Add comprehensive documentation for CLAUDE and Nostr WebSocket skills
Some checks failed
Go / build (push) Has been cancelled
Go / release (push) Has been cancelled

- Introduced CLAUDE.md to provide guidance for working with the Claude Code repository, including project overview, build commands, testing guidelines, and performance considerations.
- Added INDEX.md for a structured overview of the strfry WebSocket implementation analysis, detailing document contents and usage.
- Created SKILL.md for the nostr-websocket skill, covering WebSocket protocol fundamentals, connection management, and performance optimization techniques.
- Included multiple reference documents for Go, C++, and Rust implementations of WebSocket patterns, enhancing the knowledge base for developers.
- Updated deployment and build documentation to reflect new multi-platform capabilities and pure Go build processes.
- Bumped version to reflect the addition of extensive documentation and resources for developers working with Nostr relays and WebSocket connections.
This commit is contained in:
2025-11-06 16:18:09 +00:00
parent 27f92336ae
commit d604341a27
16 changed files with 8542 additions and 0 deletions

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

View File

@@ -0,0 +1,921 @@
# C++ WebSocket Implementation for Nostr Relays (strfry patterns)
This reference documents high-performance WebSocket patterns from the strfry Nostr relay implementation in C++.
## Repository Information
- **Project:** strfry - High-performance Nostr relay
- **Repository:** https://github.com/hoytech/strfry
- **Language:** C++ (C++20)
- **WebSocket Library:** Custom fork of uWebSockets with epoll
- **Architecture:** Single-threaded I/O with specialized thread pools
## Core Architecture
### Thread Pool Design
strfry uses 6 specialized thread pools for different operations:
```
┌─────────────────────────────────────────────────────────────┐
│ Main Thread (I/O) │
│ - epoll event loop │
│ - WebSocket message reception │
│ - Connection management │
└─────────────────────────────────────────────────────────────┘
┌───────────────────┼───────────────────┐
│ │ │
┌────▼────┐ ┌───▼────┐ ┌───▼────┐
│Ingester │ │ReqWorker│ │Negentropy│
│ (3) │ │ (3) │ │ (2) │
└─────────┘ └─────────┘ └─────────┘
│ │ │
┌────▼────┐ ┌───▼────┐
│ Writer │ │ReqMonitor│
│ (1) │ │ (3) │
└─────────┘ └─────────┘
```
**Thread Pool Responsibilities:**
1. **WebSocket (1 thread):** Main I/O loop, epoll event handling
2. **Ingester (3 threads):** Event validation, signature verification, deduplication
3. **Writer (1 thread):** Database writes, event storage
4. **ReqWorker (3 threads):** Process REQ subscriptions, query database
5. **ReqMonitor (3 threads):** Monitor active subscriptions, send real-time events
6. **Negentropy (2 threads):** NIP-77 set reconciliation
**Deterministic thread assignment:**
```cpp
int threadId = connId % numThreads;
```
**Benefits:**
- **No lock contention:** Shared-nothing architecture
- **Predictable performance:** Same connection always same thread
- **CPU cache efficiency:** Thread-local data stays hot
### Connection State
```cpp
struct ConnectionState {
uint64_t connId; // Unique connection identifier
std::string remoteAddr; // Client IP address
// Subscription state
flat_str subId; // Current subscription ID
std::shared_ptr<Subscription> sub; // Subscription filter
uint64_t latestEventSent = 0; // Latest event ID sent
// Compression state (per-message deflate)
PerMessageDeflate pmd;
// Parsing state (reused buffer)
std::string parseBuffer;
// Signature verification context (reused)
secp256k1_context *secpCtx;
};
```
**Key design decisions:**
1. **Reusable parseBuffer:** Single allocation per connection
2. **Persistent secp256k1_context:** Expensive to create, reused for all signatures
3. **Connection ID:** Enables deterministic thread assignment
4. **Flat string (flat_str):** Value-semantic string-like type for zero-copy
## WebSocket Message Reception
### Main Event Loop (epoll)
```cpp
// Pseudocode representation of strfry's I/O loop
uWS::App app;
app.ws<ConnectionState>("/*", {
.compression = uWS::SHARED_COMPRESSOR,
.maxPayloadLength = 16 * 1024 * 1024,
.idleTimeout = 120,
.maxBackpressure = 1 * 1024 * 1024,
.upgrade = nullptr,
.open = [](auto *ws) {
auto *state = ws->getUserData();
state->connId = nextConnId++;
state->remoteAddr = getRemoteAddress(ws);
state->secpCtx = secp256k1_context_create(SECP256K1_CONTEXT_VERIFY);
LI << "New connection: " << state->connId << " from " << state->remoteAddr;
},
.message = [](auto *ws, std::string_view message, uWS::OpCode opCode) {
auto *state = ws->getUserData();
// Reuse parseBuffer to avoid allocation
state->parseBuffer.assign(message.data(), message.size());
try {
// Parse JSON (nlohmann::json)
auto json = nlohmann::json::parse(state->parseBuffer);
// Extract command type
auto cmdStr = json[0].get<std::string>();
if (cmdStr == "EVENT") {
handleEventMessage(ws, std::move(json));
}
else if (cmdStr == "REQ") {
handleReqMessage(ws, std::move(json));
}
else if (cmdStr == "CLOSE") {
handleCloseMessage(ws, std::move(json));
}
else if (cmdStr == "NEG-OPEN") {
handleNegentropyOpen(ws, std::move(json));
}
else {
sendNotice(ws, "unknown command: " + cmdStr);
}
}
catch (std::exception &e) {
sendNotice(ws, "Error: " + std::string(e.what()));
}
},
.close = [](auto *ws, int code, std::string_view message) {
auto *state = ws->getUserData();
LI << "Connection closed: " << state->connId
<< " code=" << code
<< " msg=" << std::string(message);
// Cleanup
secp256k1_context_destroy(state->secpCtx);
cleanupSubscription(state->connId);
},
});
app.listen(8080, [](auto *token) {
if (token) {
LI << "Listening on port 8080";
}
});
app.run();
```
**Key patterns:**
1. **epoll-based I/O:** Single thread handles thousands of connections
2. **Buffer reuse:** `state->parseBuffer` avoids allocation per message
3. **Move semantics:** `std::move(json)` transfers ownership to handler
4. **Exception handling:** Catches parsing errors, sends NOTICE
### Message Dispatch to Thread Pools
```cpp
void handleEventMessage(auto *ws, nlohmann::json &&json) {
auto *state = ws->getUserData();
// Pack message with connection ID
auto msg = MsgIngester{
.connId = state->connId,
.payload = std::move(json),
};
// Dispatch to Ingester thread pool (deterministic assignment)
tpIngester->dispatchToThread(state->connId, std::move(msg));
}
void handleReqMessage(auto *ws, nlohmann::json &&json) {
auto *state = ws->getUserData();
// Pack message
auto msg = MsgReq{
.connId = state->connId,
.payload = std::move(json),
};
// Dispatch to ReqWorker thread pool
tpReqWorker->dispatchToThread(state->connId, std::move(msg));
}
```
**Message passing pattern:**
```cpp
// ThreadPool::dispatchToThread
void dispatchToThread(uint64_t connId, Message &&msg) {
size_t threadId = connId % threads.size();
threads[threadId]->queue.push(std::move(msg));
}
```
**Benefits:**
- **Zero-copy:** `std::move` transfers ownership without copying
- **Deterministic:** Same connection always processed by same thread
- **Lock-free:** Each thread has own queue
## Event Ingestion Pipeline
### Ingester Thread Pool
```cpp
void IngesterThread::run() {
while (running) {
Message msg;
if (!queue.pop(msg, 100ms)) continue;
// Extract event from JSON
auto event = parseEvent(msg.payload);
// Validate event ID
if (!validateEventId(event)) {
sendOK(msg.connId, event.id, false, "invalid: id mismatch");
continue;
}
// Verify signature (using thread-local secp256k1 context)
if (!verifySignature(event, secpCtx)) {
sendOK(msg.connId, event.id, false, "invalid: signature verification failed");
continue;
}
// Check for duplicate (bloom filter + database)
if (isDuplicate(event.id)) {
sendOK(msg.connId, event.id, true, "duplicate: already have this event");
continue;
}
// Send to Writer thread
auto writerMsg = MsgWriter{
.connId = msg.connId,
.event = std::move(event),
};
tpWriter->dispatch(std::move(writerMsg));
}
}
```
**Validation sequence:**
1. Parse JSON into Event struct
2. Validate event ID matches content hash
3. Verify secp256k1 signature
4. Check duplicate (bloom filter for speed)
5. Forward to Writer thread for storage
### Writer Thread
```cpp
void WriterThread::run() {
// Single thread for all database writes
while (running) {
Message msg;
if (!queue.pop(msg, 100ms)) continue;
// Write to database
bool success = db.insertEvent(msg.event);
// Send OK to client
sendOK(msg.connId, msg.event.id, success,
success ? "" : "error: failed to store");
if (success) {
// Broadcast to subscribers
broadcastEvent(msg.event);
}
}
}
```
**Single-writer pattern:**
- Only one thread writes to database
- Eliminates write conflicts
- Simplified transaction management
### Event Broadcasting
```cpp
void broadcastEvent(const Event &event) {
// Serialize event JSON once
std::string eventJson = serializeEvent(event);
// Iterate all active subscriptions
for (auto &[connId, sub] : activeSubscriptions) {
// Check if filter matches
if (!sub->filter.matches(event)) continue;
// Check if event newer than last sent
if (event.id <= sub->latestEventSent) continue;
// Send to connection
auto msg = MsgWebSocket{
.connId = connId,
.payload = eventJson, // Reuse serialized JSON
};
tpWebSocket->dispatch(std::move(msg));
// Update latest sent
sub->latestEventSent = event.id;
}
}
```
**Critical optimization:** Serialize event JSON once, send to N subscribers
**Performance impact:** For 1000 subscribers, reduces:
- JSON serialization: 1000× → 1×
- Memory allocations: 1000× → 1×
- CPU time: ~100ms → ~1ms
## Subscription Management
### REQ Processing
```cpp
void ReqWorkerThread::run() {
while (running) {
MsgReq msg;
if (!queue.pop(msg, 100ms)) continue;
// Parse REQ message: ["REQ", subId, filter1, filter2, ...]
std::string subId = msg.payload[1];
// Create subscription object
auto sub = std::make_shared<Subscription>();
sub->subId = subId;
// Parse filters
for (size_t i = 2; i < msg.payload.size(); i++) {
Filter filter = parseFilter(msg.payload[i]);
sub->filters.push_back(filter);
}
// Store subscription
activeSubscriptions[msg.connId] = sub;
// Query stored events
std::vector<Event> events = db.queryEvents(sub->filters);
// Send matching events
for (const auto &event : events) {
sendEvent(msg.connId, subId, event);
}
// Send EOSE
sendEOSE(msg.connId, subId);
// Notify ReqMonitor to watch for real-time events
auto monitorMsg = MsgReqMonitor{
.connId = msg.connId,
.subId = subId,
};
tpReqMonitor->dispatchToThread(msg.connId, std::move(monitorMsg));
}
}
```
**Query optimization:**
```cpp
std::vector<Event> Database::queryEvents(const std::vector<Filter> &filters) {
// Combine filters with OR logic
std::string sql = "SELECT * FROM events WHERE ";
for (size_t i = 0; i < filters.size(); i++) {
if (i > 0) sql += " OR ";
sql += buildFilterSQL(filters[i]);
}
sql += " ORDER BY created_at DESC LIMIT 1000";
return executeQuery(sql);
}
```
**Filter SQL generation:**
```cpp
std::string buildFilterSQL(const Filter &filter) {
std::vector<std::string> conditions;
// Event IDs
if (!filter.ids.empty()) {
conditions.push_back("id IN (" + joinQuoted(filter.ids) + ")");
}
// Authors
if (!filter.authors.empty()) {
conditions.push_back("pubkey IN (" + joinQuoted(filter.authors) + ")");
}
// Kinds
if (!filter.kinds.empty()) {
conditions.push_back("kind IN (" + join(filter.kinds) + ")");
}
// Time range
if (filter.since) {
conditions.push_back("created_at >= " + std::to_string(*filter.since));
}
if (filter.until) {
conditions.push_back("created_at <= " + std::to_string(*filter.until));
}
// Tags (requires JOIN with tags table)
if (!filter.tags.empty()) {
for (const auto &[tagName, tagValues] : filter.tags) {
conditions.push_back(
"EXISTS (SELECT 1 FROM tags WHERE tags.event_id = events.id "
"AND tags.name = '" + tagName + "' "
"AND tags.value IN (" + joinQuoted(tagValues) + "))"
);
}
}
return "(" + join(conditions, " AND ") + ")";
}
```
### ReqMonitor for Real-Time Events
```cpp
void ReqMonitorThread::run() {
// Subscribe to event broadcast channel
auto eventSubscription = subscribeToEvents();
while (running) {
Event event;
if (!eventSubscription.receive(event, 100ms)) continue;
// Check all subscriptions assigned to this thread
for (auto &[connId, sub] : mySubscriptions) {
// Only process subscriptions for this thread
if (connId % numThreads != threadId) continue;
// Check if filter matches
bool matches = false;
for (const auto &filter : sub->filters) {
if (filter.matches(event)) {
matches = true;
break;
}
}
if (matches) {
sendEvent(connId, sub->subId, event);
}
}
}
}
```
**Pattern:** Monitor thread watches event stream, sends to matching subscriptions
### CLOSE Handling
```cpp
void handleCloseMessage(auto *ws, nlohmann::json &&json) {
auto *state = ws->getUserData();
// Parse CLOSE message: ["CLOSE", subId]
std::string subId = json[1];
// Remove subscription
activeSubscriptions.erase(state->connId);
LI << "Subscription closed: connId=" << state->connId
<< " subId=" << subId;
}
```
## Performance Optimizations
### 1. Event Batching
**Problem:** Serializing same event 1000× for 1000 subscribers is wasteful
**Solution:** Serialize once, send to all
```cpp
// BAD: Serialize for each subscriber
for (auto &sub : subscriptions) {
std::string json = serializeEvent(event); // Repeated!
send(sub.connId, json);
}
// GOOD: Serialize once
std::string json = serializeEvent(event);
for (auto &sub : subscriptions) {
send(sub.connId, json); // Reuse!
}
```
**Measurement:** For 1000 subscribers, reduces broadcast time from 100ms to 1ms
### 2. Move Semantics
**Problem:** Copying large JSON objects is expensive
**Solution:** Transfer ownership with `std::move`
```cpp
// BAD: Copies JSON object
void dispatch(Message msg) {
queue.push(msg); // Copy
}
// GOOD: Moves JSON object
void dispatch(Message &&msg) {
queue.push(std::move(msg)); // Move
}
```
**Benefit:** Zero-copy message passing between threads
### 3. Pre-allocated Buffers
**Problem:** Allocating buffer for each message
**Solution:** Reuse buffer per connection
```cpp
struct ConnectionState {
std::string parseBuffer; // Reused for all messages
};
void handleMessage(std::string_view msg) {
state->parseBuffer.assign(msg.data(), msg.size());
auto json = nlohmann::json::parse(state->parseBuffer);
// ...
}
```
**Benefit:** Eliminates 10,000+ allocations/second per connection
### 4. std::variant for Message Types
**Problem:** Virtual function calls for polymorphic messages
**Solution:** `std::variant` with `std::visit`
```cpp
// BAD: Virtual function (pointer indirection, vtable lookup)
struct Message {
virtual void handle() = 0;
};
// GOOD: std::variant (no indirection, inlined)
using Message = std::variant<
MsgIngester,
MsgReq,
MsgWriter,
MsgWebSocket
>;
void handle(Message &&msg) {
std::visit([](auto &&m) { m.handle(); }, msg);
}
```
**Benefit:** Compiler inlines visit, eliminates virtual call overhead
### 5. Bloom Filter for Duplicate Detection
**Problem:** Database query for every event to check duplicate
**Solution:** In-memory bloom filter for fast negative
```cpp
class DuplicateDetector {
BloomFilter bloom; // Fast probabilistic check
bool isDuplicate(const std::string &eventId) {
// Fast negative (definitely not seen)
if (!bloom.contains(eventId)) {
bloom.insert(eventId);
return false;
}
// Possible positive (maybe seen, check database)
if (db.eventExists(eventId)) {
return true;
}
// False positive
bloom.insert(eventId);
return false;
}
};
```
**Benefit:** 99% of duplicate checks avoid database query
### 6. Batch Queue Operations
**Problem:** Lock contention on message queue
**Solution:** Batch multiple pushes with single lock
```cpp
class MessageQueue {
std::mutex mutex;
std::deque<Message> queue;
void pushBatch(std::vector<Message> &messages) {
std::lock_guard lock(mutex);
for (auto &msg : messages) {
queue.push_back(std::move(msg));
}
}
};
```
**Benefit:** Reduces lock acquisitions by 10-100×
### 7. ZSTD Dictionary Compression
**Problem:** WebSocket compression slower than desired
**Solution:** Train ZSTD dictionary on typical Nostr messages
```cpp
// Train dictionary on corpus of Nostr events
std::string corpus = collectTypicalEvents();
ZSTD_CDict *dict = ZSTD_createCDict(
corpus.data(), corpus.size(),
compressionLevel
);
// Use dictionary for compression
size_t compressedSize = ZSTD_compress_usingCDict(
cctx, dst, dstSize,
src, srcSize, dict
);
```
**Benefit:** 10-20% better compression ratio, 2× faster decompression
### 8. String Views
**Problem:** Unnecessary string copies when parsing
**Solution:** Use `std::string_view` for zero-copy
```cpp
// BAD: Copies substring
std::string extractCommand(const std::string &msg) {
return msg.substr(0, 5); // Copy
}
// GOOD: View into original string
std::string_view extractCommand(std::string_view msg) {
return msg.substr(0, 5); // No copy
}
```
**Benefit:** Eliminates allocations during parsing
## Compression (permessage-deflate)
### WebSocket Compression Configuration
```cpp
struct PerMessageDeflate {
z_stream deflate_stream;
z_stream inflate_stream;
// Sliding window for compression history
static constexpr int WINDOW_BITS = 15;
static constexpr int MEM_LEVEL = 8;
void init() {
// Initialize deflate (compression)
deflate_stream.zalloc = Z_NULL;
deflate_stream.zfree = Z_NULL;
deflate_stream.opaque = Z_NULL;
deflateInit2(&deflate_stream,
Z_DEFAULT_COMPRESSION,
Z_DEFLATED,
-WINDOW_BITS, // Negative = no zlib header
MEM_LEVEL,
Z_DEFAULT_STRATEGY);
// Initialize inflate (decompression)
inflate_stream.zalloc = Z_NULL;
inflate_stream.zfree = Z_NULL;
inflate_stream.opaque = Z_NULL;
inflateInit2(&inflate_stream, -WINDOW_BITS);
}
std::string compress(std::string_view data) {
// Compress with sliding window
deflate_stream.next_in = (Bytef*)data.data();
deflate_stream.avail_in = data.size();
std::string compressed;
compressed.resize(deflateBound(&deflate_stream, data.size()));
deflate_stream.next_out = (Bytef*)compressed.data();
deflate_stream.avail_out = compressed.size();
deflate(&deflate_stream, Z_SYNC_FLUSH);
compressed.resize(compressed.size() - deflate_stream.avail_out);
return compressed;
}
};
```
**Typical compression ratios:**
- JSON events: 60-80% reduction
- Subscription filters: 40-60% reduction
- Binary events: 10-30% reduction
## Database Schema (LMDB)
strfry uses LMDB (Lightning Memory-Mapped Database) for event storage:
```cpp
// Key-value stores
struct EventDB {
// Primary event storage (key: event ID, value: event data)
lmdb::dbi eventsDB;
// Index by pubkey (key: pubkey + created_at, value: event ID)
lmdb::dbi pubkeyDB;
// Index by kind (key: kind + created_at, value: event ID)
lmdb::dbi kindDB;
// Index by tags (key: tag_name + tag_value + created_at, value: event ID)
lmdb::dbi tagsDB;
// Deletion index (key: event ID, value: deletion event ID)
lmdb::dbi deletionsDB;
};
```
**Why LMDB?**
- Memory-mapped I/O (kernel manages caching)
- Copy-on-write (MVCC without locks)
- Ordered keys (enables range queries)
- Crash-proof (no corruption on power loss)
## Monitoring and Metrics
### Connection Statistics
```cpp
struct RelayStats {
std::atomic<uint64_t> totalConnections{0};
std::atomic<uint64_t> activeConnections{0};
std::atomic<uint64_t> eventsReceived{0};
std::atomic<uint64_t> eventsSent{0};
std::atomic<uint64_t> bytesReceived{0};
std::atomic<uint64_t> bytesSent{0};
void recordConnection() {
totalConnections.fetch_add(1, std::memory_order_relaxed);
activeConnections.fetch_add(1, std::memory_order_relaxed);
}
void recordDisconnection() {
activeConnections.fetch_sub(1, std::memory_order_relaxed);
}
void recordEventReceived(size_t bytes) {
eventsReceived.fetch_add(1, std::memory_order_relaxed);
bytesReceived.fetch_add(bytes, std::memory_order_relaxed);
}
};
```
**Atomic operations:** Lock-free updates from multiple threads
### Performance Metrics
```cpp
struct PerformanceMetrics {
// Latency histograms
Histogram eventIngestionLatency;
Histogram subscriptionQueryLatency;
Histogram eventBroadcastLatency;
// Thread pool queue depths
std::atomic<size_t> ingesterQueueDepth{0};
std::atomic<size_t> writerQueueDepth{0};
std::atomic<size_t> reqWorkerQueueDepth{0};
void recordIngestion(std::chrono::microseconds duration) {
eventIngestionLatency.record(duration.count());
}
};
```
## Configuration
### relay.conf Example
```ini
[relay]
bind = 0.0.0.0
port = 8080
maxConnections = 10000
maxMessageSize = 16777216 # 16 MB
[ingester]
threads = 3
queueSize = 10000
[writer]
threads = 1
queueSize = 1000
batchSize = 100
[reqWorker]
threads = 3
queueSize = 10000
[db]
path = /var/lib/strfry/events.lmdb
maxSizeGB = 100
```
## Deployment Considerations
### System Limits
```bash
# Increase file descriptor limit
ulimit -n 65536
# Increase maximum socket connections
sysctl -w net.core.somaxconn=4096
# TCP tuning
sysctl -w net.ipv4.tcp_fin_timeout=15
sysctl -w net.ipv4.tcp_tw_reuse=1
```
### Memory Requirements
**Per connection:**
- ConnectionState: ~1 KB
- WebSocket buffers: ~32 KB (16 KB send + 16 KB receive)
- Compression state: ~400 KB (200 KB deflate + 200 KB inflate)
**Total:** ~433 KB per connection
**For 10,000 connections:** ~4.3 GB
### CPU Requirements
**Single-core can handle:**
- 1000 concurrent connections
- 10,000 events/sec ingestion
- 100,000 events/sec broadcast (cached)
**Recommended:**
- 8+ cores for 10,000 connections
- 16+ cores for 50,000 connections
## Summary
**Key architectural patterns:**
1. **Single-threaded I/O:** epoll handles all connections in one thread
2. **Specialized thread pools:** Different operations use dedicated threads
3. **Deterministic assignment:** Connection ID determines thread assignment
4. **Move semantics:** Zero-copy message passing
5. **Event batching:** Serialize once, send to many
6. **Pre-allocated buffers:** Reuse memory per connection
7. **Bloom filters:** Fast duplicate detection
8. **LMDB:** Memory-mapped database for zero-copy reads
**Performance characteristics:**
- **50,000+ concurrent connections** per server
- **100,000+ events/sec** throughput
- **Sub-millisecond** latency for broadcasts
- **10 GB+ event database** with fast queries
**When to use strfry patterns:**
- Need maximum performance (trading complexity)
- Have C++ expertise on team
- Running large public relay (thousands of users)
- Want minimal memory footprint
- Need to scale to 50K+ connections
**Trade-offs:**
- **Complexity:** More complex than Go/Rust implementations
- **Portability:** Linux-specific (epoll, LMDB)
- **Development speed:** Slower iteration than higher-level languages
**Further reading:**
- strfry repository: https://github.com/hoytech/strfry
- uWebSockets: https://github.com/uNetworking/uWebSockets
- LMDB: http://www.lmdb.tech/doc/
- epoll: https://man7.org/linux/man-pages/man7/epoll.7.html

View File

@@ -0,0 +1,881 @@
# WebSocket Protocol (RFC 6455) - Complete Reference
## Connection Establishment
### HTTP Upgrade Handshake
The WebSocket protocol begins as an HTTP request that upgrades to WebSocket:
**Client Request:**
```http
GET /chat HTTP/1.1
Host: server.example.com
Upgrade: websocket
Connection: Upgrade
Sec-WebSocket-Key: dGhlIHNhbXBsZSBub25jZQ==
Origin: http://example.com
Sec-WebSocket-Protocol: chat, superchat
Sec-WebSocket-Version: 13
```
**Server Response:**
```http
HTTP/1.1 101 Switching Protocols
Upgrade: websocket
Connection: Upgrade
Sec-WebSocket-Accept: s3pPLMBiTxaQ9kYGzzhZRbK+xOo=
Sec-WebSocket-Protocol: chat
```
### Handshake Details
**Sec-WebSocket-Key Generation (Client):**
1. Generate 16 random bytes
2. Base64-encode the result
3. Send in `Sec-WebSocket-Key` header
**Sec-WebSocket-Accept Computation (Server):**
1. Concatenate client key with GUID: `258EAFA5-E914-47DA-95CA-C5AB0DC85B11`
2. Compute SHA-1 hash of concatenated string
3. Base64-encode the hash
4. Send in `Sec-WebSocket-Accept` header
**Example computation:**
```
Client Key: dGhlIHNhbXBsZSBub25jZQ==
Concatenated: dGhlIHNhbXBsZSBub25jZQ==258EAFA5-E914-47DA-95CA-C5AB0DC85B11
SHA-1 Hash: b37a4f2cc0cb4e7e8cf769a5f3f8f2e8e4c9f7a3
Base64: s3pPLMBiTxaQ9kYGzzhZRbK+xOo=
```
**Validation (Client):**
- Verify HTTP status is 101
- Verify `Sec-WebSocket-Accept` matches expected value
- If validation fails, do not establish connection
### Origin Header
The `Origin` header provides protection against cross-site WebSocket hijacking:
**Server-side validation:**
```go
func checkOrigin(r *http.Request) bool {
origin := r.Header.Get("Origin")
allowedOrigins := []string{
"https://example.com",
"https://app.example.com",
}
for _, allowed := range allowedOrigins {
if origin == allowed {
return true
}
}
return false
}
```
**Security consideration:** Browser-based clients MUST send Origin header. Non-browser clients MAY omit it. Servers SHOULD validate Origin for browser clients to prevent CSRF attacks.
## Frame Format
### Base Framing Protocol
WebSocket frames use a binary format with variable-length fields:
```
0 1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-------+-+-------------+-------------------------------+
|F|R|R|R| opcode|M| Payload len | Extended payload length |
|I|S|S|S| (4) |A| (7) | (16/64) |
|N|V|V|V| |S| | (if payload len==126/127) |
| |1|2|3| |K| | |
+-+-+-+-+-------+-+-------------+ - - - - - - - - - - - - - - - +
| Extended payload length continued, if payload len == 127 |
+ - - - - - - - - - - - - - - - +-------------------------------+
| |Masking-key, if MASK set to 1 |
+-------------------------------+-------------------------------+
| Masking-key (continued) | Payload Data |
+-------------------------------- - - - - - - - - - - - - - - - +
: Payload Data continued ... :
+ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - +
| Payload Data continued ... |
+---------------------------------------------------------------+
```
### Frame Header Fields
**FIN (1 bit):**
- `1` = Final fragment in message
- `0` = More fragments follow
- Used for message fragmentation
**RSV1, RSV2, RSV3 (1 bit each):**
- Reserved for extensions
- MUST be 0 unless extension negotiated
- Server MUST fail connection if non-zero with no extension
**Opcode (4 bits):**
- Defines interpretation of payload data
- See "Frame Opcodes" section below
**MASK (1 bit):**
- `1` = Payload is masked (required for client-to-server)
- `0` = Payload is not masked (required for server-to-client)
- Client MUST mask all frames sent to server
- Server MUST NOT mask frames sent to client
**Payload Length (7 bits, 7+16 bits, or 7+64 bits):**
- If 0-125: Actual payload length
- If 126: Next 2 bytes are 16-bit unsigned payload length
- If 127: Next 8 bytes are 64-bit unsigned payload length
**Masking-key (0 or 4 bytes):**
- Present if MASK bit is set
- 32-bit value used to mask payload
- MUST be unpredictable (strong entropy source)
### Frame Opcodes
**Data Frame Opcodes:**
- `0x0` - Continuation Frame
- Used for fragmented messages
- Must follow initial data frame (text/binary)
- Carries same data type as initial frame
- `0x1` - Text Frame
- Payload is UTF-8 encoded text
- MUST be valid UTF-8
- Endpoint MUST fail connection if invalid UTF-8
- `0x2` - Binary Frame
- Payload is arbitrary binary data
- Application interprets data
- `0x3-0x7` - Reserved for future non-control frames
**Control Frame Opcodes:**
- `0x8` - Connection Close
- Initiates or acknowledges connection closure
- MAY contain status code and reason
- See "Close Handshake" section
- `0x9` - Ping
- Heartbeat mechanism
- MAY contain application data
- Recipient MUST respond with Pong
- `0xA` - Pong
- Response to Ping
- MUST contain identical payload as Ping
- MAY be sent unsolicited (unidirectional heartbeat)
- `0xB-0xF` - Reserved for future control frames
### Control Frame Constraints
**Control frames are subject to strict rules:**
1. **Maximum payload:** 125 bytes
- Allows control frames to fit in single IP packet
- Reduces fragmentation
2. **No fragmentation:** Control frames MUST NOT be fragmented
- FIN bit MUST be 1
- Ensures immediate processing
3. **Interleaving:** Control frames MAY be injected in middle of fragmented message
- Enables ping/pong during long transfers
- Close frames can interrupt any operation
4. **All control frames MUST be handled immediately**
### Masking
**Purpose of masking:**
- Prevents cache poisoning attacks
- Protects against misinterpretation by intermediaries
- Makes WebSocket traffic unpredictable to proxies
**Masking algorithm:**
```
j = i MOD 4
transformed-octet-i = original-octet-i XOR masking-key-octet-j
```
**Implementation:**
```go
func maskBytes(data []byte, mask [4]byte) {
for i := range data {
data[i] ^= mask[i%4]
}
}
```
**Example:**
```
Original: [0x48, 0x65, 0x6C, 0x6C, 0x6F] // "Hello"
Masking Key: [0x37, 0xFA, 0x21, 0x3D]
Masked: [0x7F, 0x9F, 0x4D, 0x51, 0x58]
Calculation:
0x48 XOR 0x37 = 0x7F
0x65 XOR 0xFA = 0x9F
0x6C XOR 0x21 = 0x4D
0x6C XOR 0x3D = 0x51
0x6F XOR 0x37 = 0x58 (wraps around to mask[0])
```
**Security requirement:** Masking key MUST be derived from strong source of entropy. Predictable masking keys defeat the security purpose.
## Message Fragmentation
### Why Fragment?
- Send message without knowing total size upfront
- Multiplex logical channels (interleave messages)
- Keep control frames responsive during large transfers
### Fragmentation Rules
**Sender rules:**
1. First fragment has opcode (text/binary)
2. Subsequent fragments have opcode 0x0 (continuation)
3. Last fragment has FIN bit set to 1
4. Control frames MAY be interleaved
**Receiver rules:**
1. Reassemble fragments in order
2. Final message type determined by first fragment opcode
3. Validate UTF-8 across all text fragments
4. Process control frames immediately (don't wait for FIN)
### Fragmentation Example
**Sending "Hello World" in 3 fragments:**
```
Frame 1 (Text, More Fragments):
FIN=0, Opcode=0x1, Payload="Hello"
Frame 2 (Continuation, More Fragments):
FIN=0, Opcode=0x0, Payload=" Wor"
Frame 3 (Continuation, Final):
FIN=1, Opcode=0x0, Payload="ld"
```
**With interleaved Ping:**
```
Frame 1: FIN=0, Opcode=0x1, Payload="Hello"
Frame 2: FIN=1, Opcode=0x9, Payload="" <- Ping (complete)
Frame 3: FIN=0, Opcode=0x0, Payload=" Wor"
Frame 4: FIN=1, Opcode=0x0, Payload="ld"
```
### Implementation Pattern
```go
type fragmentState struct {
messageType int
fragments [][]byte
}
func (ws *WebSocket) handleFrame(fin bool, opcode int, payload []byte) {
switch opcode {
case 0x1, 0x2: // Text or Binary (first fragment)
if fin {
ws.handleCompleteMessage(opcode, payload)
} else {
ws.fragmentState = &fragmentState{
messageType: opcode,
fragments: [][]byte{payload},
}
}
case 0x0: // Continuation
if ws.fragmentState == nil {
ws.fail("Unexpected continuation frame")
return
}
ws.fragmentState.fragments = append(ws.fragmentState.fragments, payload)
if fin {
complete := bytes.Join(ws.fragmentState.fragments, nil)
ws.handleCompleteMessage(ws.fragmentState.messageType, complete)
ws.fragmentState = nil
}
case 0x8, 0x9, 0xA: // Control frames
ws.handleControlFrame(opcode, payload)
}
}
```
## Ping and Pong Frames
### Purpose
1. **Keep-alive:** Detect broken connections
2. **Latency measurement:** Time round-trip
3. **NAT traversal:** Maintain mapping in stateful firewalls
### Protocol Rules
**Ping (0x9):**
- MAY be sent by either endpoint at any time
- MAY contain application data (≤125 bytes)
- Application data arbitrary (often empty or timestamp)
**Pong (0xA):**
- MUST be sent in response to Ping
- MUST contain identical payload as Ping
- MUST be sent "as soon as practical"
- MAY be sent unsolicited (one-way heartbeat)
**No Response:**
- If Pong not received within timeout, connection assumed dead
- Application should close connection
### Implementation Patterns
**Pattern 1: Automatic Pong (most WebSocket libraries)**
```go
// Library handles pong automatically
ws.SetPingHandler(func(appData string) error {
// Custom handler if needed
return nil // Library sends pong automatically
})
```
**Pattern 2: Manual Pong**
```go
func (ws *WebSocket) handlePing(payload []byte) {
pongFrame := Frame{
FIN: true,
Opcode: 0xA,
Payload: payload, // Echo same payload
}
ws.writeFrame(pongFrame)
}
```
**Pattern 3: Periodic Client Ping**
```go
func (ws *WebSocket) pingLoop() {
ticker := time.NewTicker(30 * time.Second)
defer ticker.Stop()
for {
select {
case <-ticker.C:
if err := ws.writePing([]byte{}); err != nil {
return // Connection dead
}
case <-ws.done:
return
}
}
}
```
**Pattern 4: Timeout Detection**
```go
const pongWait = 60 * time.Second
ws.SetReadDeadline(time.Now().Add(pongWait))
ws.SetPongHandler(func(string) error {
ws.SetReadDeadline(time.Now().Add(pongWait))
return nil
})
// If no frame received in pongWait, ReadMessage returns timeout error
```
### Nostr Relay Recommendations
**Server-side:**
- Send ping every 30-60 seconds
- Close connection if no pong within 60-120 seconds
- Log timeout closures for monitoring
**Client-side:**
- Respond to pings automatically (use library handler)
- Consider sending unsolicited pongs every 30 seconds (some proxies)
- Reconnect if no frames received for 120 seconds
## Close Handshake
### Close Frame Structure
**Close frame (Opcode 0x8) payload:**
```
0 1 2 3
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
| Status Code (16) | Reason (variable length)... |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
```
**Status Code (2 bytes, optional):**
- 16-bit unsigned integer
- Network byte order (big-endian)
- See "Status Codes" section below
**Reason (variable length, optional):**
- UTF-8 encoded text
- MUST be valid UTF-8
- Typically human-readable explanation
### Close Handshake Sequence
**Initiator (either endpoint):**
1. Send Close frame with optional status/reason
2. Stop sending data frames
3. Continue processing received frames until Close frame received
4. Close underlying TCP connection
**Recipient:**
1. Receive Close frame
2. Send Close frame in response (if not already sent)
3. Close underlying TCP connection
### Status Codes
**Normal Closure Codes:**
- `1000` - Normal Closure
- Successful operation complete
- Default if no code specified
- `1001` - Going Away
- Endpoint going away (server shutdown, browser navigation)
- Client navigating to new page
**Error Closure Codes:**
- `1002` - Protocol Error
- Endpoint terminating due to protocol error
- Invalid frame format, unexpected opcode, etc.
- `1003` - Unsupported Data
- Endpoint cannot accept data type
- Server received binary when expecting text
- `1007` - Invalid Frame Payload Data
- Inconsistent data (e.g., non-UTF-8 in text frame)
- `1008` - Policy Violation
- Message violates endpoint policy
- Generic code when specific code doesn't fit
- `1009` - Message Too Big
- Message too large to process
- `1010` - Mandatory Extension
- Client expected server to negotiate extension
- Server didn't respond with extension
- `1011` - Internal Server Error
- Server encountered unexpected condition
- Prevents fulfilling request
**Reserved Codes:**
- `1004` - Reserved
- `1005` - No Status Rcvd (internal use only, never sent)
- `1006` - Abnormal Closure (internal use only, never sent)
- `1015` - TLS Handshake (internal use only, never sent)
**Custom Application Codes:**
- `3000-3999` - Library/framework use
- `4000-4999` - Application use (e.g., Nostr-specific)
### Implementation Patterns
**Graceful close (initiator):**
```go
func (ws *WebSocket) Close() error {
// Send close frame
closeFrame := Frame{
FIN: true,
Opcode: 0x8,
Payload: encodeCloseStatus(1000, "goodbye"),
}
ws.writeFrame(closeFrame)
// Wait for close frame response (with timeout)
ws.SetReadDeadline(time.Now().Add(5 * time.Second))
for {
frame, err := ws.readFrame()
if err != nil || frame.Opcode == 0x8 {
break
}
// Process other frames
}
// Close TCP connection
return ws.conn.Close()
}
```
**Handling received close:**
```go
func (ws *WebSocket) handleCloseFrame(payload []byte) {
status, reason := decodeClosePayload(payload)
log.Printf("Close received: %d %s", status, reason)
// Send close response
closeFrame := Frame{
FIN: true,
Opcode: 0x8,
Payload: payload, // Echo same status/reason
}
ws.writeFrame(closeFrame)
// Close connection
ws.conn.Close()
}
```
**Nostr relay close examples:**
```go
// Client subscription limit exceeded
ws.SendClose(4000, "subscription limit exceeded")
// Invalid message format
ws.SendClose(1002, "protocol error: invalid JSON")
// Relay shutting down
ws.SendClose(1001, "relay shutting down")
// Client rate limit exceeded
ws.SendClose(4001, "rate limit exceeded")
```
## Security Considerations
### Origin-Based Security Model
**Threat:** Malicious web page opens WebSocket to victim server using user's credentials
**Mitigation:**
1. Server checks `Origin` header
2. Reject connections from untrusted origins
3. Implement same-origin or allowlist policy
**Example:**
```go
func validateOrigin(r *http.Request) bool {
origin := r.Header.Get("Origin")
// Allow same-origin
if origin == "https://"+r.Host {
return true
}
// Allowlist trusted origins
trusted := []string{
"https://app.example.com",
"https://mobile.example.com",
}
for _, t := range trusted {
if origin == t {
return true
}
}
return false
}
```
### Masking Attacks
**Why masking is required:**
- Without masking, attacker can craft WebSocket frames that look like HTTP requests
- Proxies might misinterpret frame data as HTTP
- Could lead to cache poisoning or request smuggling
**Example attack (without masking):**
```
WebSocket payload: "GET /admin HTTP/1.1\r\nHost: victim.com\r\n\r\n"
Proxy might interpret as separate HTTP request
```
**Defense:** Client MUST mask all frames. Server MUST reject unmasked frames from client.
### Connection Limits
**Prevent resource exhaustion:**
```go
type ConnectionLimiter struct {
connections map[string]int
maxPerIP int
mu sync.Mutex
}
func (cl *ConnectionLimiter) Allow(ip string) bool {
cl.mu.Lock()
defer cl.mu.Unlock()
if cl.connections[ip] >= cl.maxPerIP {
return false
}
cl.connections[ip]++
return true
}
func (cl *ConnectionLimiter) Release(ip string) {
cl.mu.Lock()
defer cl.mu.Unlock()
cl.connections[ip]--
}
```
### TLS (WSS)
**Use WSS (WebSocket Secure) for:**
- Authentication credentials
- Private user data
- Financial transactions
- Any sensitive information
**WSS connection flow:**
1. Establish TLS connection
2. Perform TLS handshake
3. Verify server certificate
4. Perform WebSocket handshake over TLS
**URL schemes:**
- `ws://` - Unencrypted WebSocket (default port 80)
- `wss://` - Encrypted WebSocket over TLS (default port 443)
### Message Size Limits
**Prevent memory exhaustion:**
```go
const maxMessageSize = 512 * 1024 // 512 KB
ws.SetReadLimit(maxMessageSize)
// Or during frame reading:
if payloadLength > maxMessageSize {
ws.SendClose(1009, "message too large")
ws.Close()
}
```
### Rate Limiting
**Prevent abuse:**
```go
type RateLimiter struct {
limiter *rate.Limiter
}
func (rl *RateLimiter) Allow() bool {
return rl.limiter.Allow()
}
// Per-connection limiter
limiter := rate.NewLimiter(10, 20) // 10 msgs/sec, burst 20
if !limiter.Allow() {
ws.SendClose(4001, "rate limit exceeded")
}
```
## Error Handling
### Connection Errors
**Types of errors:**
1. **Network errors:** TCP connection failure, timeout
2. **Protocol errors:** Invalid frame format, wrong opcode
3. **Application errors:** Invalid message content
**Handling strategy:**
```go
for {
frame, err := ws.ReadFrame()
if err != nil {
// Check error type
if netErr, ok := err.(net.Error); ok && netErr.Timeout() {
// Timeout - connection likely dead
log.Println("Connection timeout")
ws.Close()
return
}
if err == io.EOF || err == io.ErrUnexpectedEOF {
// Connection closed
log.Println("Connection closed")
return
}
if protocolErr, ok := err.(*ProtocolError); ok {
// Protocol violation
log.Printf("Protocol error: %v", protocolErr)
ws.SendClose(1002, protocolErr.Error())
ws.Close()
return
}
// Unknown error
log.Printf("Unknown error: %v", err)
ws.Close()
return
}
// Process frame
}
```
### UTF-8 Validation
**Text frames MUST contain valid UTF-8:**
```go
func validateUTF8(data []byte) bool {
return utf8.Valid(data)
}
func handleTextFrame(payload []byte) error {
if !validateUTF8(payload) {
return fmt.Errorf("invalid UTF-8 in text frame")
}
// Process valid text
return nil
}
```
**For fragmented messages:** Validate UTF-8 across all fragments when reassembled.
## Implementation Checklist
### Client Implementation
- [ ] Generate random Sec-WebSocket-Key
- [ ] Compute and validate Sec-WebSocket-Accept
- [ ] MUST mask all frames sent to server
- [ ] Handle unmasked frames from server
- [ ] Respond to Ping with Pong
- [ ] Implement close handshake (both initiating and responding)
- [ ] Validate UTF-8 in text frames
- [ ] Handle fragmented messages
- [ ] Set reasonable timeouts
- [ ] Implement reconnection logic
### Server Implementation
- [ ] Validate Sec-WebSocket-Key format
- [ ] Compute correct Sec-WebSocket-Accept
- [ ] Validate Origin header
- [ ] MUST NOT mask frames sent to client
- [ ] Reject masked frames from server (protocol error)
- [ ] Respond to Ping with Pong
- [ ] Implement close handshake (both initiating and responding)
- [ ] Validate UTF-8 in text frames
- [ ] Handle fragmented messages
- [ ] Implement connection limits (per IP, total)
- [ ] Implement message size limits
- [ ] Implement rate limiting
- [ ] Log connection statistics
- [ ] Graceful shutdown (close all connections)
### Both Client and Server
- [ ] Handle concurrent read/write safely
- [ ] Process control frames immediately (even during fragmentation)
- [ ] Implement proper timeout mechanisms
- [ ] Log errors with appropriate detail
- [ ] Handle unexpected close gracefully
- [ ] Validate frame structure
- [ ] Check RSV bits (must be 0 unless extension)
- [ ] Support standard close status codes
- [ ] Implement proper error handling for all operations
## Common Implementation Mistakes
### 1. Concurrent Writes
**Mistake:** Writing to WebSocket from multiple goroutines without synchronization
**Fix:** Use mutex or single-writer goroutine
```go
type WebSocket struct {
conn *websocket.Conn
mutex sync.Mutex
}
func (ws *WebSocket) WriteMessage(data []byte) error {
ws.mutex.Lock()
defer ws.mutex.Unlock()
return ws.conn.WriteMessage(websocket.TextMessage, data)
}
```
### 2. Not Handling Pong
**Mistake:** Sending Ping but not updating read deadline on Pong
**Fix:**
```go
ws.SetPongHandler(func(string) error {
ws.SetReadDeadline(time.Now().Add(pongWait))
return nil
})
```
### 3. Forgetting Close Handshake
**Mistake:** Just calling `conn.Close()` without sending Close frame
**Fix:** Send Close frame first, wait for response, then close TCP
### 4. Not Validating UTF-8
**Mistake:** Accepting any bytes in text frames
**Fix:** Validate UTF-8 and fail connection on invalid text
### 5. No Message Size Limit
**Mistake:** Allowing unlimited message sizes
**Fix:** Set `SetReadLimit()` to reasonable value (e.g., 512 KB)
### 6. Blocking on Write
**Mistake:** Blocking indefinitely on slow clients
**Fix:** Set write deadline before each write
```go
ws.SetWriteDeadline(time.Now().Add(10 * time.Second))
```
### 7. Memory Leaks
**Mistake:** Not cleaning up resources on disconnect
**Fix:** Use defer for cleanup, ensure all goroutines terminate
### 8. Race Conditions in Close
**Mistake:** Multiple goroutines trying to close connection
**Fix:** Use `sync.Once` for close operation
```go
type WebSocket struct {
conn *websocket.Conn
closeOnce sync.Once
}
func (ws *WebSocket) Close() error {
var err error
ws.closeOnce.Do(func() {
err = ws.conn.Close()
})
return err
}
```