ton/NATSBridge
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity
Files
1d5c6f334813912a8ec6ac02c85076906542e958
NATSBridge/docs/architecture.md
narawat 1d5c6f3348 update
2026-02-17 21:04:32 +07:00

584 lines
20 KiB
Markdown
Raw Blame History

# Architecture Documentation: Bi-Directional Data Bridge (Julia ↔ JavaScript)
## Overview
This document describes the architecture for a high-performance, bi-directional data bridge between a Julia service and a JavaScript (Node.js) service using NATS (Core & JetStream), implementing the Claim-Check pattern for large payloads.
### File Server Handler Architecture
The system uses **handler functions** to abstract file server operations, allowing support for different file server implementations (e.g., Plik, AWS S3, custom HTTP server).
**Handler Function Signatures:**
```julia
# Upload handler - uploads data to file server and returns URL
# The handler is passed to smartsend as fileserverUploadHandler parameter
# It receives: (fileserver_url::String, dataname::String, data::Vector{UInt8})
# Returns: Dict{String, Any} with keys: "status", "uploadid", "fileid", "url"
fileserverUploadHandler(fileserver_url::String, dataname::String, data::Vector{UInt8})::Dict{String, Any}
# Download handler - fetches data from file server URL with exponential backoff
# The handler is passed to smartreceive as fileserverDownloadHandler parameter
# It receives: (url::String, max_retries::Int, base_delay::Int, max_delay::Int, correlation_id::String)
# Returns: Vector{UInt8} (the downloaded data)
fileserverDownloadHandler(url::String, max_retries::Int, base_delay::Int, max_delay::Int, correlation_id::String)::Vector{UInt8}
```
This design allows the system to support multiple file server backends without changing the core messaging logic.
### Multi-Payload Support (Standard API)
The system uses a **standardized list-of-tuples format** for all payload operations. **Even when sending a single payload, the user must wrap it in a list.**
**API Standard:**
```julia
# Input format for smartsend (always a list of tuples with type info)
[(dataname1, data1, type1), (dataname2, data2, type2), ...]
# Output format for smartreceive (always returns a list of tuples)
[(dataname1, data1), (dataname2, data2), ...]
```
**Supported Types:**
- `"text"` - Plain text
- `"dictionary"` - JSON-serializable dictionaries (Dict, NamedTuple)
- `"table"` - Tabular data (DataFrame, array of structs)
- `"image"` - Image data (Bitmap, PNG/JPG bytes)
- `"audio"` - Audio data (WAV, MP3 bytes)
- `"video"` - Video data (MP4, AVI bytes)
- `"binary"` - Generic binary data (Vector{UInt8})
This design allows per-payload type specification, enabling **mixed-content messages** where different payloads can use different serialization formats in a single message.
**Examples:**
```julia
# Single payload - still wrapped in a list
smartsend(
"/test",
[("dataname1", data1, "dictionary")], # List with one tuple (data, type)
nats_url="nats://localhost:4222",
fileserverUploadHandler=plik_oneshot_upload,
metadata=user_provided_envelope_level_metadata
)
# Multiple payloads in one message with different types
smartsend(
"/test",
[("dataname1", data1, "dictionary"), ("dataname2", data2, "table")],
nats_url="nats://localhost:4222",
fileserverUploadHandler=plik_oneshot_upload
)
# Mixed content (e.g., chat with text, image, audio)
smartsend(
"/chat",
[
("message_text", "Hello!", "text"),
("user_image", image_data, "image"),
("audio_clip", audio_data, "audio")
],
nats_url="nats://localhost:4222"
)
# Receive always returns a list
payloads = smartreceive(msg, fileserverDownloadHandler, max_retries, base_delay, max_delay)
# payloads = [("dataname1", data1), ("dataname2", data2), ...]
```
## Architecture Diagram
```mermaid
flowchart TD
subgraph Client
JS[JavaScript Client]
JSApp[Application Logic]
end
subgraph Server
Julia[Julia Service]
NATS[NATS Server]
FileServer[HTTP File Server]
end
JS -->|Control/Small Data| JSApp
JSApp -->|NATS| NATS
NATS -->|NATS| Julia
Julia -->|NATS| NATS
Julia -->|HTTP POST| FileServer
JS -->|HTTP GET| FileServer
style JS fill:#e1f5fe
style Julia fill:#e8f5e9
style NATS fill:#fff3e0
style FileServer fill:#f3e5f5
```
## System Components
### 1. msgEnvelope_v1 - Message Envelope
The `msgEnvelope_v1` structure provides a comprehensive message format for bidirectional communication between Julia and JavaScript services.
**Julia Structure:**
```julia
struct msgEnvelope_v1
correlationId::String # Unique identifier to track messages across systems
msgId::String # This message id
timestamp::String # Message published timestamp
sendTo::String # Topic/subject the sender sends to
msgPurpose::String # Purpose of this message (ACK | NACK | updateStatus | shutdown | ...)
senderName::String # Sender name (e.g., "agent-wine-web-frontend")
senderId::String # Sender id (uuid4)
receiverName::String # Message receiver name (e.g., "agent-backend")
receiverId::String # Message receiver id (uuid4 or nothing for broadcast)
replyTo::String # Topic to reply to
replyToMsgId::String # Message id this message is replying to
brokerURL::String # NATS server address
metadata::Dict{String, Any}
payloads::AbstractArray{msgPayload_v1} # Multiple payloads stored here
end
```
**JSON Schema:**
```json
{
"correlationId": "uuid-v4-string",
"msgId": "uuid-v4-string",
"timestamp": "2024-01-15T10:30:00Z",
"sendTo": "topic/subject",
"msgPurpose": "ACK | NACK | updateStatus | shutdown | chat",
"senderName": "agent-wine-web-frontend",
"senderId": "uuid4",
"receiverName": "agent-backend",
"receiverId": "uuid4",
"replyTo": "topic",
"replyToMsgId": "uuid4",
"brokerURL": "nats://localhost:4222",
"metadata": {
},
"payloads": [
{
"id": "uuid4",
"dataname": "login_image",
"type": "image",
"transport": "direct",
"encoding": "base64",
"size": 15433,
"data": "base64-encoded-string",
"metadata": {
}
},
{
"id": "uuid4",
"dataname": "large_data",
"type": "table",
"transport": "link",
"encoding": "none",
"size": 524288,
"data": "http://localhost:8080/file/UPLOAD_ID/FILE_ID/data.arrow",
"metadata": {
}
}
]
}
```
### 2. msgPayload_v1 - Payload Structure
The `msgPayload_v1` structure provides flexible payload handling for various data types.
**Julia Structure:**
```julia
struct msgPayload_v1
id::String # Id of this payload (e.g., "uuid4")
dataname::String # Name of this payload (e.g., "login_image")
type::String # "text | dictionary | table | image | audio | video | binary"
transport::String # "direct | link"
encoding::String # "none | json | base64 | arrow-ipc"
size::Integer # Data size in bytes
data::Any # Payload data in case of direct transport or a URL in case of link
metadata::Dict{String, Any} # Dict("checksum" => "sha256_hash", ...)
end
```
**Key Features:**
- Supports multiple data types: text, dictionary, table, image, audio, video, binary
- Flexible transport: "direct" (NATS) or "link" (HTTP fileserver)
- Multiple payloads per message (essential for chat with mixed content)
- Per-payload and per-envelope metadata support
### 3. Transport Strategy Decision Logic
```
┌─────────────────────────────────────────────────────────────┐
│ smartsend Function │
│ Accepts: [(dataname1, data1, type1), ...] │
│ (No standalone type parameter - type per payload) │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ For each payload: │
│ 1. Extract type from tuple │
│ 2. Serialize based on type │
│ 3. Check payload size │
└─────────────────────────────────────────────────────────────┘
│
┌────────────────┴─-────────────────┐
▼ ▼
┌─────────────────┐ ┌─────────────────┐
│ Direct Path │ │ Link Path │
│ (< 1MB) │ │ (> 1MB) │
│ │ │ │
│ • Serialize to │ │ • Serialize to │
│ IOBuffer │ │ IOBuffer │
│ • Base64 encode │ │ • Upload to │
│ • Publish to │ │ HTTP Server │
│ NATS │ │ • Publish to │
│ (with payload │ │ NATS with URL │
│ in envelope) │ │ (in envelope) │
└─────────────────┘ └─────────────────┘
```
### 4. Julia Module Architecture
```mermaid
graph TD
subgraph JuliaModule
smartsendJulia[smartsend Julia]
SizeCheck[Size Check]
DirectPath[Direct Path]
LinkPath[Link Path]
HTTPClient[HTTP Client]
end
smartsendJulia --> SizeCheck
SizeCheck -->|< 1MB| DirectPath
SizeCheck -->|>= 1MB| LinkPath
LinkPath --> HTTPClient
style JuliaModule fill:#c5e1a5
```
### 5. JavaScript Module Architecture
```mermaid
graph TD
subgraph JSModule
smartsendJS[smartsend JS]
smartreceiveJS[smartreceive JS]
JetStreamConsumer[JetStream Pull Consumer]
ApacheArrow[Apache Arrow]
end
smartsendJS --> NATS
smartreceiveJS --> JetStreamConsumer
JetStreamConsumer --> ApacheArrow
style JSModule fill:#f3e5f5
```
## Implementation Details
### Julia Implementation
#### Dependencies
- `NATS.jl` - Core NATS functionality
- `Arrow.jl` - Arrow IPC serialization
- `JSON3.jl` - JSON parsing
- `HTTP.jl` - HTTP client for file server
- `Dates.jl` - Timestamps for logging
#### smartsend Function
```julia
function smartsend(
subject::String,
data::AbstractArray{Tuple{String, Any, String}}; # No standalone type parameter
nats_url::String = "nats://localhost:4222",
fileserverUploadHandler::Function = plik_oneshot_upload,
size_threshold::Int = 1_000_000 # 1MB
)
```
**Input Format:**
- `data::AbstractArray{Tuple{String, Any, String}}` - **Must be a list of (dataname, data, type) tuples**: `[("dataname1", data1, "type1"), ("dataname2", data2, "type2"), ...]`
- Even for single payloads: `[(dataname1, data1, "type1")]`
- Each payload can have a different type, enabling mixed-content messages
**Flow:**
1. Iterate through the list of `(dataname, data, type)` tuples
2. For each payload: extract the type from the tuple and serialize accordingly
3. Check payload size
4. If < threshold: publish directly to NATS with Base64-encoded payload
5. If >= threshold: upload to HTTP server, publish NATS with URL
#### smartreceive Handler
```julia
function smartreceive(
msg::NATS.Message,
fileserverDownloadHandler::Function=_fetch_with_backoff;
max_retries::Int = 5,
base_delay::Int = 100,
max_delay::Int = 5000
)
# Parse envelope
# Iterate through all payloads
# For each payload: check transport type
# If direct: decode Base64 payload
# If link: fetch from URL with exponential backoff using fileserverDownloadHandler
# Deserialize payload based on type
# Return list of (dataname, data) tuples
end
```
**Output Format:**
- Always returns a list of tuples: `[(dataname1, data1), (dataname2, data2), ...]`
- Even for single payloads: `[(dataname1, data1)]`
**Process Flow:**
1. Parse the JSON envelope to extract the `payloads` array
2. Iterate through each payload in `payloads`
3. For each payload:
- Determine transport type (`direct` or `link`)
- If `direct`: decode Base64 data from the message
- If `link`: fetch data from URL using exponential backoff (via `fileserverDownloadHandler`)
- Deserialize based on payload type (`dictionary`, `table`, `binary`, etc.)
4. Return list of `(dataname, data)` tuples
**Note:** The `fileserverDownloadHandler` receives `(url::String, max_retries::Int, base_delay::Int, max_delay::Int, correlation_id::String)` and returns `Vector{UInt8}`.
### JavaScript Implementation
#### Dependencies
- `nats.js` - Core NATS functionality
- `apache-arrow` - Arrow IPC serialization
- `uuid` - Correlation ID generation
#### smartsend Function
```javascript
async function smartsend(subject, data, options = {})
// data format: [(dataname, data, type), ...]
// options object should include:
// - natsUrl: NATS server URL
// - fileserverUrl: base URL of the file server
// - sizeThreshold: threshold in bytes for transport selection
// - correlationId: optional correlation ID for tracing
```
**Input Format:**
- `data` - **Must be a list of (dataname, data, type) tuples**: `[(dataname1, data1, "type1"), (dataname2, data2, "type2"), ...]`
- Even for single payloads: `[(dataname1, data1, "type1")]`
- Each payload can have a different type, enabling mixed-content messages
**Flow:**
1. Iterate through the list of (dataname, data, type) tuples
2. For each payload: extract the type from the tuple and serialize accordingly
3. Check payload size
4. If < threshold: publish directly to NATS
5. If >= threshold: upload to HTTP server, publish NATS with URL
#### smartreceive Handler
```javascript
async function smartreceive(msg, options = {})
// options object should include:
// - fileserverDownloadHandler: function to fetch data from file server URL
// - max_retries: maximum retry attempts for fetching URL
// - base_delay: initial delay for exponential backoff in ms
// - max_delay: maximum delay for exponential backoff in ms
// - correlationId: optional correlation ID for tracing
```
**Process Flow:**
1. Parse the JSON envelope to extract the `payloads` array
2. Iterate through each payload in `payloads`
3. For each payload:
- Determine transport type (`direct` or `link`)
- If `direct`: decode Base64 data from the message
- If `link`: fetch data from URL using exponential backoff
- Deserialize based on payload type (`dictionary`, `table`, `binary`, etc.)
4. Return list of `(dataname, data)` tuples
## Scenario Implementations
### Scenario 1: Command & Control (Small Dictionary)
**Julia (Receiver):**
```julia
# Subscribe to control subject
# Parse JSON envelope
# Execute simulation with parameters
# Send acknowledgment
```
**JavaScript (Sender):**
```javascript
// Create small dictionary config
// Send via smartsend with type="dictionary"
```
### Scenario 2: Deep Dive Analysis (Large Arrow Table)
**Julia (Sender):**
```julia
# Create large DataFrame
# Convert to Arrow IPC stream
# Check size (> 1MB)
# Upload to HTTP server
# Publish NATS with URL
```
**JavaScript (Receiver):**
```javascript
// Receive NATS message with URL
// Fetch data from HTTP server
// Parse Arrow IPC with zero-copy
// Load into Perspective.js or D3
```
### Scenario 3: Live Audio Processing
**JavaScript (Sender):**
```javascript
// Capture audio chunk
// Send as binary with metadata headers
// Use smartsend with type="audio"
```
**Julia (Receiver):**
```julia
// Receive audio data
// Perform FFT or AI transcription
// Send results back (JSON + Arrow table)
```
### Scenario 4: Catch-Up (JetStream)
**Julia (Producer):**
```julia
# Publish to JetStream
# Include metadata for temporal tracking
```
**JavaScript (Consumer):**
```javascript
// Connect to JetStream
// Request replay from last 10 minutes
// Process historical and real-time messages
```
### Scenario 5: Selection (Low Bandwidth)
**Focus:** Small Arrow tables, Julia to JavaScript. The Action: Julia wants to send a small DataFrame to show on a JavaScript dashboard for the user to choose.
**Julia (Sender):**
```julia
# Create small DataFrame (e.g., 50KB - 500KB)
# Convert to Arrow IPC stream
# Check payload size (< 1MB threshold)
# Publish directly to NATS with Base64-encoded payload
# Include metadata for dashboard selection context
```
**JavaScript (Receiver):**
```javascript
// Receive NATS message with direct transport
// Decode Base64 payload
// Parse Arrow IPC with zero-copy
// Load into selection UI component (e.g., dropdown, table)
// User makes selection
// Send selection back to Julia
```
**Use Case:** Julia server generates a list of available options (e.g., file selections, configuration presets) as a small DataFrame and sends to JavaScript dashboard for user selection. The selection is then sent back to Julia for processing.
### Scenario 6: Chat System
**Focus:** Every conversational message is composed of any number and any combination of components, spanning the full spectrum from small to large. This includes text, images, audio, video, tables, and files—specifically accommodating everything from brief snippets to high-resolution images, large audio files, extensive tables, and massive documents. Support for claim-check delivery and full bi-directional messaging.
**Multi-Payload Support:** The system supports mixed-payload messages where a single message can contain multiple payloads with different transport strategies. The `smartreceive` function iterates through all payloads in the envelope and processes each according to its transport type.
**Julia (Sender/Receiver):**
```julia
# Build chat message with mixed payloads:
# - Text: direct transport (Base64)
# - Small images: direct transport (Base64)
# - Large images: link transport (HTTP URL)
# - Audio/video: link transport (HTTP URL)
# - Tables: direct or link depending on size
# - Files: link transport (HTTP URL)
#
# Each payload uses appropriate transport strategy:
# - Size < 1MB → direct (NATS + Base64)
# - Size >= 1MB → link (HTTP upload + NATS URL)
#
# Include claim-check metadata for delivery tracking
# Support bidirectional messaging with replyTo fields
```
**JavaScript (Sender/Receiver):**
```javascript
// Build chat message with mixed content:
// - User input text: direct transport
// - Selected image: check size, use appropriate transport
// - Audio recording: link transport for large files
// - File attachment: link transport
//
// Parse received message:
// - Direct payloads: decode Base64
// - Link payloads: fetch from HTTP with exponential backoff
// - Deserialize all payloads appropriately
//
// Render mixed content in chat interface
// Support bidirectional reply with claim-check delivery confirmation
```
**Use Case:** Full-featured chat system supporting rich media. User can send text, small images directly, or upload large files that get uploaded to HTTP server and referenced via URLs. Claim-check pattern ensures reliable delivery tracking for all message components.
**Implementation Note:** The `smartreceive` function iterates through all payloads in the envelope and processes each according to its transport type. See the standard API format in Section 1: `msgEnvelope_v1` supports `AbstractArray{msgPayload_v1}` for multiple payloads.
## Performance Considerations
### Zero-Copy Reading
- Use Arrow's memory-mapped file reading
- Avoid unnecessary data copying during deserialization
- Use Apache Arrow's native IPC reader
### Exponential Backoff
- Implement exponential backoff for HTTP link fetching
- Maximum retry count: 5
- Base delay: 100ms, max delay: 5000ms
### Correlation ID Logging
- Log correlation_id at every stage
- Include: send, receive, serialize, deserialize
- Use structured logging format
## Testing Strategy
### Unit Tests
- Test smartsend with various payload sizes
- Test smartreceive with direct and link transport
- Test Arrow IPC serialization/deserialization
### Integration Tests
- Test full flow with NATS server
- Test large data transfer (> 100MB)
- Test audio processing pipeline
### Performance Tests
- Measure throughput for small payloads
- Measure throughput for large payloads
Reference in New Issue View Git Blame Copy Permalink