to be installed (apt-get install protobuf-compiler on Debian) if you want to build the gRPC server (swh-graph-grpc-server crate)

Distribution

To distribute executables, you can statically compile the code with:

RUSTFLAGS="-C target-cpu=x86-64-v3" cargo build --release --target x86_64-unknown-linux-musl

To add the target use:

rustup target add x86_64-unknown-linux-musl

The target-cpu will limit the compatible cpus but will enable more optimizations. Some Interesting architecture are:

  • native for the compiling CPU architecture, this is the best option for performance when you are compiling and running on the same machine.
  • x86-64-v3 for Intel Haswell and newer (2013), oldest architecture that supports AVX2 and BMI2 instructions.
  • x86-64-v2 for Intel Core 2 and newer (2006), oldest reasonable architecture to compile for.

Performance consideration

At every random access, we need to query an Elias-Fano data structure to find the bit-offset at which the codes of the given nodes start. This operation has to find the i-th one in a given word of memory.

The implementation in sux-rs can exploit the pdep instruction to speed up the operation. So it's important to compile the code with the pdep feature enabled, and generally targeting the intended CPU architecture by using the RUSTFLAGS environment variable as:

RUSTFLAGS="-C target-cpu=native" cargo run --release --bin bfs

this compiles targeting the compiling CPU architecture.

For this reason, this is enabled by default in the .cargo/config.toml file. Be aware that a file compiled with pdep enabled will not run on a CPU that does not support it. Generally, running a binary compiled with a given set of CPU features on another one will result in SIGILL (illegal instruction) error.

Minimal setup for tests

Here is the minimal setup you will need to be able to execute python tests in this package. This is not the recommended build process for production-like deployments!

  1. install binary dependencies
swh-graph$ sudo apt install protobuf-compiler
swh-graph$ curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

The Rust compiler is installed from upstream because Debian stable (bookworm) only provides rust 1.63.

  1. compile the binary assets:
swh-graph$ cargo build --all-features

This should build binary assets in the local directory target/debug. Check the swh-graph-index is present there and cen be executed properly:

swh-graph$ ./target/debug/swh-graph-index --help
Commands to (re)generate `.ef` and `.offsets` files, allowing random access to BVGraph

Usage: swh-graph-index <COMMAND>

Commands:
  offsets    Reads a graph file linearly and produce a .offsets file which can be used by the Java backend to randomly access the graph
  ef         Reads either a graph file linearly or .offsets file (generated and used by the Java backend to randomly access the graph), and produces a .ef file suitable to randomly access the graph from the Rust backend
  labels-ef  Reads either a graph file linearly or .offsets file (generated and used by the Java backend to randomly access the graph), and produces a .ef file suitable to randomly access the graph from the Rust backend
  dcf        Reads either a graph file linearly, and produces a degree-cumulative function encoded as an Elias-Fano sequence in a .dcf file, suitable to distribute load while working on the graph
  help       Print this message or the help of the given subcommand(s)

Options:
  -h, --help  Print help

Deployment

Downloading a compressed graph

See the Software Heritage Dataset documentation, which lists all the exports from the Software Heritage archive; most of them provide a compressed graph representation that can be loaded with swh_graph (though some may need some conversion, see below).

Sharing mapped data across processes

Most of the data files are mmapped. You should pre-load some of them in memory to ensure they are not evicted from the in-memory cache.

Loading old graphs

The original Java (and C++) implementation of WebGraph used slightly different data structures and formats than the Rust implementation. Therefore, you need to generate new files in order to load old graphs with the Rust implementation; this takes a few hours for graphs representing full SWH exports.

The swh graph reindex command (made available with pip3 install swh.graph) takes care of running the conversion.

Additionally, the .ef format may change from time to time. If you get an error like this:

Error: Cannot map Elias-Fano pointer list ../swh/graph/example_dataset/compressed/example.ef

Caused by:
    Wrong type hash. Expected: 0x47e8ca1ab8fa94f1 Actual: 0x890ce77a9258940c.
    You are trying to deserialize a file with the wrong type.
    The serialized type is 'sux::dict::elias_fano::EliasFano<sux::rank_sel::select_fixed2::SelectFixed2<sux::bits::bit_vec::CountBitVec, alloc::vec::Vec<u64>, 8>>' and the deserialized type is 'sux::dict::elias_fano::EliasFano<sux::rank_sel::select_adapt_const::SelectAdaptConst<sux::bits::bit_vec::BitVec<alloc::boxed::Box<[usize]>>, alloc::boxed::Box<[usize]>, 12, 4>>'.

you need run swh graph reindex --ef to re-create them at the current version on your system.

Dependencies

~34–74MB
~1M SLoC