SurrealDB FFI Codec
Overview
An efficient binary codec pattern for accessing SurrealDB Embedded from other languages (Go, Swift, Kotlin, etc.) via FFI.
Core principle: eliminate JSON and optimize with a two‑layer structure:
- FlatBuffers: fixed schema (error codes, status, operation types)
- MessagePack: dynamic schema (query results, parameters)
When to Use
- Embedding SurrealDB in Rust and accessing it from other languages via FFI
- Eliminating JSON parse/serialize overhead
- Handling dynamic query results while preserving type safety
- Requiring zero‑copy access through FlatBuffers
When NOT to Use
- HTTP/WebSocket client implementations (use official SurrealDB SDKs)
- Server-side API development without FFI requirements
- Simple single-language Rust projects (use surrealdb crate directly)
- MCP server implementations (this skill provides templates, not runtime)
Important: This skill provides implementation templates and guidelines only, not executable binaries or runtime libraries.
Examples
User request: "Create an FFI layer to access SurrealDB from Go"
→ Generate Rust library with:
- FlatBuffers schemas from
resources/schemas/ - Codec implementation from
resources/codec/ - C ABI wrapper from
resources/ffi/ - Go client code following
docs/integration_guide.md
User request: "Convert SurrealDB query results to MessagePack"
→ Implement converter using resources/codec/converter.rs.template with proper TypeHints handling.
Conversion Flow
See docs/design_principle.md for detailed architecture. Summary:
sql::Value → rmpv::Value → [u8] bytes (in FlatBuffers payload)
Type Mapping
See docs/conversion_table.md for the complete type conversion reference.
Key points:
sql::Number::Decimal→f64(precision loss beyond 15-17 digits)sql::Datetime→ RFC3339 string (removed'...'wrapper)sql::Thing→ ID-only string (remove table prefix and brackets)
Instructions
Project Layout
Generate the following structure:
rust_lib/
├── Cargo.toml
├── schemas/
│ ├── request.fbs
│ └── response.fbs
└── src/
├── lib.rs
├── codec/
│ ├── mod.rs
│ ├── converter.rs
│ ├── msgpack.rs
│ └── generated/
└── ffi/
├── mod.rs
├── wrapper.rs
└── error_codes.rs
Implementation
- FlatBuffers schemas - Use
resources/schemas/*.fbs.template, replace{{NAMESPACE}} - Codec - Use
resources/codec/*.rs.templatefor type conversion - FFI wrapper - Use
resources/ffi/*.rs.templatefor C ABI exports
FlatBuffers Code Generation
flatc --rust -o src/codec/generated schemas/request.fbs schemas/response.fbs
Dependencies
[dependencies]
surrealdb = { version = "2.x", features = ["kv-mem"] } # or kv-rocksdb
flatbuffers = "24.x"
rmp-serde = "1.x"
rmpv = "1.x"
chrono = { version = "0.4", features = ["serde"] }
thiserror = "2.x"
log = "0.4"
Guidelines
Using TypeHints
Convert string parameters from clients into native SurrealDB types:
// Formats: "field_name" or "datetime:field_name" or "record:table:field"
let hints = TypeHints::parse(&[
"created_at".to_string(),
"record:user:owner_id".to_string(),
]);
let surreal_value = to_surreal_value_with_type_hints(msgpack_value, &hints);
Decimal Precision Loss
sql::Number::Decimal is converted to f64, so precision beyond 15–17 digits is lost. If exact precision is required, treat it as a string.
Thing ID Cleanup
SurrealDB 2.x wraps Thing IDs in angle brackets. Remove them during conversion:
// "user:⟨uuid-here⟩" -> "uuid-here"
let clean_id = id.trim_matches(|c| c == '\u{27E8}' || c == '\u{27E9}');
Error Handling
- status_code = 0: success
- status_code = -1: Rust panic (fatal)
- status_code > 0: application error
All FFI functions must wrap logic with std::panic::catch_unwind to prevent panics from crossing the FFI boundary.
Async Runtime Management
FFI functions are synchronous but SurrealDB SDK is async. Use OnceLock for global runtime singleton. See docs/integration_guide.md for the complete pattern.
Applying RecordFieldHint
FlatBuffers record_fields: [RecordFieldHint] should be mapped into TypeHints.record_fields.
record_fields = [
{ field: "owner_id", table_name: "user" }
]
The above is equivalent to record:user:owner_id. Ensure client hints and Rust TypeHints are consistent.
FFI Input Safety
- If
request_ptris null orrequest_len == 0, returnInvalidRequestimmediately. - FFI callers must always invoke
free_response_bufferand avoid double‑free.
Error Payload Format
Use a structured MessagePack payload on error:
{ "code": i32, "message": String, "details": Option<String> }
Clients should ignore unknown fields for forward compatibility.