|
| 1 | +use std::collections::HashMap; |
| 2 | +use std::time::{Duration, Instant}; |
| 3 | + |
| 4 | +use anyhow::*; |
| 5 | +use rivet_ups_protocol::versioned::UpsMessage; |
| 6 | +use rivet_ups_protocol::{MessageBody, MessageChunk, MessageStart, PROTOCOL_VERSION}; |
| 7 | +use versioned_data_util::OwnedVersionedData; |
| 8 | + |
| 9 | +pub const CHUNK_BUFFER_GC_INTERVAL: Duration = Duration::from_secs(60); |
| 10 | +pub const CHUNK_BUFFER_MAX_AGE: Duration = Duration::from_secs(300); |
| 11 | + |
| 12 | +#[derive(Debug)] |
| 13 | +pub struct ChunkBuffer { |
| 14 | + pub message_id: [u8; 16], |
| 15 | + pub received_chunks: u32, |
| 16 | + pub last_chunk_ts: Instant, |
| 17 | + pub buffer: Vec<u8>, |
| 18 | + pub chunk_count: u32, |
| 19 | + pub reply_subject: Option<String>, |
| 20 | +} |
| 21 | + |
| 22 | +pub struct ChunkTracker { |
| 23 | + chunks_in_process: HashMap<[u8; 16], ChunkBuffer>, |
| 24 | +} |
| 25 | + |
| 26 | +impl ChunkTracker { |
| 27 | + pub fn new() -> Self { |
| 28 | + Self { |
| 29 | + chunks_in_process: HashMap::new(), |
| 30 | + } |
| 31 | + } |
| 32 | + |
| 33 | + pub fn process_chunk( |
| 34 | + &mut self, |
| 35 | + raw_message: &[u8], |
| 36 | + ) -> Result<Option<(Vec<u8>, Option<String>)>> { |
| 37 | + let message = UpsMessage::deserialize_with_embedded_version(raw_message)?; |
| 38 | + |
| 39 | + match message.body { |
| 40 | + MessageBody::MessageStart(msg) => { |
| 41 | + // If only one chunk, return immediately |
| 42 | + if msg.chunk_count == 1 { |
| 43 | + return Ok(Some((msg.payload, msg.reply_subject))); |
| 44 | + } |
| 45 | + |
| 46 | + // Start of a multi-chunk message |
| 47 | + let buffer = ChunkBuffer { |
| 48 | + message_id: msg.message_id, |
| 49 | + received_chunks: 1, |
| 50 | + last_chunk_ts: Instant::now(), |
| 51 | + buffer: msg.payload, |
| 52 | + chunk_count: msg.chunk_count, |
| 53 | + reply_subject: msg.reply_subject, |
| 54 | + }; |
| 55 | + self.chunks_in_process.insert(msg.message_id, buffer); |
| 56 | + Ok(None) |
| 57 | + } |
| 58 | + MessageBody::MessageChunk(msg) => { |
| 59 | + // Find the matching buffer using message_id |
| 60 | + let buffer = self.chunks_in_process.get_mut(&msg.message_id); |
| 61 | + |
| 62 | + let Some(buffer) = buffer else { |
| 63 | + bail!( |
| 64 | + "received chunk {} for message {:?} but no matching buffer found", |
| 65 | + msg.chunk_index, |
| 66 | + msg.message_id |
| 67 | + ); |
| 68 | + }; |
| 69 | + |
| 70 | + // Validate chunk order |
| 71 | + if buffer.received_chunks != msg.chunk_index { |
| 72 | + bail!( |
| 73 | + "received chunk {} but expected chunk {} for message {:?}", |
| 74 | + msg.chunk_index, |
| 75 | + buffer.received_chunks, |
| 76 | + msg.message_id |
| 77 | + ); |
| 78 | + } |
| 79 | + |
| 80 | + // Update buffer |
| 81 | + buffer.buffer.extend_from_slice(&msg.payload); |
| 82 | + buffer.received_chunks += 1; |
| 83 | + buffer.last_chunk_ts = Instant::now(); |
| 84 | + let is_complete = buffer.received_chunks == buffer.chunk_count; |
| 85 | + |
| 86 | + if is_complete { |
| 87 | + let completed_buffer = self.chunks_in_process.remove(&msg.message_id).unwrap(); |
| 88 | + Ok(Some(( |
| 89 | + completed_buffer.buffer, |
| 90 | + completed_buffer.reply_subject, |
| 91 | + ))) |
| 92 | + } else { |
| 93 | + Ok(None) |
| 94 | + } |
| 95 | + } |
| 96 | + } |
| 97 | + } |
| 98 | + |
| 99 | + pub fn gc(&mut self) { |
| 100 | + let now = Instant::now(); |
| 101 | + let size_before = self.chunks_in_process.len(); |
| 102 | + self.chunks_in_process |
| 103 | + .retain(|_, buffer| now.duration_since(buffer.last_chunk_ts) < CHUNK_BUFFER_MAX_AGE); |
| 104 | + let size_after = self.chunks_in_process.len(); |
| 105 | + |
| 106 | + tracing::debug!( |
| 107 | + ?size_before, |
| 108 | + ?size_after, |
| 109 | + "performed chunk buffer garbage collection" |
| 110 | + ); |
| 111 | + } |
| 112 | +} |
| 113 | + |
| 114 | +/// Splits a payload into chunks that fit within message size limits. |
| 115 | +/// |
| 116 | +/// This function handles chunking by accounting for different overhead |
| 117 | +/// between the first chunk (MessageStart) and subsequent chunks (MessageChunk). |
| 118 | +/// |
| 119 | +/// The first chunk carries additional metadata like the reply_subject and chunk_count, |
| 120 | +/// which means it has more protocol overhead and less room for payload data. |
| 121 | +/// Subsequent chunks only carry a chunk_index, allowing them to fit more payload. |
| 122 | +/// |
| 123 | +/// This optimization ensures: |
| 124 | +/// - Reply subject is only transmitted once (in MessageStart) |
| 125 | +/// - Maximum payload utilization in each chunk |
| 126 | +/// - Efficient bandwidth usage for multi-chunk messages |
| 127 | +/// |
| 128 | +/// # Returns |
| 129 | +/// A vector of payload chunks, where each chunk is sized to fit within the message limit |
| 130 | +/// after accounting for protocol overhead. |
| 131 | +pub fn split_payload_into_chunks( |
| 132 | + payload: &[u8], |
| 133 | + max_message_size: usize, |
| 134 | + message_id: [u8; 16], |
| 135 | + reply_subject: Option<&str>, |
| 136 | +) -> Result<Vec<Vec<u8>>> { |
| 137 | + // Calculate overhead for MessageStart (first chunk) |
| 138 | + let start_message = MessageStart { |
| 139 | + message_id, |
| 140 | + chunk_count: 1, |
| 141 | + reply_subject: reply_subject.map(|s| s.to_string()), |
| 142 | + payload: vec![], |
| 143 | + }; |
| 144 | + let start_ups_message = rivet_ups_protocol::UpsMessage { |
| 145 | + body: MessageBody::MessageStart(start_message), |
| 146 | + }; |
| 147 | + let start_overhead = UpsMessage::latest(start_ups_message) |
| 148 | + .serialize_with_embedded_version(PROTOCOL_VERSION)? |
| 149 | + .len(); |
| 150 | + |
| 151 | + // Calculate overhead for MessageChunk (subsequent chunks) |
| 152 | + let chunk_message = MessageChunk { |
| 153 | + message_id, |
| 154 | + chunk_index: 0, |
| 155 | + payload: vec![], |
| 156 | + }; |
| 157 | + let chunk_ups_message = rivet_ups_protocol::UpsMessage { |
| 158 | + body: MessageBody::MessageChunk(chunk_message), |
| 159 | + }; |
| 160 | + let chunk_overhead = UpsMessage::latest(chunk_ups_message) |
| 161 | + .serialize_with_embedded_version(PROTOCOL_VERSION)? |
| 162 | + .len(); |
| 163 | + |
| 164 | + // Calculate max payload sizes |
| 165 | + let first_chunk_max_payload = max_message_size.saturating_sub(start_overhead); |
| 166 | + let other_chunk_max_payload = max_message_size.saturating_sub(chunk_overhead); |
| 167 | + |
| 168 | + if first_chunk_max_payload == 0 || other_chunk_max_payload == 0 { |
| 169 | + bail!("message overhead exceeds max message size"); |
| 170 | + } |
| 171 | + |
| 172 | + // Calculate how many chunks we need |
| 173 | + if payload.len() <= first_chunk_max_payload { |
| 174 | + // Single chunk - all data fits in first message |
| 175 | + return Ok(vec![payload.to_vec()]); |
| 176 | + } |
| 177 | + |
| 178 | + // Multi-chunk: first chunk + remaining chunks |
| 179 | + let remaining_after_first = payload.len() - first_chunk_max_payload; |
| 180 | + let additional_chunks = |
| 181 | + (remaining_after_first + other_chunk_max_payload - 1) / other_chunk_max_payload; |
| 182 | + |
| 183 | + let mut chunks = Vec::new(); |
| 184 | + |
| 185 | + // First chunk (smaller due to reply_subject overhead) |
| 186 | + chunks.push(payload[..first_chunk_max_payload].to_vec()); |
| 187 | + |
| 188 | + // Subsequent chunks |
| 189 | + let mut offset = first_chunk_max_payload; |
| 190 | + for _ in 0..additional_chunks { |
| 191 | + let end = std::cmp::min(offset + other_chunk_max_payload, payload.len()); |
| 192 | + chunks.push(payload[offset..end].to_vec()); |
| 193 | + offset = end; |
| 194 | + } |
| 195 | + |
| 196 | + Ok(chunks) |
| 197 | +} |
| 198 | + |
| 199 | +/// Encodes a chunk to the resulting BARE message. |
| 200 | +pub fn encode_chunk( |
| 201 | + payload: Vec<u8>, |
| 202 | + chunk_idx: u32, |
| 203 | + chunk_count: u32, |
| 204 | + message_id: [u8; 16], |
| 205 | + reply_subject: Option<String>, |
| 206 | +) -> Result<Vec<u8>> { |
| 207 | + let body = if chunk_idx == 0 { |
| 208 | + // First chunk - MessageStart |
| 209 | + MessageBody::MessageStart(MessageStart { |
| 210 | + message_id, |
| 211 | + chunk_count, |
| 212 | + reply_subject, |
| 213 | + payload, |
| 214 | + }) |
| 215 | + } else { |
| 216 | + // Subsequent chunks - MessageChunk |
| 217 | + MessageBody::MessageChunk(MessageChunk { |
| 218 | + message_id, |
| 219 | + chunk_index: chunk_idx, |
| 220 | + payload, |
| 221 | + }) |
| 222 | + }; |
| 223 | + |
| 224 | + let ups_message = rivet_ups_protocol::UpsMessage { body }; |
| 225 | + UpsMessage::latest(ups_message).serialize_with_embedded_version(PROTOCOL_VERSION) |
| 226 | +} |
0 commit comments