Every supported Key-Value Store has it's own wrapper:

• engine_ucset.cpp for unum-cloud/ucset in-memory AVL trees,
• engine_leveldb.cpp for google/leveldb persistent B trees,
• engine_rocksdb.cpp for facebook/rocksdb persistent LSM trees.

Every server protocol has it's own implementation file:

• flight_server.cpp for Apache Arrow Flight RPC server,
• rest_server.cpp for now deprecated REST server,
• json_rpc_server.c for the minimalistic JSON-RPC server.

Some protocols, come with pre-built wrappers:

• flight_client.cpp for Apache Arrow Flight RPC client.

On top of the binary layer, every structured modality has it's own serialization:

• modality_docs.cpp for JSON, BSON and MessagePack documents,
• modality_graph.cpp for Directed Multi-Graphs,
• modality_vectors.cpp for Approximate Vector Search,
• modality_paths.cpp for String and Path-like keys.

All implementations of all modalities try to avoid dynamic memory allocations. Every call uses its own arena, until he needs to grow beyond it. Largely for that reason we chose to use the following libraries for the implementation of logic across different modalities:

• For Documents:
  • yyjson: to update the state of documents.
  • simdjson: to sample fields from immutable documents.
  • mongo-c-driver: to support BSON inputs.
  • zlib for document compression. 🔜
• For Paths:
  • pcre2: to JIT Regular Expressions.
• For Graphs:
  • turbopfor for graph compression. 🔜

More broadly:

• jemalloc for faster NUMA-aware allocations.
• arrow for shared memory columnar representations.
• arrow_flight for gRPC implementation.
• fmt for string formatting of gRPC requests.