slipstream/kv.rs
1use async_trait::async_trait;
2use std::fmt;
3use tokio::sync::mpsc::Sender;
4
5/// Opaque position in a watch stream for resuming after disconnect.
6///
7/// Backends store whatever they need to resume (NATS: u64 revision).
8/// Callers should treat this as opaque and only pass it back to
9/// `watch_all_from` / `watch_prefix_from`.
10#[derive(Debug, Clone, Default, PartialEq, Eq)]
11pub struct WatchCursor(VersionToken);
12
13impl WatchCursor {
14 /// No cursor — forces a full watch on next connect.
15 pub fn none() -> Self {
16 Self(VersionToken::unknown())
17 }
18
19 /// Returns true if this cursor has no position (will trigger full watch).
20 pub fn is_none(&self) -> bool {
21 self.0.is_unknown()
22 }
23
24 /// Create a cursor from a version token.
25 pub fn from_version(token: VersionToken) -> Self {
26 Self(token)
27 }
28
29 /// Create a cursor from a u64 revision (convenience for NATS).
30 pub fn from_u64(rev: u64) -> Self {
31 Self(VersionToken::from_u64(rev))
32 }
33
34 /// Try to extract as u64 revision.
35 #[must_use]
36 pub fn as_u64(&self) -> Option<u64> {
37 self.0.as_u64()
38 }
39
40 /// Access the underlying version token.
41 pub(crate) fn version(&self) -> &VersionToken {
42 &self.0
43 }
44}
45
46/// Error type for KV operations.
47///
48/// `KvError` is `Clone` so a single failure can fan out to multiple waiters
49/// (e.g. callers blocked on a shared connect result). The underlying backend
50/// errors — `std::io::Error`, the `async-nats` error types — are *not* `Clone`,
51/// so their detail is flattened into the message string at this boundary rather
52/// than retained as a `#[source]` cause. Keeping `KvError: Clone` across the
53/// object-safe `async_trait` surface is the deliberate trade-off; the cost is a
54/// structured cause chain, which is why the `String` variants carry pre-rendered
55/// context instead of a nested error.
56#[derive(Debug, Clone, thiserror::Error)]
57pub enum KvError {
58 #[error("store not connected")]
59 NotConnected,
60 #[error("connection failed: {0}")]
61 ConnectionFailed(String),
62 #[error("key not found")]
63 KeyNotFound,
64 /// Key already exists (create-if-not-exists conflict).
65 #[error("key already exists")]
66 AlreadyExists,
67 /// CAS conflict: current version doesn't match expected.
68 #[error("revision mismatch")]
69 RevisionMismatch,
70 #[error("deserialization error: {0}")]
71 DeserializationError(String),
72 #[error("serialization error: {0}")]
73 SerializationError(String),
74 #[error("watch error: {0}")]
75 WatchError(String),
76 #[error("operation failed: {0}")]
77 OperationFailed(String),
78 #[error("operation timed out")]
79 Timeout,
80 /// The watch cursor/revision is too old — the backend has compacted past it.
81 /// Callers should fall back to a full scan + watch.
82 #[error("watch cursor expired (compacted)")]
83 CursorExpired,
84}
85
86/// Opaque version token that abstracts store-specific versioning.
87///
88/// Different stores use different versioning schemes:
89/// - NATS: 8-byte u64 revision
90/// - FDB: 10-byte versionstamp
91/// - Redis: could be stream ID + sequence
92///
93/// Stored inline (no heap allocation) — fits up to 10 bytes, which covers
94/// every current backend.
95#[derive(Clone, Default, PartialEq, Eq, Hash)]
96pub struct VersionToken {
97 len: u8,
98 buf: [u8; 10],
99}
100
101impl fmt::Debug for VersionToken {
102 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
103 let bytes = self.as_bytes();
104 if let Some(v) = self.as_u64() {
105 write!(f, "VersionToken(u64: {v})")
106 } else if bytes.is_empty() {
107 write!(f, "VersionToken(unknown)")
108 } else {
109 write!(f, "VersionToken({bytes:?})")
110 }
111 }
112}
113
114impl VersionToken {
115 /// Create an empty/unknown version (for entries without version info).
116 pub fn unknown() -> Self {
117 Self::default()
118 }
119
120 /// Check if this is an unknown/empty version.
121 pub fn is_unknown(&self) -> bool {
122 self.len == 0
123 }
124
125 /// Create from NATS u64 revision.
126 pub fn from_u64(rev: u64) -> Self {
127 let mut buf = [0u8; 10];
128 buf[..8].copy_from_slice(&rev.to_be_bytes());
129 Self { len: 8, buf }
130 }
131
132 /// Create from FDB versionstamp (10 bytes).
133 ///
134 /// `cfg(test)` until a FoundationDB backend ships and the round-trip is
135 /// tested end-to-end: a 10-byte token has no `as_u64()`, so handing one to
136 /// the NATS backend's CAS path yields an unactionable `OperationFailed`.
137 /// Today it exists only for the snapshot length-prefixed-version tests; an
138 /// FDB backend should lift the gate (and the visibility) rather than add a
139 /// second constructor.
140 #[cfg(test)]
141 pub(crate) fn from_fdb_versionstamp(vs: &[u8; 10]) -> Self {
142 Self { len: 10, buf: *vs }
143 }
144
145 /// Try to extract as u64 (for NATS compatibility).
146 #[must_use]
147 pub fn as_u64(&self) -> Option<u64> {
148 if self.len == 8 {
149 Some(u64::from_be_bytes(self.buf[..8].try_into().unwrap_or_else(
150 |_| unreachable!("len == 8 guarantees an 8-byte slice"),
151 )))
152 } else {
153 None
154 }
155 }
156
157 /// Get the raw bytes.
158 pub fn as_bytes(&self) -> &[u8] {
159 &self.buf[..self.len as usize]
160 }
161
162 /// Create from raw bytes (crate-internal, e.g. snapshot deserialization).
163 ///
164 /// Returns `None` if `bytes` exceeds the 10-byte inline capacity. Silently
165 /// truncating instead would store a version that differs from the real
166 /// revision, causing every later CAS to fail with `RevisionMismatch` and no
167 /// actionable error — so an oversized token is rejected at the boundary
168 /// rather than absorbed. Callers parse a length-prefixed field that is
169 /// structurally bounded to 10 bytes, so `None` is unreachable in practice;
170 /// returning it (instead of panicking) keeps the failure mode a recoverable
171 /// format error for any future caller that lacks that guard.
172 #[must_use]
173 pub(crate) fn from_raw(bytes: &[u8]) -> Option<Self> {
174 if bytes.len() > 10 {
175 return None;
176 }
177 let len = bytes.len() as u8;
178 let mut buf = [0u8; 10];
179 buf[..len as usize].copy_from_slice(bytes);
180 Some(Self { len, buf })
181 }
182}
183
184/// A single key-value entry with metadata.
185#[derive(Debug, Clone)]
186pub struct KvEntry {
187 pub key: String,
188 pub value: Vec<u8>,
189 pub version: VersionToken,
190}
191
192/// Update event from a watch stream.
193#[derive(Debug, Clone)]
194pub enum KvUpdate {
195 /// Key was created or updated.
196 Put(KvEntry),
197 /// Key was deleted.
198 Delete { key: String, version: VersionToken },
199 /// Key was purged (NATS-specific: all history removed).
200 /// Stores without purge semantics should map this to Delete.
201 Purge { key: String, version: VersionToken },
202}
203
204impl KvUpdate {
205 /// Get the key affected by this update.
206 pub fn key(&self) -> &str {
207 match self {
208 KvUpdate::Put(e) => &e.key,
209 KvUpdate::Delete { key, .. } => key,
210 KvUpdate::Purge { key, .. } => key,
211 }
212 }
213
214 /// Get the version of this update.
215 pub fn version(&self) -> &VersionToken {
216 match self {
217 KvUpdate::Put(e) => &e.version,
218 KvUpdate::Delete { version, .. } => version,
219 KvUpdate::Purge { version, .. } => version,
220 }
221 }
222}
223
224/// Core read-only KV operations - the minimal interface every store must implement.
225#[async_trait]
226pub trait KvReader: Send + Sync {
227 /// Get a value by key. Returns `None` if the key doesn't exist.
228 ///
229 /// Backends that use empty-value tombstones (NATS: `delete_with_version`
230 /// writes an empty-value Put so concurrent CAS writers still conflict) also
231 /// return `None` for a *stored* empty value — `get()` cannot tell a real
232 /// `b""` apart from a tombstone. A caller using zero-length values as a
233 /// presence signal (locks, feature flags) must use [`entry`](Self::entry),
234 /// which exposes the raw record including empty-value Puts.
235 async fn get(&self, key: &str) -> Result<Option<KvEntry>, KvError>;
236
237 /// Get all keys matching a prefix. Returns keys only, not values.
238 async fn keys(&self, prefix: &str) -> Result<Vec<String>, KvError>;
239
240 /// Get multiple entries by prefix. Useful for bulk loading.
241 async fn scan(&self, prefix: &str) -> Result<Vec<KvEntry>, KvError>;
242
243 /// Get the raw entry for a key, including tombstones (empty-value Put
244 /// entries written by `delete_with_version`). Most callers should use
245 /// `get()` instead, which filters tombstones for consistency with `scan()`.
246 ///
247 /// REQUIRED (no default) — deliberately. A default delegating to `get()`
248 /// silently hid tombstones on any backend that forgot to override it,
249 /// which breaks CAS callers that need the tombstone's version: e.g.
250 /// [`ExportLease::try_acquire`](crate::ExportLease::try_acquire) reads an
251 /// abandoned (CAS-deleted) lease's version through `entry()` for its
252 /// takeover write — with a `get()` default it would see `None` and report
253 /// the round as live instead of stealing it. Backends without empty-value
254 /// tombstone semantics (where delete genuinely removes the key) should
255 /// implement this as a delegation to `get()` — explicitly, so the choice
256 /// is a reviewed decision rather than an inherited footgun.
257 async fn entry(&self, key: &str) -> Result<Option<KvEntry>, KvError>;
258}
259
260/// Watch capability - optional, not all stores support real-time updates.
261///
262/// The non-`_from` watches are **state-sync** streams: they first deliver the
263/// current value of every matching key (the "re-list", as a stream of puts plus
264/// any surviving delete markers), then live updates. A consumer starting with
265/// no cursor therefore converges on the full bucket state without a separate
266/// scan — and without the scan-to-watch race a separate scan would open. The
267/// `_from` variants skip the re-list and deliver only the delta past the cursor.
268#[async_trait]
269pub trait KvWatcher: Send + Sync {
270 /// Watch all keys: current state first, then live changes. Sends updates
271 /// through the channel. Returns when the watch ends or an error occurs.
272 async fn watch_all(&self, tx: Sender<KvUpdate>) -> Result<(), KvError>;
273
274 /// Watch keys matching a prefix: current state first, then live changes.
275 async fn watch_prefix(&self, prefix: &str, tx: Sender<KvUpdate>) -> Result<(), KvError>;
276
277 /// Watch keys matching ANY of `prefixes`, delivered through one channel.
278 ///
279 /// The contract is exactly the union of the prefixes — no other keys. A
280 /// backend with native multi-filter consumers (NATS server 2.10+) serves all
281 /// `prefixes` from a SINGLE consumer; that matters because consumers are a
282 /// per-stream resource (measured at ~tens of KB of server state each, growing
283 /// super-linearly past a few thousand on one stream), so a watcher scoped to N
284 /// prefixes must not cost N consumers.
285 async fn watch_prefixes(&self, prefixes: &[&str], tx: Sender<KvUpdate>) -> Result<(), KvError>;
286
287 /// Resume watching all keys from a previously saved cursor position.
288 ///
289 /// Returns `KvError::CursorExpired` if the backend has compacted past the
290 /// cursor — callers should fall back to a full `watch_all()`.
291 ///
292 /// Default implementation ignores the cursor and delegates to `watch_all()`.
293 async fn watch_all_from(
294 &self,
295 cursor: &WatchCursor,
296 tx: Sender<KvUpdate>,
297 ) -> Result<(), KvError> {
298 let _ = cursor;
299 self.watch_all(tx).await
300 }
301
302 /// Resume watching keys with a prefix from a previously saved cursor.
303 ///
304 /// Default implementation ignores the cursor and delegates to `watch_prefix()`.
305 async fn watch_prefix_from(
306 &self,
307 prefix: &str,
308 cursor: &WatchCursor,
309 tx: Sender<KvUpdate>,
310 ) -> Result<(), KvError> {
311 let _ = cursor;
312 self.watch_prefix(prefix, tx).await
313 }
314
315 /// Resume watching the union of `prefixes` from a previously saved cursor.
316 ///
317 /// Same single-consumer contract as [`watch_prefixes`](Self::watch_prefixes),
318 /// same delta semantics as the other `_from` variants: only updates past the
319 /// cursor are delivered, or [`KvError::CursorExpired`] if the backend has
320 /// compacted past it.
321 ///
322 /// Default implementation ignores the cursor and delegates to
323 /// `watch_prefixes()` — correct (the state-sync re-list is a superset of any
324 /// delta) but a full replay; backends that can seek a multi-filter stream
325 /// should override it.
326 async fn watch_prefixes_from(
327 &self,
328 prefixes: &[&str],
329 cursor: &WatchCursor,
330 tx: Sender<KvUpdate>,
331 ) -> Result<(), KvError> {
332 let _ = cursor;
333 self.watch_prefixes(prefixes, tx).await
334 }
335}
336
337/// Write operations - optional, edge proxy is primarily read-only.
338#[async_trait]
339pub trait KvWriter: Send + Sync {
340 /// Put a value. Returns the new version token.
341 async fn put(&self, key: &str, value: &[u8]) -> Result<VersionToken, KvError>;
342
343 /// Delete a key. Best-effort: may return `true` even if the key did not
344 /// exist (NATS does not report pre-existence). Use `get()` first if you
345 /// need to distinguish "deleted something" from "nothing to delete".
346 async fn delete(&self, key: &str) -> Result<bool, KvError>;
347
348 /// Create a key only if it doesn't exist.
349 /// Returns `AlreadyExists` if the key has a live value.
350 async fn create(&self, key: &str, value: &[u8]) -> Result<VersionToken, KvError>;
351
352 /// Compare-and-swap: update only if current version matches `expected`.
353 /// Returns `RevisionMismatch` on conflict.
354 async fn update(
355 &self,
356 key: &str,
357 value: &[u8],
358 expected: &VersionToken,
359 ) -> Result<VersionToken, KvError>;
360
361 /// CAS-gated delete: delete only if current version matches `expected`.
362 /// Returns `RevisionMismatch` on conflict.
363 /// Writes an empty value (logical delete) so concurrent writers get a conflict.
364 async fn delete_with_version(
365 &self,
366 key: &str,
367 expected: &VersionToken,
368 ) -> Result<bool, KvError>;
369}
370
371/// TTL support - optional, for stores that support key expiration.
372#[async_trait]
373pub trait KvTtl: KvWriter {
374 /// Put a value with TTL. Value expires after duration.
375 async fn put_with_ttl(
376 &self,
377 key: &str,
378 value: &[u8],
379 ttl: std::time::Duration,
380 ) -> Result<VersionToken, KvError>;
381}
382
383/// Purge support - optional, for stores that can reclaim a key's storage.
384///
385/// Unlike [`KvWriter::delete`] (which writes a delete marker) and
386/// [`KvWriter::delete_with_version`] (which writes an empty-value tombstone),
387/// `purge` removes a key *and reclaims its bytes*. On NATS this issues a
388/// rollup (`Nats-Rollup: sub`) that drops all prior revisions of the subject,
389/// so the bytes stop counting against the stream's `max_bytes`.
390///
391/// Use this to bound a bucket that has no `max_age`: dead keys deleted with
392/// `delete`/`delete_with_version` accumulate forever, but purged keys are
393/// reclaimed.
394#[async_trait]
395pub trait KvPurge: KvWriter {
396 /// Purge a key, reclaiming its storage. Idempotent: purging an absent key
397 /// is not an error.
398 async fn purge(&self, key: &str) -> Result<(), KvError>;
399}
400
401#[cfg(test)]
402mod tests {
403 use super::*;
404
405 #[test]
406 fn from_raw_roundtrips_within_capacity() {
407 // The largest token any backend uses is a 10-byte FDB versionstamp.
408 let bytes = [1u8, 2, 3, 4, 5, 6, 7, 8, 9, 10];
409 let token = VersionToken::from_raw(&bytes).expect("10 bytes is within capacity");
410 assert_eq!(token.as_bytes(), &bytes);
411
412 // An 8-byte token is still interpretable as a NATS u64 revision.
413 let rev = 0x0102_0304_0506_0708u64;
414 let token = VersionToken::from_raw(&rev.to_be_bytes()).expect("8 bytes is within capacity");
415 assert_eq!(token.as_u64(), Some(rev));
416
417 // Empty input is the "unknown" token.
418 assert!(
419 VersionToken::from_raw(&[])
420 .expect("empty is within capacity")
421 .is_unknown()
422 );
423 }
424
425 #[test]
426 fn from_raw_rejects_above_capacity() {
427 // 11 bytes exceeds the 10-byte inline buffer. This guards against a
428 // loosened `parse_cursor` bound ever feeding oversized data through —
429 // returning `None` surfaces the format/backend mismatch at its origin
430 // instead of silently truncating into a wrong revision.
431 assert!(VersionToken::from_raw(&[0u8; 11]).is_none());
432 }
433}