1use std::collections::{BTreeMap, VecDeque};
18use std::num::NonZeroUsize;
19use std::sync::Arc;
20use std::time::{Duration, SystemTime};
21
22use serde::{Deserialize, Serialize};
24use tokio::time::Instant;
25
26use super::config::MUTATION_RETENTION;
28use super::{FileProducer, MutationId, MutationWindow, StorageManager, StoreHistory};
29use crate::{Error, Identifier, MemorySize, Result};
30
31const ID_COPIES: usize = 5;
34const COMPLETION_BYTES: usize = 2 * std::mem::size_of::<usize>();
35const NODE_SHARE: usize = node_bytes().div_ceil(4);
36const LEDGER_BASE: usize = std::mem::size_of::<BTreeMap<MutationId, Entry>>() + std::mem::size_of::<VecDeque<MutationId>>() + node_bytes();
38const ENTRY_MINIMUM: usize =
40 ID_COPIES * std::mem::size_of::<MutationId>() + std::mem::size_of::<MutationValue>() + NODE_SHARE + COMPLETION_BYTES;
41
42#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
49#[serde(deny_unknown_fields)]
50pub struct MutationReceipt {
51 pub mutation: MutationId,
52 pub fingerprint: [u8; 32],
53 pub expires_at: SystemTime,
54 pub value: MutationValue,
55}
56
57pub(super) struct Receipts {
58 execution: crate::execution::Execution,
59 window: Identifier,
60 deadline: Instant,
61 expires_at: SystemTime,
62 entries: BTreeMap<MutationId, Entry>,
63 order: VecDeque<MutationId>,
65 bytes: usize,
66 bound: NonZeroUsize,
67}
68
69struct Entry {
70 fingerprint: [u8; 32],
71 deadline: Instant,
72 expires_at: SystemTime,
73 value: Option<MutationValue>,
74 bytes: usize,
75 completion: Arc<()>,
76}
77
78pub(super) struct Completion(Arc<()>);
79
80#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
83#[serde(tag = "kind", content = "value", rename_all = "snake_case", deny_unknown_fields)]
84pub enum MutationValue {
85 Cursor(super::StorageCursor),
86 Deleted(super::StagedFile),
87 Producer(FileProducer),
88 Reference(Box<super::FileReference>),
89 Progress(super::FileProgress),
90 Rejected,
91}
92
93pub(super) enum ReceiptState {
94 Missing,
95 Pending,
96 Complete(MutationValue),
97}
98
99impl Entry {
104 fn reclaimable(&self) -> bool {
105 self.value.is_some() && Arc::strong_count(&self.completion) == 1
106 }
107}
108
109impl Receipts {
110 pub(super) fn new(retention: Duration, bound: NonZeroUsize, execution: crate::execution::Execution) -> Result<Self> {
111 let now = execution.now();
112 let wall = SystemTime::now();
113 Ok(Self {
114 execution,
115 window: epoch()?,
116 deadline: now + retention,
117 expires_at: wall + retention,
118 entries: BTreeMap::new(),
119 order: VecDeque::new(),
120 bytes: 0,
121 bound,
122 })
123 }
124
125 pub(super) fn recovery_limits(&self) -> super::ReadLimits {
126 super::ReadLimits {
127 rows: NonZeroUsize::new(self.bound.get() / ENTRY_MINIMUM).unwrap_or(NonZeroUsize::MIN),
128 bytes: self.bound.saturating_add(std::mem::size_of::<Vec<MutationReceipt>>()),
129 }
130 }
131
132 pub(super) fn window(&mut self, history: &StoreHistory, retention: Duration) -> Result<MutationWindow> {
133 let now = self.execution.now();
134 self.purge(now);
135 if now >= self.deadline {
136 self.window = epoch()?;
137 self.deadline = now + retention;
138 self.expires_at = SystemTime::now().checked_add(retention).ok_or_else(expired)?;
139 }
140 Ok(MutationWindow { history: history.clone(), id: self.window.clone(), expires_at: self.expires_at })
141 }
142
143 pub(super) fn lookup(&self, id: &MutationId, fingerprint: &[u8; 32], completion: Option<&Completion>) -> Result<ReceiptState> {
144 let now = self.execution.now();
145 if let Some(entry) = self.entries.get(id) {
146 let admitted = completion.is_some_and(|completion| Arc::ptr_eq(&completion.0, &entry.completion));
147 if !admitted && now >= entry.deadline {
148 tracing::error!(window = %id.window, mutation = %id.key, reason = "retry window expired", "storage mutation rejected");
149 return Err(expired());
150 }
151 if entry.fingerprint != *fingerprint {
152 tracing::error!(window = %id.window, mutation = %id.key, reason = "identity reused with different request", "storage mutation rejected");
153 return Err(Error::InvalidInput { field: "storage mutation", reason: "identity reused with different request" });
154 }
155 return Ok(match &entry.value {
156 Some(value) => ReceiptState::Complete(value.clone()),
157 None => ReceiptState::Pending,
158 });
159 }
160 if id.window != self.window || now >= self.deadline {
161 tracing::error!(window = %id.window, mutation = %id.key, reason = "retry window expired or unknown", "storage mutation rejected");
162 return Err(expired());
163 }
164 Ok(ReceiptState::Missing)
165 }
166
167 pub(super) fn completion(&self, id: &MutationId) -> Completion {
168 Completion(self.entries.get(id).expect("admitted receipt").completion.clone())
169 }
170
171 pub(super) fn insert(&mut self, id: &MutationId, fingerprint: [u8; 32], outcome_bytes: usize) -> Result<()> {
172 self.purge(self.execution.now());
173 if !matches!(self.lookup(id, &fingerprint, None)?, ReceiptState::Missing) {
174 return Err(Error::AlreadyRegistered { domain: "storage mutation", name: id.key.to_string() });
175 }
176 let bytes = charge(id, outcome_bytes)?;
177 while self.bytes.saturating_add(LEDGER_BASE).saturating_add(bytes) > self.bound.get() {
178 self.evict_oldest()?;
179 }
180 self.order.push_back(id.clone());
181 self.entries.insert(
182 id.clone(),
183 Entry { fingerprint, deadline: self.deadline, expires_at: self.expires_at, value: None, bytes, completion: Arc::new(()) },
184 );
185 self.bytes += bytes;
186 Ok(())
187 }
188
189 pub(super) fn prepared(&self, id: &MutationId, value: MutationValue) -> MutationReceipt {
190 let entry = self.entries.get(id).expect("admitted receipt");
191 MutationReceipt { mutation: id.clone(), fingerprint: entry.fingerprint, expires_at: entry.expires_at, value }
192 }
193
194 pub(super) fn complete(&mut self, id: &MutationId, value: MutationValue) {
195 let entry = self.entries.get_mut(id).expect("pending receipt retained through completion");
196 if entry.value.is_some() {
197 return;
198 }
199 assert!(charge(id, value.memory_size()).is_ok_and(|bytes| bytes <= entry.bytes));
201 entry.value = Some(value);
202 }
203
204 pub(super) fn snapshot_bytes(&self) -> usize {
205 if self.entries.is_empty() { 0 } else { self.bytes + LEDGER_BASE }
206 }
207
208 pub(super) fn clear(&mut self) {
209 self.entries = BTreeMap::new();
210 self.order = VecDeque::new();
211 self.bytes = 0;
212 }
213
214 fn evict_oldest(&mut self) -> Result<()> {
216 let index = self.order.iter().position(|id| self.entries.get(id).is_some_and(Entry::reclaimable)).ok_or_else(exhausted)?;
217 let id = self.order.remove(index).expect("ordered receipt");
218 let entry = self.entries.remove(&id).expect("ordered receipt");
219 self.bytes -= entry.bytes;
220 Ok(())
221 }
222
223 pub(super) fn snapshot(&self, through: Option<&super::StorageCursor>) -> Vec<MutationReceipt> {
224 let now = self.execution.now();
225 let count = self.entries.values().filter(|entry| entry.value.is_some() && now < entry.deadline).count();
226 let mut receipts = Vec::with_capacity(count);
227 for (mutation, entry) in &self.entries {
228 if let Some(value) = &entry.value
229 && now < entry.deadline
230 && match value {
231 MutationValue::Reference(reference) => {
232 through.is_none_or(|through| through.covers(&reference.position).unwrap_or(false))
233 }
234 MutationValue::Cursor(cursor) => through.is_none_or(|through| through.covers(cursor).unwrap_or(false)),
235 _ => true,
236 }
237 {
238 receipts.push(MutationReceipt {
239 mutation: mutation.clone(),
240 fingerprint: entry.fingerprint,
241 expires_at: entry.expires_at,
242 value: value.clone(),
243 });
244 }
245 }
246 receipts
247 }
248
249 pub(super) fn restore(&mut self, mut receipts: Vec<MutationReceipt>) -> Result<()> {
250 let now = self.execution.now();
251 let wall = SystemTime::now();
252 receipts.sort_by_key(|receipt| receipt.expires_at);
253 let mut restored = BTreeMap::new();
254 let mut order = VecDeque::with_capacity(receipts.len());
255 let mut total = 0usize;
256 for receipt in receipts {
257 let Ok(remaining) = receipt.expires_at.duration_since(wall) else {
258 continue;
259 };
260 let remaining = remaining.min(MUTATION_RETENTION);
261 if remaining.is_zero() {
262 continue;
263 }
264 let bytes = charge(&receipt.mutation, receipt.value.memory_size())?;
265 total = total.checked_add(bytes).ok_or_else(exhausted)?;
266 let entry = Entry {
267 fingerprint: receipt.fingerprint,
268 deadline: now + remaining,
269 expires_at: receipt.expires_at,
270 value: Some(receipt.value),
271 bytes,
272 completion: Arc::new(()),
273 };
274 order.push_back(receipt.mutation.clone());
275 if restored.insert(receipt.mutation, entry).is_some() {
276 return Err(Error::InvalidInput { field: "mutation recovery", reason: "duplicate receipt identity" });
277 }
278 }
279 self.entries = restored;
280 self.order = order;
281 self.bytes = total;
282 while self.snapshot_bytes() > self.bound.get() {
283 self.evict_oldest()?;
284 }
285 Ok(())
286 }
287
288 pub(super) fn purge(&mut self, now: Instant) {
291 self.order.retain(|id| {
292 let remove = self.entries.get(id).is_none_or(|entry| now >= entry.deadline && entry.reclaimable());
293 if remove && let Some(entry) = self.entries.remove(id) {
294 self.bytes -= entry.bytes;
295 }
296 !remove
297 });
298 if self.entries.is_empty() {
299 self.clear();
300 }
301 }
302}
303
304impl<C> StorageManager<C> {
305 #[tracing::instrument(name = "storage.receipts.recover_receipts", level = "debug", skip_all)]
306 pub(super) async fn recover_receipts(&self) -> Result<()> {
307 if !self.inner.primary.supports_mutation_receipts() {
308 return Ok(());
309 }
310 let limits = self.inner.state.lock().unwrap().receipts.recovery_limits();
311 let bytes = limits.bytes.get();
312 let memory = super::capacity::Reservation::lifecycle(&self.inner, super::capacity::Pool::Read, bytes)?.retain();
313 let receipts =
314 self.inner.bounded("primary.recover_mutations", self.inner.primary.recover_mutations(limits, memory.clone())).await?;
315 if receipts.len() > limits.rows.get() || receipts.memory_size() > bytes {
316 return Err(Error::InvalidInput { field: "mutation recovery", reason: "backend exceeded recovery bounds" });
317 }
318 for receipt in &receipts {
319 if let MutationValue::Deleted(file) = &receipt.value {
320 self.inner.primary.history().check(&file.history)?;
321 }
322 if let MutationValue::Cursor(cursor) = &receipt.value {
323 let state = self.inner.state.lock().unwrap();
324 if !state.status.saved.covers(cursor)? {
325 return Err(Error::InvalidInput { field: "mutation recovery", reason: "receipt exceeds saved data frontier" });
326 }
327 }
328 if let MutationValue::Producer(producer) = &receipt.value {
329 self.inner.primary.history().check(&producer.session.history)?;
330 if producer.session.generation == 0 {
331 return Err(Error::InvalidInput { field: "mutation recovery", reason: "zero producer generation" });
332 }
333 }
334 }
335 for receipt in &receipts {
336 if let MutationValue::Reference(reference) = &receipt.value {
337 self.inner.primary.history().check(&reference.position.history)?;
338 reference
339 .check_native()
340 .map_err(|_| Error::InvalidInput { field: "mutation recovery", reason: "invalid completed file reference" })?;
341 }
342 }
343 let mut state = self.inner.state.lock().unwrap();
344 state.receipts.restore(receipts)?;
345 state.collect();
346 Ok(())
347 }
348
349 pub fn info(&self) -> Result<crate::storage::StorageInfo> {
350 let mut state = self.inner.state.lock().unwrap();
351 if !state.ready || state.status.closing {
352 return Err(Error::NotReady { domain: "storage" });
353 }
354 let mutations = if self.inner.primary.supports_mutation_receipts() {
355 Some(state.receipts.window(self.inner.primary.history(), MUTATION_RETENTION)?)
356 } else {
357 None
358 };
359 Ok(crate::storage::StorageInfo {
360 accepted: state.status.accepted.clone(),
361 saved: state.status.saved.clone(),
362 mutations,
363 file_batch_bytes: self.inner.settings.response.batch_bytes.get() as u64,
364 })
365 }
366
367 pub fn mutation_window(&self) -> Result<MutationWindow> {
368 if !self.inner.primary.supports_mutation_receipts() {
369 return Err(Error::InvalidInput { field: "storage mutations", reason: "backend does not support mutation receipts" });
370 }
371 let mut state = self.inner.state.lock().unwrap();
372 if !state.ready || state.status.closing {
373 return Err(Error::NotReady { domain: "storage mutations" });
374 }
375 let window = state.receipts.window(self.inner.primary.history(), MUTATION_RETENTION)?;
376 state.collect();
377 Ok(window)
378 }
379}
380
381impl MutationValue {
382 pub(super) fn deleted(self) -> Result<super::StagedFile> {
383 match self {
384 Self::Deleted(file) => Ok(file),
385 _ => Err(rejected()),
386 }
387 }
388
389 pub(super) fn cursor(self) -> Result<super::StorageCursor> {
390 match self {
391 Self::Cursor(value) => Ok(value),
392 _ => Err(rejected()),
393 }
394 }
395
396 pub(super) fn producer(self) -> Result<FileProducer> {
397 match self {
398 Self::Producer(value) => Ok(value),
399 _ => Err(rejected()),
400 }
401 }
402
403 pub(super) fn reference(self) -> Result<super::FileReference> {
404 match self {
405 Self::Reference(value) => Ok(*value),
406 _ => Err(rejected()),
407 }
408 }
409
410 pub(super) fn progress(self) -> Result<super::FileProgress> {
411 match self {
412 Self::Progress(value) => Ok(value),
413 _ => Err(rejected()),
414 }
415 }
416}
417
418impl MemorySize for MutationValue {
419 fn memory_size(&self) -> usize {
420 std::mem::size_of::<Self>()
421 + match self {
422 Self::Deleted(value) => value.memory_size() - std::mem::size_of::<super::StagedFile>(),
423 Self::Cursor(value) => value.memory_size() - std::mem::size_of::<super::StorageCursor>(),
424 Self::Producer(value) => value.memory_size() - std::mem::size_of::<FileProducer>(),
425 Self::Reference(value) => value.as_ref().memory_size(),
426 Self::Progress(_) | Self::Rejected => 0,
427 }
428 }
429}
430impl MemorySize for MutationReceipt {
431 fn memory_size(&self) -> usize {
432 std::mem::size_of::<Self>() + self.mutation.memory_size() - std::mem::size_of::<MutationId>() + self.value.memory_size()
433 - std::mem::size_of::<MutationValue>()
434 }
435}
436
437fn epoch() -> Result<Identifier> {
442 let mut bytes = [0u8; 16];
443 getrandom::fill(&mut bytes)
444 .map_err(|source| Error::OperationFailed { operation: "storage retry window".into(), source: Arc::new(source) })?;
445 let name = bytes.iter().map(|byte| format!("{byte:02x}")).collect::<String>();
446 Identifier::new(name).map_err(|_| Error::InvalidInput { field: "storage retry window", reason: "invalid generated identity" })
447}
448
449fn charge(id: &MutationId, outcome_bytes: usize) -> Result<usize> {
454 id.memory_size()
455 .checked_mul(ID_COPIES)
456 .and_then(|bytes| bytes.checked_add(outcome_bytes))
457 .and_then(|bytes| bytes.checked_add(NODE_SHARE))
458 .and_then(|bytes| bytes.checked_add(COMPLETION_BYTES))
459 .ok_or_else(exhausted)
460}
461
462const fn node_bytes() -> usize {
463 let alignment = max(max(std::mem::align_of::<MutationId>(), std::mem::align_of::<Entry>()), std::mem::align_of::<usize>());
464 (std::mem::size_of::<MutationId>() + std::mem::size_of::<Entry>()) * 11
465 + 13 * std::mem::size_of::<usize>()
466 + 2 * std::mem::size_of::<u16>()
467 + 5 * alignment
468}
469
470const fn max(left: usize, right: usize) -> usize {
471 if left > right { left } else { right }
472}
473
474fn rejected() -> Error {
475 Error::InvalidInput { field: "file mutation", reason: "owned production ended without the requested result" }
476}
477fn exhausted() -> Error {
478 Error::ResourceExhausted { resource: "storage mutation receipts" }
479}
480fn expired() -> Error {
481 Error::InvalidInput { field: "storage mutation", reason: "retry window expired or unknown; reconcile state before issuing new work" }
482}
483
484