Advanced similarity and duplicate source code proof of concept for our research efforts.
Switch branches/tags
Nothing to show
Clone or download
vmarkovtsev Merge pull request #71 from zurk/feature/skip-member-dco
Skipping sign-off for organization members
Latest commit dd520a9 Oct 10, 2018

README.md

Apollo

Advanced code deduplicator from hell. Powered by source{d} ML, source{d} engine and minhashcuda. Agnostic to the analysed language thanks to Babelfish. Python 3, PySpark, CUDA inside.

What is this?

source{d}'s effort to research and solve the code deduplication problem. At scale, as usual. A code clone is several snippets of code with few differences. For now this project focuses on find near-duplicate projects and files; it will eventually support functions and snippets in the future.

Should I use it?

If you've got hundreds of thousands of files or more, consider. Otherwise, use one of the many existing tools which may be already integrated into your IDE.

Difference from src-d/gemini?

This guy is my brother. Apollo focuses on research, extensibility, flexibility and rapid changes, while Gemini focuses on performance and serious production usage. All the proven and tested features will be eventually ported to Gemini. At the same time, Gemini may reuse some of Apollo's code.

Algorithm

Apollo takes the "hash'em all" approach. We extract unordered weighted features from code aka "weighted bags", apply Weighted MinHash and then design the Locality Sensitive Hashing index. All items which appear in the same hashtable bucket are considered the same. The size of the hash and the number of hashtables depend on the weighted Jaccard similarity threshold (hence Weighted MinHash).

The features include identifiers such as variable, function or class names, literal values and structural elements. The latter carries the topological information, and we currently support several variants: "node2vec", "deterministic node2vec" and "role-children atoms". Graphlets are upcoming. Different features have different weights which will be tuned by a hyperparameter optimization algorithm or even an SGD (not yet implemented).

It's not all unfortunately! Dumping the huge graph of pairwise similarities is of little practicality. We need to group (cluster) the neighborhoods of densely connected nodes. Apollo solves this problem in two steps:

  1. Run connected components analysis to find disjoint parts in the similarity graph.
  2. Run community detection to cluster the components. The clusters are with overlaps.

Implementation

Apollo is structured as a series of commands in CLI. It stores data in Cassandra (compatible with Scylla) and writes MinHashCuda batches on disk. Community detection is delegated to igraph.

  • resetdb (erases) and initializes a Cassandra keyspace.
  • bags extracts the features, stores them in the database and writes MinHashCuda batches on disk. Runs source{d} engine through PySpark.
  • hash performs the hashing, writes the hashtables to the database and hashing parameters on disk in Modelforge format.
  • cc fetches the buckets, runs the connected component analysis and writes the result on disk in Modelforge format. Uses PySpark.
  • cmd reads the connected components and performs the community detection (by default, walktrap). Uses PySpark.
  • query outputs items similar to the specified. In case of files, the path or the sha1 are accepted.
  • dumpcmd outputs the groups of similar items.

Installation

mount -o bind /path/to/sourced-ml bundle/ml
mount -o bind /path/to/spark-2.2.0-bin-hadoop2.7 bundle/spark
mount -o bind /path/to/sourced-engine bundle/engine
docker build -t srcd/apollo .
docker run --name scylla -p 9042:9042 -v /var/lib/scylla:/var/lib/scylla -d scylladb/scylla --developer-mode=1
docker run -it --rm --link scylla srcd/apollo resetdb --cassandra scylla
docker run -d --name bblfshd --privileged -p 9432:9432 -v /var/lib/bblfshd:/var/lib/bblfshd bblfsh/bblfshd
docker exec -it bblfshd bblfshctl driver install --all

You are going to need grip to instantly render Markdown reports in your browser. There multiple Docker options available, e.g. 1, 2, 3.

Contributions

...are welcome! See CONTRIBUTING and code of conduct.

License

GPL.

Docker command snippets

Bags

docker run -it --rm -v /path/to/io:/io --link bblfshd --link scylla srcd/apollo bags -r /io/siva \
--bow /io/bags/bow.asdf --docfreq /io/bags/docfreq.asdf -f id lit uast2seq --uast2seq-seq-len 4 \
-l Java Python -s 'local[*]' --min-docfreq 5 --bblfsh bblfshd --cassandra scylla --persist MEMORY_ONLY \
--config spark.executor.memory=4G spark.driver.memory=10G spark.driver.maxResultSize=4G

Hash

docker run -it --rm -v /path/to/io:/io --link scylla srcd/apollo hash /io/batches/bow*.asdf -p /io/bags/params.asdf \
-t 0.8 --cassandra scylla

Query sha1

docker run -it --rm -v /path/to/io:/io --link scylla srcd/apollo query -i <sha1> --precise \
--docfreq /io/bags/docfreq.asdf -t 0.8 --cassandra scylla

Query file

docker run -it --rm -v /path/to/io:/io -v .:/q --link bblfshd --link scylla srcd/apollo query \
-f /q/myfile.java --bblfsh bblfshd --cassandra scylla --precise --docfreq /io/docfreq.asdf \
--params /io/params.asdf -t 0.9 | grip -b -

Connected components

docker run -it --rm -v /path/to/io:/io --link scylla srcd/apollo cc -o /io/ccs.asdf

Dump connected components

docker run -it --rm -v /path/to/io:/io srcd/apollo dumpcc -o /io/ccs.asdf

Community detection

docker run -it --rm -v /path/to/io:/io srcd/apollo cmd -i /io/ccs.asdf -o /io/communities.asdf -s 'local[*]'

Dump communities (final report)

docker run -it --rm -v /path/to/io:/io srcd/apollo dumpcmd /io/communities.asdf | grip -b -