Skip to main content

xs/store/
mod.rs

1//! The event stream store.
2//!
3//! A [`Store`] is an append-only log of [`Frame`]s persisted to a directory on
4//! disk. Append events with [`Store::append`], replay and follow them with
5//! [`Store::read`], and stash payload bytes in the content-addressed store with
6//! the `cas_*` methods.
7//!
8//! ## Topics
9//!
10//! Every frame has a dot-delimited `topic` (for example `clip.add`). Topics form
11//! a hierarchy: a reader can ask for an exact topic, a prefix wildcard like
12//! `clip.*`, `*` for everything, or a comma-separated list of such patterns
13//! (`game.move.*,game.create`). See [`validate_topic`](crate::store::validate_topic)
14//! for the allowed characters and [`ReadOptions::topic`] for querying.
15//!
16//! ## Retention
17//!
18//! Each frame carries a [`TTL`] that controls how long it is kept: forever, for
19//! a fixed duration, only the last N per topic, or ephemeral (broadcast to live
20//! readers but never stored). A `time:` frame is indexed by its expiry and a
21//! background sweeper removes it within one sweep interval (the default, one
22//! second, is set with [`StoreOptions::ttl_sweep`]). Reads never return an
23//! expired frame, even one the sweeper has not reached yet.
24//!
25//! ## Durability
26//!
27//! Every append commits to the on-disk journal before it returns, so a process
28//! crash loses nothing. Only power loss or a kernel crash can lose the part of
29//! the journal not yet fsynced, and recovery then yields a stable prefix of the
30//! stream. An [`Fsync`] policy, set with [`Store::with_fsync`], decides when
31//! that fsync happens: after every append, on a timer (the default, one
32//! second), or never. [`Store::flush`] fsyncs on demand.
33
34mod fsync;
35pub use fsync::*;
36mod ttl;
37pub use ttl::*;
38
39#[cfg(test)]
40mod tests;
41
42use std::collections::{HashMap, HashSet};
43use std::ops::Bound;
44use std::path::PathBuf;
45use std::time::Duration;
46
47use tokio::sync::broadcast;
48use tokio::sync::mpsc::{self, UnboundedReceiver, UnboundedSender};
49
50use std::sync::atomic::{AtomicBool, AtomicU64, AtomicUsize, Ordering};
51use std::sync::{Arc, Mutex};
52
53use nu_protocol::engine::EngineState;
54use scru128::Scru128Id;
55
56use serde::{Deserialize, Deserializer, Serialize};
57
58use fjall::{
59    config::{BlockSizePolicy, HashRatioPolicy, RestartIntervalPolicy},
60    Database, Error as FjallError, Keyspace, KeyspaceCreateOptions, OwnedWriteBatch, PersistMode,
61};
62
63/// Error returned when opening a [`Store`].
64#[derive(Debug)]
65pub enum StoreError {
66    /// The store directory is already open in another process.
67    Locked,
68    /// An error from the underlying `fjall` database.
69    Other(FjallError),
70}
71
72impl std::fmt::Display for StoreError {
73    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
74        match self {
75            StoreError::Locked => write!(f, "Store is locked by another process"),
76            StoreError::Other(e) => write!(f, "{e}"),
77        }
78    }
79}
80
81impl std::error::Error for StoreError {}
82
83/// A single event in the stream.
84///
85/// A frame is metadata; the payload bytes (if any) live in the content-addressed
86/// store and are referenced by [`hash`](Frame::hash). Build one with the
87/// [`bon`](https://docs.rs/bon) builder, where the topic is the required
88/// starting argument:
89///
90/// ```
91/// use xs::{Frame, TTL};
92///
93/// let frame = Frame::builder("clip.add")
94///     .meta(serde_json::json!({ "source": "keyboard" }))
95///     .ttl(TTL::Last(100))
96///     .build();
97///
98/// assert_eq!(frame.topic, "clip.add");
99/// ```
100///
101/// The [`id`](Frame::id) is assigned by [`Store::append`] at write time, so the
102/// value you set on a builder is ignored when appending.
103#[derive(PartialEq, Eq, Serialize, Deserialize, Clone, Default, bon::Builder)]
104pub struct Frame {
105    /// Dot-delimited topic this frame belongs to (for example `clip.add`).
106    ///
107    /// Must satisfy [`validate_topic`]; [`Store::append`] rejects invalid topics.
108    #[builder(start_fn, into)]
109    pub topic: String,
110    /// Time-sortable identifier. Assigned by [`Store::append`]; any value set
111    /// before appending is overwritten.
112    #[builder(default)]
113    pub id: Scru128Id,
114    /// Integrity hash of the payload in the content-addressed store, if this
115    /// frame has one. Produce it with [`Store::cas_insert`].
116    pub hash: Option<ssri::Integrity>,
117    /// Arbitrary JSON metadata carried inline with the frame.
118    pub meta: Option<serde_json::Value>,
119    /// Retention policy for this frame. Defaults to [`TTL::Forever`] when unset.
120    pub ttl: Option<TTL>,
121}
122
123use std::fmt;
124
125impl fmt::Debug for Frame {
126    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
127        f.debug_struct("Frame")
128            .field("id", &format!("{id}", id = self.id))
129            .field("topic", &self.topic)
130            .field("hash", &self.hash.as_ref().map(|x| format!("{x}")))
131            .field("meta", &self.meta)
132            .field("ttl", &self.ttl)
133            .finish()
134    }
135}
136
137impl<'de> Deserialize<'de> for FollowOption {
138    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
139    where
140        D: Deserializer<'de>,
141    {
142        let s: String = Deserialize::deserialize(deserializer)?;
143        if s.is_empty() || s == "yes" {
144            Ok(FollowOption::On)
145        } else if let Ok(duration) = s.parse::<u64>() {
146            Ok(FollowOption::WithHeartbeat(Duration::from_millis(duration)))
147        } else {
148            match s.as_str() {
149                "true" => Ok(FollowOption::On),
150                "false" | "no" => Ok(FollowOption::Off),
151                _ => Err(serde::de::Error::custom("Invalid value for follow option")),
152            }
153        }
154    }
155}
156
157fn deserialize_bool<'de, D>(deserializer: D) -> Result<bool, D::Error>
158where
159    D: Deserializer<'de>,
160{
161    let s: String = Deserialize::deserialize(deserializer)?;
162    match s.as_str() {
163        "false" | "no" | "0" => Ok(false),
164        _ => Ok(true),
165    }
166}
167
168/// Options controlling a [`Store::read`] or [`Store::read_sync`] call.
169///
170/// Defaults replay every stored frame once, oldest first, then stop. Build a
171/// query with the [`bon`](https://docs.rs/bon) builder:
172///
173/// ```
174/// use xs::{ReadOptions, FollowOption};
175///
176/// // Replay history for one topic, then keep streaming new appends.
177/// let opts = ReadOptions::builder()
178///     .topic("clip.*".to_string())
179///     .follow(FollowOption::On)
180///     .build();
181/// ```
182#[derive(PartialEq, Deserialize, Clone, Debug, Default, bon::Builder)]
183pub struct ReadOptions {
184    /// Whether to keep streaming live appends after history is replayed.
185    /// Defaults to [`FollowOption::Off`].
186    #[serde(default)]
187    #[builder(default)]
188    pub follow: FollowOption,
189    /// Skip historical frames and emit only appends made after the read starts.
190    #[serde(default, deserialize_with = "deserialize_bool")]
191    #[builder(default)]
192    pub new: bool,
193    /// Start after this ID (exclusive).
194    #[serde(rename = "after")]
195    pub after: Option<Scru128Id>,
196    /// Start from this ID (inclusive).
197    pub from: Option<Scru128Id>,
198    /// Stop after emitting this many historical frames.
199    pub limit: Option<usize>,
200    /// Return the last N frames (most recent), in chronological order.
201    pub last: Option<usize>,
202    /// Restrict to one or more topic patterns, comma-separated. Each pattern
203    /// is an exact name, a `prefix.*` wildcard, or `*` for all, e.g.
204    /// `clip.add` or `game.move.*,game.create`. Commas are a safe separator
205    /// because [`validate_topic`] forbids them in topic names.
206    pub topic: Option<String>,
207}
208
209/// A single topic pattern: an exact topic name or a `prefix.*` wildcard.
210#[derive(Clone, Debug, PartialEq)]
211pub enum Pattern {
212    /// Matches the topic exactly.
213    Exact(String),
214    /// Matches any topic starting with this prefix. Stored with the trailing
215    /// dot, so `a.*` becomes `Prefix("a.")` and matches `a.b` but not `a`.
216    Prefix(String),
217}
218
219impl Pattern {
220    fn matches(&self, topic: &str) -> bool {
221        match self {
222            Pattern::Exact(t) => t == topic,
223            Pattern::Prefix(p) => topic.starts_with(p.as_str()),
224        }
225    }
226}
227
228/// A parsed topic filter: the value of [`ReadOptions::topic`] split on commas
229/// into individual [`Pattern`]s.
230///
231/// Note that synthetic control frames emitted by a following read --
232/// `xs.threshold` and heartbeat `xs.pulse` -- are delivered directly on the
233/// read channel and are never subject to this filter.
234#[derive(Clone, Debug, PartialEq)]
235pub enum TopicFilter {
236    /// Match every topic (no filter, or some element was `*`).
237    All,
238    /// Match topics satisfying at least one of these patterns.
239    Patterns(Vec<Pattern>),
240}
241
242impl TopicFilter {
243    /// Parse a comma-separated pattern list. Empty elements are ignored; an
244    /// empty or all-`*` spec yields [`TopicFilter::All`].
245    pub fn parse(spec: &str) -> TopicFilter {
246        let mut patterns = Vec::new();
247        for part in spec.split(',') {
248            let part = part.trim();
249            if part.is_empty() {
250                continue;
251            }
252            if part == "*" {
253                return TopicFilter::All;
254            }
255            if let Some(prefix) = part.strip_suffix(".*") {
256                patterns.push(Pattern::Prefix(format!("{prefix}.")));
257            } else {
258                patterns.push(Pattern::Exact(part.to_string()));
259            }
260        }
261        if patterns.is_empty() {
262            TopicFilter::All
263        } else {
264            TopicFilter::Patterns(patterns)
265        }
266    }
267
268    /// Parse an optional spec; `None` means no filter.
269    pub fn from_option(spec: Option<&str>) -> TopicFilter {
270        match spec {
271            Some(s) => TopicFilter::parse(s),
272            None => TopicFilter::All,
273        }
274    }
275
276    /// Does `topic` match this filter?
277    pub fn matches(&self, topic: &str) -> bool {
278        match self {
279            TopicFilter::All => true,
280            TopicFilter::Patterns(patterns) => patterns.iter().any(|p| p.matches(topic)),
281        }
282    }
283}
284
285/// K-way merge of per-pattern frame iterators, ordered by frame id, with
286/// dedupe when overlapping patterns yield the same frame from more than one
287/// iterator. Each input iterator must itself be sorted by id (ascending when
288/// `descending` is false, descending otherwise).
289struct MergeById<'a> {
290    iters: Vec<Box<dyn Iterator<Item = Frame> + 'a>>,
291    heap: std::collections::BinaryHeap<MergeEntry>,
292    descending: bool,
293    last_emitted: Option<Scru128Id>,
294}
295
296struct MergeEntry {
297    key: u128,
298    idx: usize,
299    frame: Frame,
300}
301
302impl PartialEq for MergeEntry {
303    fn eq(&self, other: &Self) -> bool {
304        self.key == other.key && self.idx == other.idx
305    }
306}
307impl Eq for MergeEntry {}
308impl PartialOrd for MergeEntry {
309    fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
310        Some(self.cmp(other))
311    }
312}
313impl Ord for MergeEntry {
314    fn cmp(&self, other: &Self) -> std::cmp::Ordering {
315        self.key.cmp(&other.key).then(self.idx.cmp(&other.idx))
316    }
317}
318
319/// BinaryHeap is a max-heap; for ascending merges flip the bits so the
320/// smallest id pops first.
321fn merge_key(id: &Scru128Id, descending: bool) -> u128 {
322    let raw = u128::from_be_bytes(id.to_bytes());
323    if descending {
324        raw
325    } else {
326        !raw
327    }
328}
329
330impl<'a> MergeById<'a> {
331    fn new(mut iters: Vec<Box<dyn Iterator<Item = Frame> + 'a>>, descending: bool) -> Self {
332        let mut heap = std::collections::BinaryHeap::with_capacity(iters.len());
333        for (idx, iter) in iters.iter_mut().enumerate() {
334            if let Some(frame) = iter.next() {
335                heap.push(MergeEntry {
336                    key: merge_key(&frame.id, descending),
337                    idx,
338                    frame,
339                });
340            }
341        }
342        MergeById {
343            iters,
344            heap,
345            descending,
346            last_emitted: None,
347        }
348    }
349}
350
351impl Iterator for MergeById<'_> {
352    type Item = Frame;
353
354    fn next(&mut self) -> Option<Frame> {
355        loop {
356            let entry = self.heap.pop()?;
357            if let Some(frame) = self.iters[entry.idx].next() {
358                self.heap.push(MergeEntry {
359                    key: merge_key(&frame.id, self.descending),
360                    idx: entry.idx,
361                    frame,
362                });
363            }
364            // Overlapping patterns can yield the same frame from more than
365            // one iterator; emit it once.
366            if self.last_emitted == Some(entry.frame.id) {
367                continue;
368            }
369            self.last_emitted = Some(entry.frame.id);
370            return Some(entry.frame);
371        }
372    }
373}
374
375impl ReadOptions {
376    /// Parse options from a URL query string (the form used by the HTTP API),
377    /// for example `follow=true&topic=clip.*&last=10`. `None` yields the
378    /// defaults.
379    pub fn from_query(query: Option<&str>) -> Result<Self, crate::error::Error> {
380        match query {
381            Some(q) => Ok(serde_urlencoded::from_str(q)?),
382            None => Ok(Self::default()),
383        }
384    }
385
386    /// Render these options back into a URL query string, the inverse of
387    /// [`from_query`](ReadOptions::from_query). Returns an empty string when no
388    /// options are set.
389    pub fn to_query_string(&self) -> String {
390        let mut params = Vec::new();
391
392        // Add follow parameter with heartbeat if specified
393        match self.follow {
394            FollowOption::Off => {}
395            FollowOption::On => params.push(("follow", "true".to_string())),
396            FollowOption::WithHeartbeat(duration) => {
397                params.push(("follow", duration.as_millis().to_string()));
398            }
399        }
400
401        // Add new if true
402        if self.new {
403            params.push(("new", "true".to_string()));
404        }
405
406        // Add after if present
407        if let Some(after) = self.after {
408            params.push(("after", after.to_string()));
409        }
410
411        // Add from if present
412        if let Some(from) = self.from {
413            params.push(("from", from.to_string()));
414        }
415
416        // Add limit if present
417        if let Some(limit) = self.limit {
418            params.push(("limit", limit.to_string()));
419        }
420
421        // Add last if present
422        if let Some(last) = self.last {
423            params.push(("last", last.to_string()));
424        }
425
426        if let Some(topic) = &self.topic {
427            params.push(("topic", topic.clone()));
428        }
429
430        // Return empty string if no params
431        if params.is_empty() {
432            String::new()
433        } else {
434            url::form_urlencoded::Serializer::new(String::new())
435                .extend_pairs(params)
436                .finish()
437        }
438    }
439}
440
441/// Whether a read keeps streaming after history is replayed.
442#[derive(Default, PartialEq, Clone, Debug)]
443pub enum FollowOption {
444    /// Stop once historical frames are exhausted.
445    #[default]
446    Off,
447    /// Replay history, then stream live appends indefinitely.
448    On,
449    /// Like [`On`](FollowOption::On), but also emit a periodic heartbeat frame
450    /// at the given interval so idle readers can detect a live connection.
451    WithHeartbeat(Duration),
452}
453
454#[derive(Debug)]
455enum GCTask {
456    CheckLastTTL {
457        topic: String,
458        keep: u32,
459    },
460    /// Remove every `time:` frame whose expiry has passed, up to
461    /// `MAX_SWEEP_PER_DRAIN`. Sent by the sweep tick.
462    Sweep,
463    Drain(tokio::sync::oneshot::Sender<()>),
464}
465
466/// A frame the gc worker is about to delete.
467enum Removal {
468    /// Only the id is known. The frame is fetched to learn its keys.
469    Id(Scru128Id),
470    /// An index hit that already carries every key the removal needs, so no
471    /// point read into `stream` is required. `expires_at` is `None` when the
472    /// frame has no `time:` ttl and so has no expiry index entry.
473    Indexed {
474        id: Scru128Id,
475        topic: String,
476        expires_at: Option<u64>,
477    },
478}
479
480impl Removal {
481    fn id(&self) -> Scru128Id {
482        match self {
483            Removal::Id(id) | Removal::Indexed { id, .. } => *id,
484        }
485    }
486}
487
488/// Where the next expiry sweep starts scanning. Everything below it has been
489/// swept, so the sweep steps over the tombstones earlier sweeps left instead
490/// of walking them again.
491///
492/// It only ever moves up to a key a sweep actually removed, and a sweep never
493/// looks past the frames that are due, so it can never come to rest above an
494/// entry that is still waiting for its expiry to pass. The one way an entry
495/// can appear below it is an index write, which
496/// [`note_indexed_expiry`](Store::note_indexed_expiry) moves it back for.
497#[derive(Clone, Copy, PartialEq, Eq, Debug)]
498struct SweepStart {
499    key: [u8; 24],
500    /// Whether `key` itself still has to be swept: false on the key a sweep
501    /// removed, true on one a later index write put back in play.
502    inclusive: bool,
503}
504
505impl SweepStart {
506    fn bound(&self) -> Bound<[u8; 24]> {
507        if self.inclusive {
508            Bound::Included(self.key)
509        } else {
510            Bound::Excluded(self.key)
511        }
512    }
513
514    /// Sort position, so two starts can be compared. An inclusive start sits
515    /// before an exclusive one on the same key: it covers one more entry.
516    fn rank(&self) -> ([u8; 24], u8) {
517        (self.key, u8::from(!self.inclusive))
518    }
519}
520
521/// Where a `last:n` trim scan of a topic starts. Below it the topic holds no
522/// live frame, only the tombstones of ones a trim or another removal path
523/// took, so a scan that starts here counts and trims exactly what a scan of
524/// the whole topic would.
525///
526/// A trim proves that of everything it removed. A scan that reaches the end of
527/// the topic without finding `keep` live frames proves it of everything under
528/// the oldest entry it saw, which is why the bound can be inclusive.
529///
530/// The one way a live frame can appear below it is an index write, which
531/// [`note_indexed_topic`](Store::note_indexed_topic) drops the floor for.
532#[derive(Clone, Copy, PartialEq, Eq, Debug)]
533struct TrimFloor {
534    id: Scru128Id,
535    /// Whether `id` itself is still on the topic and in scope for the scan.
536    inclusive: bool,
537}
538
539/// Options for opening a [`Store`]. Build one with the
540/// [`bon`](https://docs.rs/bon) builder; every field has a default.
541///
542/// ```
543/// use std::time::Duration;
544/// use xs::{Fsync, StoreOptions};
545///
546/// let options = StoreOptions::builder()
547///     .fsync(Fsync::Always)
548///     .ttl_sweep(Duration::from_millis(250))
549///     .build();
550/// assert_eq!(options.fsync, Fsync::Always);
551/// assert_eq!(StoreOptions::default().ttl_sweep, Duration::from_millis(1000));
552/// ```
553#[derive(Clone, Debug, PartialEq, Eq, bon::Builder)]
554pub struct StoreOptions {
555    /// When the journal is fsynced. See [`Fsync`].
556    #[builder(default)]
557    pub fsync: Fsync,
558    /// How often expired `time:` frames are swept from disk. Default 1000 ms.
559    /// Zero disables the sweeper; expired frames are then only hidden from
560    /// reads, never removed.
561    #[builder(default = Duration::from_millis(1000))]
562    pub ttl_sweep: Duration,
563}
564
565impl Default for StoreOptions {
566    fn default() -> Self {
567        StoreOptions::builder().build()
568    }
569}
570
571/// An append-only event stream backed by a directory on disk.
572///
573/// Open one with [`new`](Store::new). `Store` is cheaply [`Clone`]able: every
574/// clone shares the same underlying database, broadcast channel, and
575/// content-addressed store, so clone it freely across tasks and threads instead
576/// of wrapping it in an `Arc`.
577///
578/// See the [module docs](crate::store) for the topic and retention model.
579#[derive(Clone)]
580pub struct Store {
581    /// Directory backing this store (the path passed to [`new`](Store::new)).
582    pub path: PathBuf,
583    db: Database,
584    stream: Keyspace,
585    idx_topic: Keyspace,
586    /// `time:` frames keyed by expiry then id, valued by topic. The sweeper
587    /// range-scans it from [`sweep_start`](Store::sweep_start) up to now. See
588    /// [`idx_expiry_key`].
589    idx_expiry: Keyspace,
590    /// Where the next sweep resumes in `idx_expiry`; `None` is the start of
591    /// the index. In memory only: a reopened store sweeps from the start
592    /// again, which costs one walk of whatever tombstones outlived it and can
593    /// never skip a frame. See [`SweepStart`].
594    sweep_start: Arc<Mutex<Option<SweepStart>>>,
595    /// Per-topic `last:n` trim floor: every frame on the topic with an id at
596    /// or below it has been removed, so a trim scan stops there rather than
597    /// running on into the tombstones of the frames it trimmed last time. In
598    /// memory only, on the same reasoning as
599    /// [`sweep_start`](Store::sweep_start). See
600    /// [`last_overflow`](Store::last_overflow).
601    trim_floor: Arc<Mutex<HashMap<String, TrimFloor>>>,
602    broadcast_tx: broadcast::Sender<Frame>,
603    gc_tx: UnboundedSender<GCTask>,
604    /// Joins the background workers when the last caller-held clone drops.
605    /// Declared after `gc_tx` on purpose: fields drop in order, so the gc
606    /// channel closes before the join, and the gc worker can exit.
607    workers: Option<Arc<Workers>>,
608    append_lock: Arc<Mutex<()>>,
609    /// Runtime handle captured at [`new`](Store::new), used to spawn the follow
610    /// and heartbeat tasks from the now-sync [`read`](Store::read). `Handle` is
611    /// cheaply `Clone`, so every `Store` clone shares the same runtime.
612    rt: Option<tokio::runtime::Handle>,
613    /// Optional base engine the processors clone, set via [`with_base_engine`].
614    /// `None` falls back to `Engine::new()`. See
615    /// [`prepared_base`](crate::nu::prepared_base) and ADR 0007.
616    base_engine: Option<Arc<EngineState>>,
617    /// Fsync policy and dirty flag, shared by every clone and the fsync worker.
618    fsync: Arc<FsyncState>,
619    /// Counters about this store's own work, shared by every clone and by
620    /// the gc worker. See [`StoreStats`].
621    stats: Arc<StoreStats>,
622}
623
624/// Counters the store keeps about its own work, always on.
625///
626/// Relaxed atomics on paths that already touch the database, so the cost is
627/// noise next to the work being counted. Cumulative for the life of the
628/// process, with no way to reset: a caller wanting a rate subtracts two
629/// reads.
630#[derive(Debug, Default)]
631pub struct StoreStats {
632    /// Frames accepted by [`append`](Store::append), including ephemeral ones
633    /// that were broadcast but never stored.
634    pub appends: AtomicUsize,
635    /// Frames pulled from a keyspace iterator by a read. Larger than the
636    /// number returned whenever a read filters: the gap is wasted work.
637    pub frames_scanned: AtomicUsize,
638    /// Point reads into `stream`. A topic read does one per frame it returns,
639    /// so this over `frames_scanned` says how well the index is filtering.
640    pub stream_reads: AtomicUsize,
641    /// Frames removed by [`remove_many`](Store::remove_many).
642    pub removed_frames: AtomicUsize,
643    /// Batches committed by [`remove_many`](Store::remove_many). One per gc
644    /// drain that had anything to do.
645    pub remove_commits: AtomicUsize,
646    /// Times the gc worker drained its queue.
647    pub gc_drains: AtomicUsize,
648    /// Nanoseconds the gc worker spent inside a drain. Against `gc_drains`
649    /// this is the cost of retention, which is otherwise invisible.
650    pub gc_nanos: AtomicU64,
651}
652
653/// A serializable read of [`StoreStats`]. See its fields for what each counts.
654#[derive(Debug, Clone, serde::Serialize)]
655pub struct StoreStatsSnapshot {
656    pub appends: u64,
657    pub frames_scanned: u64,
658    pub stream_reads: u64,
659    pub removed_frames: u64,
660    pub remove_commits: u64,
661    pub gc_drains: u64,
662    pub gc_nanos: u64,
663}
664
665impl StoreStats {
666    fn snapshot(&self) -> StoreStatsSnapshot {
667        let n = |c: &AtomicUsize| c.load(Ordering::Relaxed) as u64;
668        StoreStatsSnapshot {
669            appends: n(&self.appends),
670            frames_scanned: n(&self.frames_scanned),
671            stream_reads: n(&self.stream_reads),
672            removed_frames: n(&self.removed_frames),
673            remove_commits: n(&self.remove_commits),
674            gc_drains: n(&self.gc_drains),
675            gc_nanos: self.gc_nanos.load(Ordering::Relaxed),
676        }
677    }
678}
679
680/// Cache hit rates for one keyspace, each `None` until that layer has been
681/// asked for something.
682#[derive(Debug, Clone, serde::Serialize)]
683pub struct KeyspaceRates {
684    pub block_cache_hit: Option<f64>,
685    pub data_block_cache_hit: Option<f64>,
686    pub index_block_cache_hit: Option<f64>,
687    pub filter_block_cache_hit: Option<f64>,
688    pub table_file_cache_hit: Option<f64>,
689    /// Share of filter lookups that saved a read. Only counted when the
690    /// filter reports a miss: a filter costs nothing on a hit, so this is
691    /// what it is worth, not how often it ran.
692    pub filter_efficiency: Option<f64>,
693}
694
695/// A rate is meaningless with no observations behind it.
696fn rate(observations: usize, value: f64) -> Option<f64> {
697    (observations > 0).then_some(value)
698}
699
700/// A point-in-time read of one keyspace's cache and filter counters.
701///
702/// Every rate is a fraction in `0.0..=1.0`, or `0.0` when nothing has been
703/// asked of that layer yet. The counters are cumulative since the process
704/// started; fjall has no way to reset them, so a caller comparing two reads
705/// should subtract.
706#[derive(Debug, Clone, serde::Serialize)]
707pub struct KeyspaceMetrics {
708    pub keyspace: &'static str,
709
710    /// Hit rates, or `null` where nothing has been asked of that layer yet.
711    /// fjall returns 1.0 for a rate with no observations, which reads as a
712    /// perfect cache on a store that has done no work.
713    pub rates: KeyspaceRates,
714
715    pub filter_queries: usize,
716    pub io_skipped_by_filter: usize,
717
718    /// Blocks asked for, and how many of those reached the filesystem.
719    pub block_loads: usize,
720    pub block_load_io: usize,
721    pub data_block_io: u64,
722    pub index_block_io: u64,
723    pub filter_block_io: u64,
724}
725
726impl KeyspaceMetrics {
727    fn read(keyspace: &'static str, ks: &Keyspace) -> Self {
728        // `Keyspace::metrics` hands back an `lsm_tree::Metrics`, a type fjall
729        // does not re-export, so the reference is never named.
730        let m = ks.metrics();
731        Self {
732            keyspace,
733            rates: KeyspaceRates {
734                block_cache_hit: rate(m.block_loads(), m.block_cache_hit_rate()),
735                data_block_cache_hit: rate(
736                    m.data_block_load_count(),
737                    m.data_block_cache_hit_rate(),
738                ),
739                index_block_cache_hit: rate(
740                    m.index_block_load_count(),
741                    m.index_block_cache_hit_rate(),
742                ),
743                filter_block_cache_hit: rate(
744                    m.filter_block_load_count(),
745                    m.filter_block_cache_hit_rate(),
746                ),
747                // No public count of table opens, so borrow the block loads:
748                // no block was loaded means no file was opened either.
749                table_file_cache_hit: rate(m.block_loads(), m.table_file_cache_hit_rate()),
750                filter_efficiency: rate(m.filter_queries(), m.filter_efficiency()),
751            },
752            filter_queries: m.filter_queries(),
753            io_skipped_by_filter: m.io_skipped_by_filter(),
754            block_loads: m.block_loads(),
755            block_load_io: m.block_load_io_count(),
756            data_block_io: m.data_block_io(),
757            index_block_io: m.index_block_io(),
758            filter_block_io: m.filter_block_io(),
759        }
760    }
761}
762
763/// Joins the background workers when the last caller-held `Store` drops.
764/// The gc worker's own clone carries `None`, so it never joins itself. The
765/// ticks stop first: the sweep tick holds a gc sender, and the gc worker
766/// only exits once every sender is gone. Once the joins return every worker
767/// has dropped its database handle, the database is closed, and the path
768/// can be reopened.
769struct Workers {
770    gc: Option<std::thread::JoinHandle<()>>,
771    fsync: Option<Tick>,
772    sweep: Option<Tick>,
773}
774
775impl Drop for Workers {
776    fn drop(&mut self) {
777        drop(self.fsync.take());
778        drop(self.sweep.take());
779        if let Some(handle) = self.gc.take() {
780            let _ = handle.join();
781        }
782    }
783}
784
785/// A thread that runs a closure every `interval` until dropped. Dropping it
786/// sets `stop`, unparks the thread, and joins it, so the join waits for at
787/// most one in-flight run of the closure.
788struct Tick {
789    stop: Arc<AtomicBool>,
790    handle: Option<std::thread::JoinHandle<()>>,
791}
792
793impl Tick {
794    fn spawn(interval: Duration, mut run: impl FnMut() + Send + 'static) -> Tick {
795        let stop = Arc::new(AtomicBool::new(false));
796        let handle = std::thread::spawn({
797            let stop = stop.clone();
798            move || loop {
799                // park_timeout may wake early; the flags checked by each
800                // closure make that harmless.
801                std::thread::park_timeout(interval);
802                if stop.load(Ordering::Acquire) {
803                    break;
804                }
805                run();
806            }
807        });
808        Tick {
809            stop,
810            handle: Some(handle),
811        }
812    }
813}
814
815impl Drop for Tick {
816    fn drop(&mut self) {
817        self.stop.store(true, Ordering::Release);
818        if let Some(handle) = self.handle.take() {
819            handle.thread().unpark();
820            let _ = handle.join();
821        }
822    }
823}
824
825/// Shared fsync bookkeeping.
826struct FsyncState {
827    mode: Fsync,
828    /// Set after every commit under `interval`; the tick clears it.
829    dirty: AtomicBool,
830}
831
832impl Store {
833    /// Open the store at `path`, creating the directory layout if it does not
834    /// exist. Spawns a background worker that garbage-collects expired frames.
835    /// Uses the default [`Fsync`] policy; see [`with_fsync`](Store::with_fsync)
836    /// to choose one.
837    ///
838    /// # Errors
839    ///
840    /// Returns [`StoreError::Locked`] if another process already holds the
841    /// store open, or [`StoreError::Other`] for any other database error.
842    ///
843    /// ```no_run
844    /// use xs::Store;
845    ///
846    /// let store = Store::new("./clipboard-store".into())?;
847    /// # Ok::<(), xs::StoreError>(())
848    /// ```
849    pub fn new(path: PathBuf) -> Result<Store, StoreError> {
850        Store::open(path, StoreOptions::default())
851    }
852
853    /// Open the store at `path` with an explicit [`Fsync`] policy. Otherwise
854    /// the same as [`new`](Store::new).
855    ///
856    /// ```no_run
857    /// use xs::{Store, Fsync};
858    ///
859    /// let store = Store::with_fsync("./clipboard-store".into(), Fsync::Always)?;
860    /// # Ok::<(), xs::StoreError>(())
861    /// ```
862    pub fn with_fsync(path: PathBuf, fsync: Fsync) -> Result<Store, StoreError> {
863        Store::open(path, StoreOptions::builder().fsync(fsync).build())
864    }
865
866    /// Open the store at `path` with explicit [`StoreOptions`]. Otherwise the
867    /// same as [`new`](Store::new).
868    ///
869    /// ```no_run
870    /// use std::time::Duration;
871    /// use xs::{Store, StoreOptions};
872    ///
873    /// let options = StoreOptions::builder()
874    ///     .ttl_sweep(Duration::from_millis(250))
875    ///     .build();
876    /// let store = Store::open("./clipboard-store".into(), options)?;
877    /// # Ok::<(), xs::StoreError>(())
878    /// ```
879    pub fn open(path: PathBuf, options: StoreOptions) -> Result<Store, StoreError> {
880        let fsync = options.fsync;
881        let db = match Database::builder(path.join("fjall"))
882            .cache_size(32 * 1024 * 1024) // 32 MiB
883            // One worker cannot keep up: every append writes three keyspaces, each
884            // calls check_memtable_rotate, and all of them queue into the single
885            // shared channel that worker also has to drain. At 1, flushes starve --
886            // 1M appends put only 10.7% of their bytes into segments, and 1.0% with
887            // the sweeper also committing. See BUGS-FOUND.md.
888            .worker_threads(2)
889            .open()
890        {
891            Ok(db) => db,
892            Err(FjallError::Locked) => return Err(StoreError::Locked),
893            Err(e) => return Err(StoreError::Other(e)),
894        };
895
896        // Options for stream keyspace: point reads by frame ID
897        let stream_opts = || {
898            KeyspaceCreateOptions::default()
899                .max_memtable_size(8 * 1024 * 1024) // 8 MiB
900                .data_block_size_policy(BlockSizePolicy::all(16 * 1024)) // 16 KiB
901                .data_block_hash_ratio_policy(HashRatioPolicy::all(8.0))
902                .expect_point_read_hits(true)
903        };
904
905        // Options for idx_topic keyspace: prefix scans only
906        let idx_opts = || {
907            KeyspaceCreateOptions::default()
908                .max_memtable_size(8 * 1024 * 1024) // 8 MiB
909                .data_block_size_policy(BlockSizePolicy::all(16 * 1024)) // 16 KiB
910                .data_block_hash_ratio_policy(HashRatioPolicy::all(0.0)) // no point reads
911                .expect_point_read_hits(true)
912        };
913
914        // An idx_topic key is `<topic>\0<id>`, so one topic repeats across
915        // every entry it has. A restart point writes a key in full instead of
916        // against the one before it, so fewer of them is a smaller index, and
917        // nothing binary-searches inside these blocks to pay for it.
918        let idx_topic_opts =
919            || idx_opts().data_block_restart_interval_policy(RestartIntervalPolicy::all(32));
920
921        // An idx_expiry key is `<expiry><id>`: consecutive keys share a prefix
922        // only when frames expire in the same millisecond, so the same change
923        // buys much less. Left at the default.
924        let stream = db.keyspace("stream", stream_opts).unwrap();
925        let idx_topic = db.keyspace("idx_topic", idx_topic_opts).unwrap();
926        let idx_expiry = db.keyspace("idx_expiry", idx_opts).unwrap();
927
928        let (broadcast_tx, _) = broadcast::channel(FOLLOW_BUFFER);
929        let (gc_tx, gc_rx) = mpsc::unbounded_channel();
930
931        let fsync_state = Arc::new(FsyncState {
932            mode: fsync,
933            dirty: AtomicBool::new(false),
934        });
935
936        let mut store = Store {
937            path: path.clone(),
938            db,
939            stream,
940            idx_topic,
941            idx_expiry,
942            sweep_start: Arc::new(Mutex::new(None)),
943            trim_floor: Arc::new(Mutex::new(HashMap::new())),
944            broadcast_tx,
945            gc_tx,
946            workers: None,
947            append_lock: Arc::new(Mutex::new(())),
948            rt: tokio::runtime::Handle::try_current().ok(),
949            base_engine: None,
950            fsync: fsync_state,
951            stats: Arc::new(StoreStats::default()),
952        };
953
954        // Spawn the workers. The gc worker's clone is taken while `workers` is
955        // still `None`, so only callers' clones join them.
956        let gc = spawn_gc_worker(gc_rx, store.clone());
957        let fsync_tick = match fsync {
958            Fsync::Interval(interval) => Some(spawn_fsync_tick(
959                store.db.clone(),
960                store.fsync.clone(),
961                interval,
962            )),
963            Fsync::Always | Fsync::Never => None,
964        };
965        let sweep_tick = if options.ttl_sweep.is_zero() {
966            None
967        } else {
968            Some(spawn_sweep_tick(store.gc_tx.clone(), options.ttl_sweep))
969        };
970        store.workers = Some(Arc::new(Workers {
971            gc: Some(gc),
972            fsync: fsync_tick,
973            sweep: sweep_tick,
974        }));
975
976        Ok(store)
977    }
978
979    /// Counters about this store's own work: appends, how much a read
980    /// scanned against how much it returned, and what retention cost.
981    pub fn stats(&self) -> StoreStatsSnapshot {
982        self.stats.snapshot()
983    }
984
985    /// Cache and filter counters for each keyspace, for tuning
986    /// `cache_size`, `max_cached_files`, and the filter and pinning policies.
987    ///
988    /// Cumulative since the process started. See [`KeyspaceMetrics`].
989    pub fn metrics(&self) -> Vec<KeyspaceMetrics> {
990        vec![
991            KeyspaceMetrics::read("stream", &self.stream),
992            KeyspaceMetrics::read("idx_topic", &self.idx_topic),
993            KeyspaceMetrics::read("idx_expiry", &self.idx_expiry),
994        ]
995    }
996
997    /// The [`Fsync`] policy this store was opened with.
998    pub fn fsync(&self) -> Fsync {
999        self.fsync.mode
1000    }
1001
1002    /// Fsync the journal now, whatever the [`Fsync`] policy. Every append
1003    /// that has returned is durable once this returns. Use it before a
1004    /// planned shutdown, or in tests that reopen a store.
1005    pub fn flush(&self) -> Result<(), crate::error::Error> {
1006        // Clear before persisting: a write that lands during the persist sets
1007        // the flag again, so the next tick still covers it.
1008        self.fsync.dirty.store(false, Ordering::Release);
1009        self.db.persist(PersistMode::SyncAll)?;
1010        Ok(())
1011    }
1012
1013    /// Apply the [`Fsync`] policy after a committed batch: sync inline under
1014    /// `always`, mark the journal dirty for the tick under `interval`, and do
1015    /// nothing under `never`.
1016    fn after_commit(&self) -> Result<(), crate::error::Error> {
1017        match self.fsync.mode {
1018            Fsync::Always => self.db.persist(PersistMode::SyncAll)?,
1019            Fsync::Interval(_) => self.fsync.dirty.store(true, Ordering::Release),
1020            Fsync::Never => {}
1021        }
1022        Ok(())
1023    }
1024
1025    /// Set a base engine the processor engines clone.
1026    ///
1027    /// A program that embeds xs prepares an [`EngineState`] with its own
1028    /// commands, env, and consts;
1029    /// [`prepared_base`](crate::nu::prepared_base) clones it per spawn and adds
1030    /// the store commands, so the actor, service, and action processors get
1031    /// those commands. Shared across `Store` clones, so set it once, before the
1032    /// processors spawn. No base set: `Engine::new()`.
1033    pub fn with_base_engine(mut self, base: EngineState) -> Self {
1034        self.base_engine = Some(Arc::new(base));
1035        self
1036    }
1037
1038    /// The base engine an embedder attached via [`with_base_engine`], if any.
1039    pub fn base_engine(&self) -> Option<&EngineState> {
1040        self.base_engine.as_deref()
1041    }
1042
1043    /// Wait until the background garbage-collection worker has processed every
1044    /// task queued so far. Useful in tests to observe TTL eviction
1045    /// deterministically.
1046    pub async fn wait_for_gc(&self) {
1047        let (tx, rx) = tokio::sync::oneshot::channel();
1048        let _ = self.gc_tx.send(GCTask::Drain(tx));
1049        let _ = rx.await;
1050    }
1051
1052    /// Queue a sweep now, as the sweep tick would. Follow it with
1053    /// [`wait_for_gc`](Store::wait_for_gc) to observe the removals.
1054    #[cfg(test)]
1055    fn sweep(&self) {
1056        let _ = self.gc_tx.send(GCTask::Sweep);
1057    }
1058
1059    /// Number of entries in the expiry index.
1060    #[cfg(test)]
1061    fn idx_expiry_len(&self) -> usize {
1062        self.idx_expiry.len().unwrap()
1063    }
1064
1065    /// Read frames into an async channel according to `options`.
1066    ///
1067    /// By default this replays matching historical frames oldest-first and then
1068    /// closes the channel. With [`FollowOption::On`] it instead keeps the
1069    /// channel open and streams new appends as they arrive. When following, a
1070    /// single ephemeral `xs.threshold` frame is emitted to mark the boundary
1071    /// between replayed history and live events.
1072    ///
1073    /// The returned [`Receiver`](tokio::sync::mpsc::Receiver) is bounded;
1074    /// dropping it stops the read. For a blocking, non-async caller use
1075    /// [`read_sync`](Store::read_sync).
1076    ///
1077    /// ```no_run
1078    /// use xs::{Store, ReadOptions, FollowOption};
1079    ///
1080    /// # async fn run(store: Store) {
1081    /// let mut rx = store.read(ReadOptions::builder().follow(FollowOption::On).build());
1082    /// while let Some(frame) = rx.recv().await {
1083    ///     if frame.topic == "xs.threshold" {
1084    ///         // caught up to live; everything after this is new
1085    ///         continue;
1086    ///     }
1087    ///     println!("{} {}", frame.id, frame.topic);
1088    /// }
1089    /// # }
1090    /// ```
1091    #[tracing::instrument(skip(self))]
1092    pub fn read(&self, options: ReadOptions) -> tokio::sync::mpsc::Receiver<Frame> {
1093        let (tx, rx) = tokio::sync::mpsc::channel(100);
1094
1095        let should_follow = matches!(
1096            options.follow,
1097            FollowOption::On | FollowOption::WithHeartbeat(_)
1098        );
1099
1100        // Only take broadcast subscription if following. We initate the subscription here to
1101        // ensure we don't miss any messages between historical processing and starting the
1102        // broadcast subscription.
1103        let broadcast_rx = if should_follow {
1104            Some(self.broadcast_tx.subscribe())
1105        } else {
1106            None
1107        };
1108
1109        // Only create done channel if we're doing historical processing
1110        let done_rx = if !options.new {
1111            let (done_tx, done_rx) = tokio::sync::oneshot::channel();
1112            let tx_clone = tx.clone();
1113            let store = self.clone();
1114            let options = options.clone();
1115            let should_follow_clone = should_follow;
1116
1117            // Spawn OS thread to handle historical events
1118            std::thread::spawn(move || {
1119                let mut last_id = None;
1120                let mut count = 0;
1121
1122                // Handle --last N: get the N most recent frames
1123                let filter = TopicFilter::from_option(options.topic.as_deref());
1124
1125                if let Some(last_n) = options.last {
1126                    let iter = store.iter_for_filter_rev(filter);
1127
1128                    // Collect last N frames (in reverse order), skipping expired
1129                    let mut frames: Vec<Frame> = Vec::with_capacity(last_n);
1130                    for frame in iter {
1131                        if is_expired(&frame) {
1132                            continue;
1133                        }
1134                        frames.push(frame);
1135                        if frames.len() >= last_n {
1136                            break;
1137                        }
1138                    }
1139
1140                    // Reverse to chronological order and send
1141                    for frame in frames.into_iter().rev() {
1142                        last_id = Some(frame.id);
1143                        count += 1;
1144                        if tx_clone.blocking_send(frame).is_err() {
1145                            return;
1146                        }
1147                    }
1148                } else {
1149                    // Normal forward iteration
1150                    // Determine start bound: from (inclusive) takes precedence over after (exclusive)
1151                    let start_bound = options
1152                        .from
1153                        .map(|id| (id, true))
1154                        .or_else(|| options.after.map(|id| (id, false)));
1155
1156                    let iter = store.iter_for_filter(filter, start_bound);
1157
1158                    for frame in iter {
1159                        if is_expired(&frame) {
1160                            continue;
1161                        }
1162
1163                        last_id = Some(frame.id);
1164
1165                        if let Some(limit) = options.limit {
1166                            if count >= limit {
1167                                return; // Exit early if limit reached
1168                            }
1169                        }
1170
1171                        if tx_clone.blocking_send(frame).is_err() {
1172                            return;
1173                        }
1174                        count += 1;
1175                    }
1176                }
1177
1178                // Send threshold message if following
1179                if should_follow_clone {
1180                    let threshold = Frame::builder("xs.threshold")
1181                        .id(scru128::new())
1182                        .ttl(TTL::Ephemeral)
1183                        .build();
1184                    if tx_clone.blocking_send(threshold).is_err() {
1185                        return;
1186                    }
1187                }
1188
1189                // Signal completion with the last seen ID and count
1190                let _ = done_tx.send((last_id, count));
1191            });
1192
1193            Some(done_rx)
1194        } else {
1195            None
1196        };
1197
1198        // Handle broadcast subscription and heartbeat
1199        if let Some(broadcast_rx) = broadcast_rx {
1200            let handle = self
1201                .rt
1202                .clone()
1203                .expect("read() with follow needs a runtime handle");
1204
1205            {
1206                let tx = tx.clone();
1207                let limit = options.limit;
1208
1209                handle.spawn(async move {
1210                    // If we have a done_rx, wait for historical processing
1211                    let (last_id, mut count) = match done_rx {
1212                        Some(done_rx) => match done_rx.await {
1213                            Ok((id, count)) => (id, count),
1214                            Err(_) => return, // Historical processing failed/cancelled
1215                        },
1216                        None => (None, 0),
1217                    };
1218
1219                    let filter = TopicFilter::from_option(options.topic.as_deref());
1220
1221                    let mut broadcast_rx = broadcast_rx;
1222                    loop {
1223                        tokio::select! {
1224                            _ = tx.closed() => break,
1225                            r = broadcast_rx.recv() => match r {
1226                                Ok(frame) => {
1227                                    // Filter by topic (any-match against the parsed patterns)
1228                                    if !filter.matches(&frame.topic) {
1229                                        continue;
1230                                    }
1231
1232                                    // Skip if we've already seen this frame during historical scan
1233                                    if let Some(last_scanned_id) = last_id {
1234                                        if frame.id <= last_scanned_id {
1235                                            continue;
1236                                        }
1237                                    }
1238
1239                                    if tx.send(frame).await.is_err() {
1240                                        break;
1241                                    }
1242
1243                                    if let Some(limit) = limit {
1244                                        count += 1;
1245                                        if count >= limit {
1246                                            break;
1247                                        }
1248                                    }
1249                                }
1250                                // The reader fell more than FOLLOW_BUFFER frames
1251                                // behind and the channel overwrote what it had not
1252                                // taken. Carrying on would hand it the frames after
1253                                // the gap as though nothing were missing, and the
1254                                // read channel carries frames only, so there is no
1255                                // way to say otherwise in band. End the read instead:
1256                                // a closed stream is a signal the reader cannot
1257                                // mistake for a complete one. It resumes with
1258                                // `after` set to the last id it received.
1259                                Err(broadcast::error::RecvError::Lagged(skipped)) => {
1260                                    tracing::warn!(
1261                                        skipped,
1262                                        "follow fell behind the live buffer, ending the read"
1263                                    );
1264                                    break;
1265                                }
1266                                Err(broadcast::error::RecvError::Closed) => break,
1267                            }
1268                        }
1269                    }
1270                });
1271            }
1272
1273            // Handle heartbeat if requested
1274            if let FollowOption::WithHeartbeat(duration) = options.follow {
1275                let heartbeat_tx = tx;
1276                handle.spawn(async move {
1277                    loop {
1278                        tokio::select! {
1279                            _ = heartbeat_tx.closed() => break,
1280                            _ = tokio::time::sleep(duration) => {
1281                                let frame = Frame::builder("xs.pulse")
1282                                    .id(scru128::new())
1283                                    .ttl(TTL::Ephemeral)
1284                                    .build();
1285                                if heartbeat_tx.send(frame).await.is_err() {
1286                                    break;
1287                                }
1288                            }
1289                        }
1290                    }
1291                });
1292            }
1293        }
1294
1295        rx
1296    }
1297
1298    /// Replay matching historical frames as a blocking iterator.
1299    ///
1300    /// This honours the `topic`, `from`, `after`, `limit`, and `last` parts of
1301    /// [`ReadOptions`] but ignores [`follow`](ReadOptions::follow): it never
1302    /// streams live appends. Use [`read`](Store::read) when you need to follow.
1303    ///
1304    /// The iterator is lazy: it reads a frame per `next`, so a caller that
1305    /// stops early only pays for what it pulled. The exception is `last`,
1306    /// which reads in reverse and buffers its N frames to return them in
1307    /// chronological order.
1308    ///
1309    /// ```no_run
1310    /// use xs::{Store, ReadOptions};
1311    ///
1312    /// # fn run(store: Store) {
1313    /// let opts = ReadOptions::builder().topic("clip.*".to_string()).last(10).build();
1314    /// for frame in store.read_sync(opts) {
1315    ///     println!("{} {}", frame.id, frame.topic);
1316    /// }
1317    /// # }
1318    /// ```
1319    pub fn read_sync(&self, options: ReadOptions) -> impl Iterator<Item = Frame> + '_ {
1320        // Hide expired frames the sweeper has not reached yet
1321        let filter_expired = |frame: Frame| (!is_expired(&frame)).then_some(frame);
1322
1323        let filter = TopicFilter::from_option(options.topic.as_deref());
1324
1325        if let Some(last_n) = options.last {
1326            // Handle --last N: get the N most recent frames. This branch reads
1327            // in reverse and re-orders, so it has to collect, but it is bounded
1328            // by last_n.
1329            let iter = self.iter_for_filter_rev(filter);
1330
1331            // Collect last N frames (in reverse order), skipping expired
1332            let mut frames: Vec<Frame> = Vec::with_capacity(last_n);
1333            for frame in iter {
1334                if let Some(frame) = filter_expired(frame) {
1335                    frames.push(frame);
1336                    if frames.len() >= last_n {
1337                        break;
1338                    }
1339                }
1340            }
1341
1342            // Reverse to chronological order
1343            frames.reverse();
1344            Box::new(frames.into_iter()) as Box<dyn Iterator<Item = Frame> + '_>
1345        } else {
1346            // Normal forward iteration, streamed: the caller only pays for the
1347            // frames it pulls.
1348            let start_bound = options
1349                .from
1350                .map(|id| (id, true))
1351                .or_else(|| options.after.map(|id| (id, false)));
1352
1353            Box::new(
1354                self.iter_for_filter(filter, start_bound)
1355                    .filter_map(filter_expired)
1356                    .take(options.limit.unwrap_or(usize::MAX)),
1357            )
1358        }
1359    }
1360
1361    /// Returns the current module state as of a given point in the stream.
1362    ///
1363    /// Scans all frames up to (and including) `as_of` and returns a mapping of
1364    /// module name to CAS hash for the latest frame on each `xs.module.<name>`
1365    /// topic.
1366    /// Resolve the set of registered Nushell modules as of a given frame ID.
1367    ///
1368    /// Scans `xs.module.<name>` frames up to and including `as_of` and returns a
1369    /// map from module name to the content hash of its latest definition. Used
1370    /// by the scripting runtime; rarely needed when embedding the store
1371    /// directly.
1372    pub fn nu_modules_at(
1373        &self,
1374        as_of: &Scru128Id,
1375    ) -> std::collections::HashMap<String, ssri::Integrity> {
1376        let mut modules = std::collections::HashMap::new();
1377        let options = ReadOptions::builder().follow(FollowOption::Off).build();
1378        for frame in self.read_sync(options) {
1379            if frame.id > *as_of {
1380                break;
1381            }
1382            if let Some(hash) = frame.hash {
1383                if let Some(name) = frame.topic.strip_prefix("xs.module.") {
1384                    if !name.is_empty() {
1385                        modules.insert(name.to_string(), hash);
1386                    }
1387                }
1388            }
1389        }
1390        modules
1391    }
1392
1393    /// Fetch a single frame by ID, or `None` if no such frame exists.
1394    pub fn get(&self, id: &Scru128Id) -> Option<Frame> {
1395        self.stats.stream_reads.fetch_add(1, Ordering::Relaxed);
1396        self.stream
1397            .get(id.to_bytes())
1398            .unwrap()
1399            .map(|value| deserialize_frame((id.as_bytes(), value)))
1400    }
1401
1402    /// Delete a frame and its topic index entries. Removing a frame that does
1403    /// not exist is a no-op and returns `Ok(())`.
1404    ///
1405    /// This removes the stream entry only; any payload bytes in the
1406    /// content-addressed store are left in place.
1407    #[tracing::instrument(skip(self), fields(id = %id.to_string()))]
1408    pub fn remove(&self, id: &Scru128Id) -> Result<(), crate::error::Error> {
1409        self.remove_many([Removal::Id(*id)])
1410    }
1411
1412    /// Delete every frame in `removals`, with their index entries, in one
1413    /// batch with one commit. A [`Removal::Id`] whose frame is already gone
1414    /// is skipped, and nothing is committed when none remain.
1415    fn remove_many(
1416        &self,
1417        removals: impl IntoIterator<Item = Removal>,
1418    ) -> Result<(), crate::error::Error> {
1419        let mut batch = self.db.batch();
1420        for removal in removals {
1421            match removal {
1422                Removal::Id(id) => {
1423                    if let Some(frame) = self.get(&id) {
1424                        self.remove_frame_keys(
1425                            &mut batch,
1426                            &frame.id,
1427                            &frame.topic,
1428                            expires_at(&frame),
1429                        );
1430                    }
1431                }
1432                Removal::Indexed {
1433                    id,
1434                    topic,
1435                    expires_at,
1436                } => self.remove_frame_keys(&mut batch, &id, &topic, expires_at),
1437            }
1438        }
1439        if batch.is_empty() {
1440            return Ok(());
1441        }
1442        batch.commit()?;
1443        self.stats.remove_commits.fetch_add(1, Ordering::Relaxed);
1444        // No inline fsync in any mode. The tombstones reach disk with the next
1445        // append's sync or the next tick. If power loss brings a frame back,
1446        // its index entries come back with it, so it is trimmed or swept
1447        // again.
1448        if matches!(self.fsync.mode, Fsync::Interval(_)) {
1449            self.fsync.dirty.store(true, Ordering::Release);
1450        }
1451        Ok(())
1452    }
1453
1454    /// Queue the stream entry and index keys of a frame for removal on
1455    /// `batch`. `expires_at` is the frame's expiry index slot, `Some` only for
1456    /// a `time:` frame. The topic key is built without validation: the frame
1457    /// exists.
1458    fn remove_frame_keys(
1459        &self,
1460        batch: &mut OwnedWriteBatch,
1461        id: &Scru128Id,
1462        topic: &str,
1463        expires_at: Option<u64>,
1464    ) {
1465        self.stats.removed_frames.fetch_add(1, Ordering::Relaxed);
1466        let mut topic_key = idx_topic_key_prefix(topic);
1467        topic_key.extend(id.as_bytes());
1468        batch.remove(&self.stream, id.as_bytes());
1469        batch.remove(&self.idx_topic, topic_key);
1470        for prefix_key in idx_topic_prefix_keys(topic, id) {
1471            batch.remove(&self.idx_topic, prefix_key);
1472        }
1473        if let Some(expires_at) = expires_at {
1474            batch.remove(&self.idx_expiry, idx_expiry_key(expires_at, id));
1475        }
1476    }
1477
1478    /// The oldest `time:` frames whose expiry is at or before `now_ms`, at
1479    /// most `limit` of them, straight from the expiry index. A full `limit`
1480    /// means there may be more.
1481    ///
1482    /// The scan resumes at [`sweep_start`](Store::sweep_start) rather than at
1483    /// the start of the index. Everything below that point is the tombstones
1484    /// of frames earlier sweeps removed, and walking them again on every sweep
1485    /// is what made draining a backlog cost O(n^2/batch).
1486    fn expired_at(&self, now_ms: u64, limit: usize) -> Vec<Removal> {
1487        let last = idx_expiry_key(now_ms, &Scru128Id::from_bytes([0xff; 16]));
1488        let from = *self.sweep_start.lock().unwrap();
1489        // A clock that stepped backwards can leave the resume point above the
1490        // cut-off. Nothing below it is unswept, so there is nothing to do
1491        // until the clock catches up.
1492        if from.is_some_and(|start| start.key > last) {
1493            return Vec::new();
1494        }
1495        let start = from.map_or(Bound::Unbounded, |start| start.bound());
1496
1497        let mut removals = Vec::new();
1498        let mut swept = None;
1499        for guard in self
1500            .idx_expiry
1501            .range((start, Bound::Included(last)))
1502            .take(limit)
1503        {
1504            // Stop on a read error rather than skipping the entry: the resume
1505            // point may never pass an entry this sweep did not remove.
1506            let Ok((key, value)) = guard.into_inner() else {
1507                break;
1508            };
1509            let Ok(key) = <[u8; 24]>::try_from(key.as_ref()) else {
1510                break;
1511            };
1512            let (expires_at, id) = idx_expiry_parse_key(&key);
1513            removals.push(Removal::Indexed {
1514                id,
1515                topic: String::from_utf8_lossy(&value).into_owned(),
1516                expires_at: Some(expires_at),
1517            });
1518            swept = Some(key);
1519        }
1520
1521        if let Some(key) = swept {
1522            let mut start = self.sweep_start.lock().unwrap();
1523            // Only advance from the point this scan started at. An index write
1524            // that moved it back while the scan ran wins: the next sweep
1525            // covers this ground again, and covering it twice is cheap next to
1526            // leaving a frame behind.
1527            if *start == from {
1528                *start = Some(SweepStart {
1529                    key,
1530                    inclusive: false,
1531                });
1532            }
1533        }
1534        removals
1535    }
1536
1537    /// Ids on `topic` past the newest `keep`, oldest first, at most `limit` of
1538    /// them, ignoring frames already in `pending` so a batch drops exactly
1539    /// what applying its tasks one by one would. A full `limit` means there
1540    /// may be more.
1541    ///
1542    /// Two bounded scans, both floored at the topic's
1543    /// [`trim_floor`](Store::trim_floor), below which the topic holds nothing
1544    /// but the tombstones of frames an earlier trim removed:
1545    ///
1546    /// 1. backwards from the newest entry for the `keep`-th newest live id,
1547    ///    the high-water mark everything older than which is overflow;
1548    /// 2. forwards from the floor for the ids to remove.
1549    ///
1550    /// Taking the removals in forward order is what keeps trimming a backlog
1551    /// linear. A truncated pass leaves the floor on the newest id it removed,
1552    /// so the next pass starts above its own tombstones instead of walking
1553    /// them to reach what it left.
1554    ///
1555    /// Both scans read index keys only. The expiry an `idx_topic` value now
1556    /// carries is for the read path: a trim counts entries, so a `time:` frame
1557    /// the sweeper has not reached yet still holds one of the `keep` slots,
1558    /// exactly as it did when those values were empty.
1559    fn last_overflow(
1560        &self,
1561        topic: &str,
1562        keep: u32,
1563        pending: &HashSet<Scru128Id>,
1564        limit: usize,
1565    ) -> Vec<Overflow> {
1566        let floor = self.trim_floor.lock().unwrap().get(topic).copied();
1567        // The floor is a bound on the id, so it is a bound on the key: both
1568        // scans open at it and never reach the tombstones below it. See
1569        // [`idx_topic_range`].
1570        let (low, top) = idx_topic_range(
1571            idx_topic_key_prefix(topic),
1572            floor.map(|floor| (floor.id, floor.inclusive)),
1573        );
1574
1575        // The keep-th newest live id. Everything below it is overflow; fewer
1576        // than keep live entries means there is nothing to trim.
1577        let high = if keep == 0 {
1578            top
1579        } else {
1580            let mut seen = 0;
1581            let mut hwm = None;
1582            let mut oldest = None;
1583            for guard in self.idx_topic.range((low.clone(), top)).rev() {
1584                // A read error is not the end of the topic: give up on this
1585                // pass rather than record a floor the scan never reached.
1586                let Ok(key) = guard.key() else {
1587                    return Vec::new();
1588                };
1589                let id = idx_topic_frame_id_from_key(&key);
1590                oldest = Some(id);
1591                if pending.contains(&id) {
1592                    continue;
1593                }
1594                seen += 1;
1595                if seen == keep {
1596                    hwm = Some(key.to_vec());
1597                    break;
1598                }
1599            }
1600            match hwm {
1601                Some(key) => Bound::Excluded(key),
1602                None => {
1603                    // The scan reached the end of the topic without finding
1604                    // `keep` live frames. Nothing to trim, but everything
1605                    // below the oldest entry it saw is tombstone, so the next
1606                    // scan can start there instead of walking it again.
1607                    if let Some(id) = oldest {
1608                        self.set_trim_floor(
1609                            topic,
1610                            floor,
1611                            TrimFloor {
1612                                id,
1613                                inclusive: true,
1614                            },
1615                        );
1616                    }
1617                    return Vec::new();
1618                }
1619            }
1620        };
1621
1622        let mut overflow = Vec::new();
1623        let mut trimmed = None;
1624        for guard in self.idx_topic.range((low, high)) {
1625            // Take the value too: it carries the frame's expiry, so the removal
1626            // needs no point read into `stream` to learn its keys.
1627            let Ok((key, value)) = guard.into_inner() else {
1628                break;
1629            };
1630            let id = idx_topic_frame_id_from_key(&key);
1631            if !pending.contains(&id) {
1632                overflow.push(Overflow {
1633                    id,
1634                    expires_at: idx_topic_expiry_of(&value),
1635                });
1636            }
1637            // A pending id counts: this drain removes it too, so the floor may
1638            // pass it.
1639            trimmed = Some(id);
1640            if overflow.len() == limit {
1641                break;
1642            }
1643        }
1644
1645        if let Some(id) = trimmed {
1646            self.set_trim_floor(
1647                topic,
1648                floor,
1649                TrimFloor {
1650                    id,
1651                    inclusive: false,
1652                },
1653            );
1654        }
1655        overflow
1656    }
1657
1658    /// Move a topic's trim floor to `to`, unless something moved it since the
1659    /// scan read `from`. An index write that dropped it while the scan ran
1660    /// wins: the next trim covers the topic from further back, and covering
1661    /// ground twice is cheap next to leaving a frame behind.
1662    fn set_trim_floor(&self, topic: &str, from: Option<TrimFloor>, to: TrimFloor) {
1663        let mut floors = self.trim_floor.lock().unwrap();
1664        if floors.get(topic).copied() == from {
1665            floors.insert(topic.to_string(), to);
1666        }
1667    }
1668
1669    /// Note that a `time:` frame's expiry entry has been committed at `key`.
1670    ///
1671    /// A frame restored with an id from the past, or appended in the instant a
1672    /// sweep passed its slot, lands below the resume point, where no later
1673    /// sweep would look. Move the resume point back onto it.
1674    ///
1675    /// Must be called after the commit: called before, a sweep could run,
1676    /// find nothing at the key yet, and leave the resume point above it again.
1677    fn note_indexed_expiry(&self, key: [u8; 24]) {
1678        let mut start = self.sweep_start.lock().unwrap();
1679        // Already at the start of the index: nothing can be below it.
1680        let Some(current) = *start else { return };
1681        let candidate = SweepStart {
1682            key,
1683            inclusive: true,
1684        };
1685        if candidate.rank() < current.rank() {
1686            *start = Some(candidate);
1687        }
1688    }
1689
1690    /// Note that a frame's topic index entry has been committed. An id at or
1691    /// below the topic's trim floor is in ground a trim scan no longer covers,
1692    /// so drop the floor and let the next trim scan the topic whole.
1693    ///
1694    /// Called after the commit, as
1695    /// [`note_indexed_expiry`](Store::note_indexed_expiry) is, and for every
1696    /// frame: `last:n` counts what is on the topic, whatever ttl put it there.
1697    fn note_indexed_topic(&self, topic: &str, id: &Scru128Id) {
1698        let mut floors = self.trim_floor.lock().unwrap();
1699        if floors.is_empty() {
1700            return;
1701        }
1702        if floors.get(topic).is_some_and(|floor| *id <= floor.id) {
1703            floors.remove(topic);
1704        }
1705    }
1706
1707    /// Forget where the sweep and the `last:n` trims had got to, so each next
1708    /// scan starts from the beginning of its index again. Called when a
1709    /// removal batch fails to commit: its frames are still there, but the
1710    /// resume points have already moved as though they were gone.
1711    fn reset_gc_resume(&self) {
1712        *self.sweep_start.lock().unwrap() = None;
1713        self.trim_floor.lock().unwrap().clear();
1714    }
1715
1716    // --- Content-addressed store (CAS) ---
1717    //
1718    // Frame payloads live here, keyed by an integrity hash. The typical flow is
1719    // `cas_insert` to store bytes, stash the returned hash on a `Frame`, then
1720    // `cas_read` to retrieve them later. Each method has a `_sync` twin for
1721    // blocking callers; the streaming `cas_reader`/`cas_writer` variants avoid
1722    // buffering the whole payload in memory.
1723
1724    /// Open a streaming reader for the payload identified by `hash`.
1725    pub async fn cas_reader(&self, hash: ssri::Integrity) -> cacache::Result<cacache::Reader> {
1726        cacache::Reader::open_hash(&self.path.join("cacache"), hash).await
1727    }
1728
1729    /// Blocking variant of [`cas_reader`](Store::cas_reader).
1730    pub fn cas_reader_sync(&self, hash: ssri::Integrity) -> cacache::Result<cacache::SyncReader> {
1731        cacache::SyncReader::open_hash(self.path.join("cacache"), hash)
1732    }
1733
1734    /// Open a streaming writer; finish it to obtain the payload's integrity hash.
1735    pub async fn cas_writer(&self) -> cacache::Result<cacache::Writer> {
1736        cacache::WriteOpts::new()
1737            .open_hash(&self.path.join("cacache"))
1738            .await
1739    }
1740
1741    /// Blocking variant of [`cas_writer`](Store::cas_writer).
1742    pub fn cas_writer_sync(&self) -> cacache::Result<cacache::SyncWriter> {
1743        cacache::WriteOpts::new().open_hash_sync(self.path.join("cacache"))
1744    }
1745
1746    /// Store `content` and return its integrity hash, ready to attach to a
1747    /// [`Frame::hash`].
1748    pub async fn cas_insert(&self, content: impl AsRef<[u8]>) -> cacache::Result<ssri::Integrity> {
1749        cacache::write_hash(&self.path.join("cacache"), content).await
1750    }
1751
1752    /// Blocking variant of [`cas_insert`](Store::cas_insert).
1753    pub fn cas_insert_sync(&self, content: impl AsRef<[u8]>) -> cacache::Result<ssri::Integrity> {
1754        cacache::write_hash_sync(self.path.join("cacache"), content)
1755    }
1756
1757    /// Convenience wrapper over [`cas_insert`](Store::cas_insert) for a byte slice.
1758    pub async fn cas_insert_bytes(&self, bytes: &[u8]) -> cacache::Result<ssri::Integrity> {
1759        self.cas_insert(bytes).await
1760    }
1761
1762    /// Blocking variant of [`cas_insert_bytes`](Store::cas_insert_bytes).
1763    pub fn cas_insert_bytes_sync(&self, bytes: &[u8]) -> cacache::Result<ssri::Integrity> {
1764        self.cas_insert_sync(bytes)
1765    }
1766
1767    /// Read back the full payload for `hash` into a `Vec<u8>`.
1768    pub async fn cas_read(&self, hash: &ssri::Integrity) -> cacache::Result<Vec<u8>> {
1769        cacache::read_hash(&self.path.join("cacache"), hash).await
1770    }
1771
1772    /// Blocking variant of [`cas_read`](Store::cas_read).
1773    pub fn cas_read_sync(&self, hash: &ssri::Integrity) -> cacache::Result<Vec<u8>> {
1774        cacache::read_hash_sync(self.path.join("cacache"), hash)
1775    }
1776
1777    /// Persist a frame exactly as given, including its existing
1778    /// [`id`](Frame::id), without broadcasting it to live readers or scheduling
1779    /// TTL garbage collection.
1780    ///
1781    /// Most callers want [`append`](Store::append) instead, which assigns a
1782    /// fresh ID, handles ephemeral and `Last` retention, and notifies
1783    /// subscribers. Use `insert_frame` only when you are reconstructing a stream
1784    /// with predetermined IDs (for example when restoring a backup).
1785    #[tracing::instrument(skip(self))]
1786    pub fn insert_frame(&self, frame: &Frame) -> Result<(), crate::error::Error> {
1787        let encoded: Vec<u8> = serde_json::to_vec(&frame).unwrap();
1788
1789        // Get the index topic key (also validates topic)
1790        let topic_key = idx_topic_key_from_frame(frame)?;
1791
1792        // Get prefix index keys for hierarchical queries
1793        let prefix_keys = idx_topic_prefix_keys(&frame.topic, &frame.id);
1794
1795        // Every idx_topic entry for this frame carries the same expiry, so a
1796        // topic scan can drop an expired id without reading the frame.
1797        let idx_value = idx_topic_expiry_value(frame);
1798
1799        // The id may already be here under a different topic or ttl. This is
1800        // an upsert, so the stream row is replaced, but the index keys carry
1801        // the topic and the expiry, so the old ones would survive their frame
1802        // and answer for a topic it no longer has.
1803        let superseded = self.get(&frame.id);
1804
1805        let mut batch = self.db.batch();
1806        batch.insert(&self.stream, frame.id.as_bytes(), encoded);
1807        batch.insert(&self.idx_topic, topic_key.clone(), idx_value);
1808        for prefix_key in &prefix_keys {
1809            batch.insert(&self.idx_topic, prefix_key, idx_value);
1810        }
1811        // A time: frame is also indexed by its expiry, for the sweeper
1812        let expiry_key = expires_at(frame).map(|at| idx_expiry_key(at, &frame.id));
1813        if let Some(key) = expiry_key {
1814            batch.insert(&self.idx_expiry, key, frame.topic.as_bytes());
1815        }
1816
1817        if let Some(old) = superseded.as_ref().filter(|old| old.topic != frame.topic) {
1818            let mut old_key = idx_topic_key_prefix(&old.topic);
1819            old_key.extend(frame.id.as_bytes());
1820            // "a.b" and "a.c" share the prefix key "a.", so drop only the keys
1821            // the write above is not putting straight back: removing and
1822            // inserting one key in one batch has no defined winner.
1823            for key in std::iter::once(old_key).chain(idx_topic_prefix_keys(&old.topic, &frame.id))
1824            {
1825                if key != topic_key && !prefix_keys.contains(&key) {
1826                    batch.remove(&self.idx_topic, key);
1827                }
1828            }
1829        }
1830        // A changed ttl moves the expiry entry, because the expiry is half of
1831        // its key. An unchanged one is the same key, already rewritten above
1832        // with whatever topic the frame now carries.
1833        if let Some(stale) = superseded
1834            .as_ref()
1835            .and_then(expires_at)
1836            .map(|at| idx_expiry_key(at, &frame.id))
1837            .filter(|stale| Some(*stale) != expiry_key)
1838        {
1839            batch.remove(&self.idx_expiry, stale);
1840        }
1841
1842        batch.commit()?;
1843        // Both gc scans resume where they last stopped, and an id from the
1844        // past can land behind them. Tell them, once the entries are there to
1845        // be found.
1846        if let Some(key) = expiry_key {
1847            self.note_indexed_expiry(key);
1848        }
1849        self.note_indexed_topic(&frame.topic, &frame.id);
1850        self.after_commit()
1851    }
1852
1853    /// Append a frame to the stream and return it with its freshly assigned
1854    /// [`id`](Frame::id).
1855    ///
1856    /// This is the primary write path. It:
1857    ///
1858    /// - assigns a new time-sortable ID (overwriting any ID on the input);
1859    /// - validates the topic (see [`validate_topic`]);
1860    /// - persists the frame, unless its [`TTL`] is [`TTL::Ephemeral`], in which
1861    ///   case it is only broadcast to live readers;
1862    /// - schedules garbage collection for [`TTL::Last`] retention;
1863    /// - broadcasts the frame to everyone currently in a following
1864    ///   [`read`](Store::read).
1865    ///
1866    /// Appends are serialized internally, so frames are assigned IDs and
1867    /// delivered to subscribers in a consistent order.
1868    ///
1869    /// # Errors
1870    ///
1871    /// Returns an error if the topic is invalid or the underlying write fails.
1872    ///
1873    /// ```no_run
1874    /// use xs::{Store, Frame, TTL};
1875    ///
1876    /// # async fn run(store: Store) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
1877    /// let hash = store.cas_insert("hello clipboard").await?;
1878    /// let frame = store.append(
1879    ///     Frame::builder("clip.add").hash(hash).ttl(TTL::Last(100)).build(),
1880    /// )?;
1881    /// println!("appended {}", frame.id);
1882    /// # Ok(())
1883    /// # }
1884    /// ```
1885    pub fn append(&self, mut frame: Frame) -> Result<Frame, crate::error::Error> {
1886        // Serialize all appends to ensure ID generation, write, and broadcast
1887        // happen atomically. This guarantees subscribers receive frames in
1888        // scru128 ID order.
1889        let _guard = self.append_lock.lock().unwrap();
1890
1891        frame.id = scru128::new();
1892
1893        // Check for null byte in topic (in case we're not storing the frame)
1894        idx_topic_key_from_frame(&frame)?;
1895
1896        // only store the frame if it's not ephemeral
1897        if frame.ttl != Some(TTL::Ephemeral) {
1898            self.insert_frame(&frame)?;
1899
1900            // If this is a Last TTL, schedule a gc task
1901            if let Some(TTL::Last(n)) = frame.ttl {
1902                let _ = self.gc_tx.send(GCTask::CheckLastTTL {
1903                    topic: frame.topic.clone(),
1904                    keep: n,
1905                });
1906            }
1907        }
1908
1909        self.stats.appends.fetch_add(1, Ordering::Relaxed);
1910        let _ = self.broadcast_tx.send(frame.clone());
1911        Ok(frame)
1912    }
1913
1914    /// Iterate frames starting from a bound.
1915    /// `start` is `(id, inclusive)` where inclusive=true means >= and inclusive=false means >.
1916    fn iter_frames(
1917        &self,
1918        start: Option<(Scru128Id, bool)>,
1919    ) -> Box<dyn Iterator<Item = Frame> + '_> {
1920        let range = match start {
1921            Some((id, true)) => (Bound::Included(id.as_bytes().to_vec()), Bound::Unbounded),
1922            Some((id, false)) => (Bound::Excluded(id.as_bytes().to_vec()), Bound::Unbounded),
1923            None => (Bound::Unbounded, Bound::Unbounded),
1924        };
1925
1926        let scanned = self.stats.clone();
1927        Box::new(self.stream.range(range).filter_map(move |guard| {
1928            scanned.frames_scanned.fetch_add(1, Ordering::Relaxed);
1929            let (key, value) = guard.into_inner().ok()?;
1930            Some(deserialize_frame((key, value)))
1931        }))
1932    }
1933
1934    /// Iterate frames in reverse order (most recent first).
1935    fn iter_frames_rev(&self) -> Box<dyn Iterator<Item = Frame> + '_> {
1936        let scanned = self.stats.clone();
1937        Box::new(self.stream.iter().rev().filter_map(move |guard| {
1938            scanned.frames_scanned.fetch_add(1, Ordering::Relaxed);
1939            let (key, value) = guard.into_inner().ok()?;
1940            Some(deserialize_frame((key, value)))
1941        }))
1942    }
1943
1944    /// Iterate frames by topic in reverse order (most recent first).
1945    fn iter_frames_by_topic_rev(&self, topic: &str) -> Box<dyn Iterator<Item = Frame> + '_> {
1946        let prefix = idx_topic_key_prefix(topic);
1947        let now = now_ms();
1948        let scanned = self.stats.clone();
1949        Box::new(
1950            self.idx_topic
1951                .prefix(prefix)
1952                .rev()
1953                .filter_map(move |guard| {
1954                    scanned.frames_scanned.fetch_add(1, Ordering::Relaxed);
1955                    self.get(&idx_topic_hit_id(guard, now)?)
1956                }),
1957        )
1958    }
1959
1960    /// Iterate frames by topic prefix in reverse order (most recent first).
1961    fn iter_frames_by_topic_prefix_rev(
1962        &self,
1963        prefix: &str,
1964    ) -> Box<dyn Iterator<Item = Frame> + '_> {
1965        let mut index_prefix = Vec::with_capacity(prefix.len() + 1);
1966        index_prefix.extend(prefix.as_bytes());
1967        index_prefix.push(NULL_DELIMITER);
1968
1969        let now = now_ms();
1970        let scanned = self.stats.clone();
1971
1972        Box::new(
1973            self.idx_topic
1974                .prefix(index_prefix)
1975                .rev()
1976                .filter_map(move |guard| {
1977                    scanned.frames_scanned.fetch_add(1, Ordering::Relaxed);
1978                    self.get(&idx_topic_hit_id(guard, now)?)
1979                }),
1980        )
1981    }
1982
1983    /// Iterate frames on an exact topic, ascending by frame id, from `start`.
1984    fn iter_frames_by_topic(
1985        &self,
1986        topic: &str,
1987        start: Option<(Scru128Id, bool)>,
1988    ) -> Box<dyn Iterator<Item = Frame> + '_> {
1989        let range = idx_topic_range(idx_topic_key_prefix(topic), start);
1990        let now = now_ms();
1991        let scanned = self.stats.clone();
1992        Box::new(self.idx_topic.range(range).filter_map(move |guard| {
1993            scanned.frames_scanned.fetch_add(1, Ordering::Relaxed);
1994            self.get(&idx_topic_hit_id(guard, now)?)
1995        }))
1996    }
1997
1998    /// Iterate frames matching a topic prefix (for wildcard queries like "user.*").
1999    /// The prefix should include the trailing dot (e.g., "user." for "user.*").
2000    fn iter_frames_by_topic_prefix(
2001        &self,
2002        prefix: &str,
2003        start: Option<(Scru128Id, bool)>,
2004    ) -> Box<dyn Iterator<Item = Frame> + '_> {
2005        // Build index prefix: "user.\0" for scanning all "user.*" entries
2006        let mut index_prefix = Vec::with_capacity(prefix.len() + 1);
2007        index_prefix.extend(prefix.as_bytes());
2008        index_prefix.push(NULL_DELIMITER);
2009
2010        let range = idx_topic_range(index_prefix, start);
2011        let now = now_ms();
2012        let scanned = self.stats.clone();
2013
2014        Box::new(self.idx_topic.range(range).filter_map(move |guard| {
2015            scanned.frames_scanned.fetch_add(1, Ordering::Relaxed);
2016            self.get(&idx_topic_hit_id(guard, now)?)
2017        }))
2018    }
2019
2020    /// Forward iterator for a single pattern, using the topic index.
2021    fn iter_for_pattern(
2022        &self,
2023        pattern: Pattern,
2024        start: Option<(Scru128Id, bool)>,
2025    ) -> Box<dyn Iterator<Item = Frame> + '_> {
2026        match pattern {
2027            Pattern::Exact(topic) => self.iter_frames_by_topic(&topic, start),
2028            Pattern::Prefix(prefix) => self.iter_frames_by_topic_prefix(&prefix, start),
2029        }
2030    }
2031
2032    /// Reverse (most recent first) iterator for a single pattern.
2033    fn iter_for_pattern_rev(&self, pattern: Pattern) -> Box<dyn Iterator<Item = Frame> + '_> {
2034        match pattern {
2035            Pattern::Exact(topic) => self.iter_frames_by_topic_rev(&topic),
2036            Pattern::Prefix(prefix) => self.iter_frames_by_topic_prefix_rev(&prefix),
2037        }
2038    }
2039
2040    /// Forward iterator for a parsed topic filter, ascending by frame id.
2041    /// A single pattern uses the indexed path directly; multiple patterns
2042    /// are k-way merged with dedupe for overlapping patterns.
2043    fn iter_for_filter(
2044        &self,
2045        filter: TopicFilter,
2046        start: Option<(Scru128Id, bool)>,
2047    ) -> Box<dyn Iterator<Item = Frame> + '_> {
2048        match filter {
2049            TopicFilter::All => self.iter_frames(start),
2050            TopicFilter::Patterns(mut patterns) if patterns.len() == 1 => {
2051                self.iter_for_pattern(patterns.remove(0), start)
2052            }
2053            TopicFilter::Patterns(patterns) => {
2054                let iters = patterns
2055                    .into_iter()
2056                    .map(|p| self.iter_for_pattern(p, start))
2057                    .collect();
2058                Box::new(MergeById::new(iters, false))
2059            }
2060        }
2061    }
2062
2063    /// Reverse variant of [`iter_for_filter`](Store::iter_for_filter),
2064    /// descending by frame id (most recent first).
2065    fn iter_for_filter_rev(&self, filter: TopicFilter) -> Box<dyn Iterator<Item = Frame> + '_> {
2066        match filter {
2067            TopicFilter::All => self.iter_frames_rev(),
2068            TopicFilter::Patterns(mut patterns) if patterns.len() == 1 => {
2069                self.iter_for_pattern_rev(patterns.remove(0))
2070            }
2071            TopicFilter::Patterns(patterns) => {
2072                let iters = patterns
2073                    .into_iter()
2074                    .map(|p| self.iter_for_pattern_rev(p))
2075                    .collect();
2076                Box::new(MergeById::new(iters, true))
2077            }
2078        }
2079    }
2080}
2081
2082fn spawn_gc_worker(
2083    mut gc_rx: UnboundedReceiver<GCTask>,
2084    mut store: Store,
2085) -> std::thread::JoinHandle<()> {
2086    // The worker's own clone must not hold a live `gc_tx`, or the channel
2087    // never closes and the worker, the database, and the fsync worker outlive
2088    // every caller's handle. Give it a sender whose receiver is already gone.
2089    let (closed_tx, _) = mpsc::unbounded_channel();
2090    store.gc_tx = closed_tx;
2091
2092    // One drain per wake: take the task that woke us and whatever else is
2093    // queued, up to the cap, and apply every removal in one batch with one
2094    // commit. A Drain is answered after that commit, so every task queued
2095    // before it has been applied by the time its reply lands.
2096    const MAX_TASKS_PER_DRAIN: usize = 1024;
2097    // A sweep removes at most this many frames. When it fills the cap the
2098    // next drain starts with another sweep, without waiting for the tick.
2099    const MAX_SWEEP_PER_DRAIN: usize = 4096;
2100    // A last:n trim removes at most this many frames per topic per drain, and
2101    // is carried over the same way a full sweep is. Without a cap the first
2102    // trim of a topic with a backlog collects every id past `keep` at once.
2103    const MAX_TRIM_PER_DRAIN: usize = 4096;
2104
2105    std::thread::spawn(move || {
2106        // Tasks a full batch left unfinished, run first in the next drain.
2107        let mut carried: Vec<GCTask> = Vec::new();
2108        loop {
2109            if carried.is_empty() {
2110                match gc_rx.blocking_recv() {
2111                    Some(task) => carried.push(task),
2112                    None => break,
2113                }
2114            }
2115
2116            let drain_started = std::time::Instant::now();
2117
2118            // Ids in this drain, so a last:n scan and a sweep never count or
2119            // queue a frame twice.
2120            let mut pending = HashSet::new();
2121            let mut removals = Vec::new();
2122            let mut drains = Vec::new();
2123            let mut sweep_again = false;
2124            // Topics whose trim filled its cap, one entry each however many
2125            // tasks for them this drain took.
2126            let mut trim_again: HashMap<String, u32> = HashMap::new();
2127            // Topics this drain must trim, collapsed from however many tasks
2128            // asked. Scanned after the take loop so every Sweep has already
2129            // run: an expired frame is hidden from reads, so it must not hold
2130            // a keep slot.
2131            let mut trims: HashMap<String, u32> = HashMap::new();
2132            let mut taken = 0;
2133            while taken < MAX_TASKS_PER_DRAIN {
2134                let Some(task) = carried.pop().or_else(|| gc_rx.try_recv().ok()) else {
2135                    break;
2136                };
2137                match task {
2138                    GCTask::CheckLastTTL { topic, keep } => {
2139                        // One scan per topic per drain, at the smallest keep
2140                        // asked for. A second scan for the same topic filters
2141                        // out everything the first collected, skips keep, and
2142                        // walks the rest to return nothing.
2143                        trims
2144                            .entry(topic)
2145                            .and_modify(|k| *k = (*k).min(keep))
2146                            .or_insert(keep);
2147                    }
2148                    GCTask::Sweep => {
2149                        let expired = store.expired_at(now_ms(), MAX_SWEEP_PER_DRAIN);
2150                        sweep_again |= expired.len() == MAX_SWEEP_PER_DRAIN;
2151                        for removal in expired {
2152                            if pending.insert(removal.id()) {
2153                                removals.push(removal);
2154                            }
2155                        }
2156                    }
2157                    GCTask::Drain(tx) => drains.push(tx),
2158                }
2159                taken += 1;
2160            }
2161
2162            // Now that every Sweep in this drain has run, so `pending` holds
2163            // the expired frames it is removing, trim each topic once.
2164            for (topic, keep) in trims {
2165                let overflow = store.last_overflow(&topic, keep, &pending, MAX_TRIM_PER_DRAIN);
2166                let again = overflow.len() == MAX_TRIM_PER_DRAIN;
2167                for frame in overflow {
2168                    if pending.insert(frame.id) {
2169                        removals.push(Removal::Indexed {
2170                            id: frame.id,
2171                            topic: topic.clone(),
2172                            expires_at: frame.expires_at,
2173                        });
2174                    }
2175                }
2176                if again {
2177                    trim_again.insert(topic, keep);
2178                }
2179            }
2180
2181            if let Err(e) = store.remove_many(removals) {
2182                tracing::error!("gc remove failed: {e}");
2183                // The batch did not land, so every frame this drain counted as
2184                // gone is still there, below resume points that have already
2185                // moved past it. Start both scans over.
2186                store.reset_gc_resume();
2187            }
2188            store.stats.gc_drains.fetch_add(1, Ordering::Relaxed);
2189            store
2190                .stats
2191                .gc_nanos
2192                .fetch_add(drain_started.elapsed().as_nanos() as u64, Ordering::Relaxed);
2193            for tx in drains {
2194                let _ = tx.send(());
2195            }
2196            carried.extend(
2197                trim_again
2198                    .into_iter()
2199                    .map(|(topic, keep)| GCTask::CheckLastTTL { topic, keep }),
2200            );
2201            if sweep_again {
2202                carried.push(GCTask::Sweep);
2203            }
2204        }
2205    })
2206}
2207
2208/// Fsync the journal every `interval` when something was written since the
2209/// last tick. No final persist is needed on stop: fjall syncs the journal
2210/// when the database closes.
2211fn spawn_fsync_tick(db: Database, state: Arc<FsyncState>, interval: Duration) -> Tick {
2212    Tick::spawn(interval, move || {
2213        if state.dirty.swap(false, Ordering::AcqRel) {
2214            if let Err(e) = db.persist(PersistMode::SyncAll) {
2215                tracing::error!("fsync failed: {e}");
2216            }
2217        }
2218    })
2219}
2220
2221/// Queue a [`GCTask::Sweep`] every `interval`. The gc worker does the scan
2222/// and the removals, so a slow sweep never blocks the tick.
2223fn spawn_sweep_tick(gc_tx: UnboundedSender<GCTask>, interval: Duration) -> Tick {
2224    Tick::spawn(interval, move || {
2225        let _ = gc_tx.send(GCTask::Sweep);
2226    })
2227}
2228
2229fn now_ms() -> u64 {
2230    std::time::SystemTime::now()
2231        .duration_since(std::time::UNIX_EPOCH)
2232        .unwrap()
2233        .as_millis() as u64
2234}
2235
2236/// When a `time:` frame expires, in ms since the epoch: its id's timestamp
2237/// plus its ttl. `None` for any other ttl.
2238fn expires_at(frame: &Frame) -> Option<u64> {
2239    match frame.ttl.as_ref() {
2240        Some(TTL::Time(ttl)) => Some(frame.id.timestamp().saturating_add(ttl.as_millis() as u64)),
2241        _ => None,
2242    }
2243}
2244
2245/// Has this frame's `time:` ttl passed? Matches the sweeper's cut-off, so a
2246/// frame is hidden from reads from the moment it becomes sweepable.
2247fn is_expired(frame: &Frame) -> bool {
2248    matches!(expires_at(frame), Some(at) if now_ms() >= at)
2249}
2250
2251/// Expiry index key: expiry in ms as big-endian u64, then the frame id. Sorts
2252/// by expiry, so a range scan up to now yields exactly the expired frames.
2253fn idx_expiry_key(expires_at: u64, id: &Scru128Id) -> [u8; 24] {
2254    let mut key = [0u8; 24];
2255    key[..8].copy_from_slice(&expires_at.to_be_bytes());
2256    key[8..].copy_from_slice(id.as_bytes());
2257    key
2258}
2259
2260fn idx_expiry_parse_key(key: &[u8]) -> (u64, Scru128Id) {
2261    let expires_at = u64::from_be_bytes(key[..8].try_into().unwrap());
2262    let id = Scru128Id::from_bytes(key[8..24].try_into().unwrap());
2263    (expires_at, id)
2264}
2265
2266/// Live frames held for a following read that has not taken them yet.
2267///
2268/// A reader that falls further behind than this has its read ended rather
2269/// than silently skipped: see the `Lagged` arm in [`read`](Store::read).
2270/// Raising it buys a slow reader more room before that happens, at a frame
2271/// of memory per slot per follower.
2272const FOLLOW_BUFFER: usize = 4096;
2273
2274const NULL_DELIMITER: u8 = 0;
2275const MAX_TOPIC_LENGTH: usize = 255;
2276
2277/// Validate a frame topic (per ADR 0001).
2278///
2279/// A topic must be non-empty and at most 255 bytes, start with an ASCII letter
2280/// or `_`, and contain only `a-z A-Z 0-9 _ - .`. It may not end with `.` or
2281/// contain consecutive dots. [`Store::append`] runs this automatically; call it
2282/// directly to validate user input before building a [`Frame`].
2283///
2284/// ```
2285/// use xs::store::validate_topic;
2286///
2287/// assert!(validate_topic("clip.add").is_ok());
2288/// assert!(validate_topic("clip.").is_err());
2289/// ```
2290pub fn validate_topic(topic: &str) -> Result<(), crate::error::Error> {
2291    if topic.is_empty() {
2292        return Err("Topic cannot be empty".to_string().into());
2293    }
2294    if topic.len() > MAX_TOPIC_LENGTH {
2295        return Err(format!("Topic exceeds max length of {MAX_TOPIC_LENGTH} bytes").into());
2296    }
2297    if topic.ends_with('.') {
2298        return Err("Topic cannot end with '.'".to_string().into());
2299    }
2300    if topic.contains("..") {
2301        return Err("Topic cannot contain consecutive dots".to_string().into());
2302    }
2303
2304    let bytes = topic.as_bytes();
2305    let first = bytes[0];
2306    if !first.is_ascii_alphabetic() && first != b'_' {
2307        return Err("Topic must start with a-z, A-Z, or _".to_string().into());
2308    }
2309
2310    for &b in bytes {
2311        if !b.is_ascii_alphanumeric() && b != b'_' && b != b'-' && b != b'.' {
2312            return Err(format!(
2313                "Topic contains invalid character: '{}'. Allowed: a-z A-Z 0-9 _ - .",
2314                b as char
2315            )
2316            .into());
2317        }
2318    }
2319
2320    Ok(())
2321}
2322
2323/// Validates a topic query (for --topic flag).
2324/// Accepts a comma-separated list of patterns; each element is an exact
2325/// topic, a `prefix.*` wildcard, or `*` (match all).
2326pub fn validate_topic_query(spec: &str) -> Result<(), crate::error::Error> {
2327    let mut seen = false;
2328    for part in spec.split(',') {
2329        let part = part.trim();
2330        if part.is_empty() {
2331            continue;
2332        }
2333        seen = true;
2334        validate_topic_pattern(part)?;
2335    }
2336    if !seen {
2337        return Err("Topic query cannot be empty".to_string().into());
2338    }
2339    Ok(())
2340}
2341
2342/// Validate a single topic pattern: "*", "prefix.*", or an exact topic.
2343fn validate_topic_pattern(topic: &str) -> Result<(), crate::error::Error> {
2344    if topic == "*" {
2345        return Ok(());
2346    }
2347    if let Some(prefix) = topic.strip_suffix(".*") {
2348        // Validate the prefix part (e.g., "user" in "user.*")
2349        // Prefix can be empty edge case: ".*" is not valid
2350        if prefix.is_empty() {
2351            return Err("Wildcard '.*' requires a prefix".to_string().into());
2352        }
2353        validate_topic(prefix)
2354    } else {
2355        validate_topic(topic)
2356    }
2357}
2358
2359/// Generate prefix index keys for hierarchical topic queries.
2360/// For topic "user.id1.messages", returns keys for prefixes "user." and "user.id1."
2361fn idx_topic_prefix_keys(topic: &str, frame_id: &scru128::Scru128Id) -> Vec<Vec<u8>> {
2362    let mut keys = Vec::new();
2363    let mut pos = 0;
2364    while let Some(dot_pos) = topic[pos..].find('.') {
2365        let prefix = &topic[..pos + dot_pos + 1]; // include the dot
2366        let mut key = Vec::with_capacity(prefix.len() + 1 + 16);
2367        key.extend(prefix.as_bytes());
2368        key.push(NULL_DELIMITER);
2369        key.extend(frame_id.as_bytes());
2370        keys.push(key);
2371        pos += dot_pos + 1;
2372    }
2373    keys
2374}
2375
2376/// An `idx_topic` value meaning "this frame has no `time:` ttl". Distinct from
2377/// the empty value an older store wrote, which means "unknown".
2378const IDX_TOPIC_NEVER: u64 = u64::MAX;
2379
2380/// One frame a `last:n` trim is going to remove, with everything
2381/// [`Store::remove_many`] needs to delete its keys. The topic is the one being
2382/// trimmed and the expiry came off the index entry, so no point read is needed.
2383#[derive(Debug, Clone, Copy)]
2384struct Overflow {
2385    id: Scru128Id,
2386    /// `Some` when the frame has a `time:` ttl and so an expiry index entry.
2387    /// `None` when it has none, or when the entry predates the index carrying
2388    /// one -- see [`idx_topic_expiry_of`].
2389    expires_at: Option<u64>,
2390}
2391
2392/// The expiry an `idx_topic` value carries, if it carries one.
2393///
2394/// `None` covers two cases the caller must treat alike: a frame with no `time:`
2395/// ttl, which has no expiry index entry to delete, and an entry written before
2396/// the index carried an expiry, whose frame may have one. Removing a key that
2397/// is not there is a no-op, so both are safe to treat as "nothing to delete" --
2398/// but the second leaves a stale expiry entry, which the sweeper drops on its
2399/// next pass when the point read finds no frame.
2400#[inline]
2401fn idx_topic_expiry_of(value: &[u8]) -> Option<u64> {
2402    match <[u8; 8]>::try_from(value) {
2403        Ok(bytes) => match u64::from_be_bytes(bytes) {
2404            IDX_TOPIC_NEVER => None,
2405            at => Some(at),
2406        },
2407        Err(_) => None,
2408    }
2409}
2410
2411/// The `idx_topic` value for a frame: its expiry in ms as a big-endian u64, or
2412/// [`IDX_TOPIC_NEVER`]. Every entry for the frame -- the exact topic key and
2413/// each prefix key -- gets the same eight bytes, so a topic scan can drop an
2414/// expired id without a point read into `stream`.
2415fn idx_topic_expiry_value(frame: &Frame) -> [u8; 8] {
2416    expires_at(frame).unwrap_or(IDX_TOPIC_NEVER).to_be_bytes()
2417}
2418
2419/// Has the frame behind this `idx_topic` value expired as of `now`? A value
2420/// that is not eight bytes -- an entry written before the index carried an
2421/// expiry -- reads as "unknown" and answers `false`, so the caller falls back
2422/// to reading the frame and letting [`is_expired`] decide. Both shapes can sit
2423/// in one index.
2424///
2425/// The arithmetic matches [`expires_at`] exactly, so this never hides a frame
2426/// [`is_expired`] would keep.
2427#[inline]
2428fn idx_topic_is_expired(value: &[u8], now: u64) -> bool {
2429    match <[u8; 8]>::try_from(value) {
2430        Ok(bytes) => {
2431            let expires_at = u64::from_be_bytes(bytes);
2432            expires_at != IDX_TOPIC_NEVER && now >= expires_at
2433        }
2434        Err(_) => false,
2435    }
2436}
2437
2438/// The frame id an `idx_topic` hit points at, or `None` to skip the hit
2439/// entirely.
2440///
2441/// The entry's value carries the frame's expiry, so an already expired id is
2442/// dropped here without the point read into `stream` that seeing its ttl would
2443/// otherwise take. `now` is sampled once when the iterator is built, which only
2444/// ever skips fewer frames than a fresh sample would: [`is_expired`] still runs
2445/// on everything that gets through, so the two never disagree about what is
2446/// visible.
2447#[inline]
2448fn idx_topic_hit_id(guard: fjall::Guard, now: u64) -> Option<Scru128Id> {
2449    let (key, value) = guard.into_inner().ok()?;
2450    if idx_topic_is_expired(&value, now) {
2451        return None;
2452    }
2453    Some(idx_topic_frame_id_from_key(&key))
2454}
2455
2456/// The `idx_topic` key range covering one index prefix, opening at `start`
2457/// when there is one.
2458///
2459/// Every key under a prefix is that prefix followed by the frame's 16 byte id,
2460/// and ids sort ascending, so a bound on the id is a bound on the key. Seeking
2461/// to it beats walking the topic from its first entry and discarding
2462/// everything below: on a topic holding 500k frames, reading `after` a recent
2463/// id touches a handful of index entries instead of all 500k.
2464///
2465/// The upper bound is the prefix plus all ones, the largest key the prefix can
2466/// hold. A topic never contains the NUL delimiter, so no other topic's keys
2467/// fall inside the range.
2468fn idx_topic_range(
2469    prefix: Vec<u8>,
2470    start: Option<(Scru128Id, bool)>,
2471) -> (Bound<Vec<u8>>, Bound<Vec<u8>>) {
2472    fn key_at(prefix: &[u8], id: &[u8; 16]) -> Vec<u8> {
2473        let mut key = Vec::with_capacity(prefix.len() + 16);
2474        key.extend_from_slice(prefix);
2475        key.extend_from_slice(id);
2476        key
2477    }
2478
2479    let hi = Bound::Included(key_at(&prefix, &[0xff; 16]));
2480    let lo = match start {
2481        Some((id, true)) => Bound::Included(key_at(&prefix, id.as_bytes())),
2482        Some((id, false)) => Bound::Excluded(key_at(&prefix, id.as_bytes())),
2483        // Shorter than every key it covers, so it sorts before all of them.
2484        None => Bound::Included(prefix),
2485    };
2486    (lo, hi)
2487}
2488
2489fn idx_topic_key_prefix(topic: &str) -> Vec<u8> {
2490    let mut v = Vec::with_capacity(topic.len() + 1); // topic bytes + delimiter
2491    v.extend(topic.as_bytes()); // topic string as UTF-8 bytes
2492    v.push(NULL_DELIMITER); // Delimiter for variable-sized keys
2493    v
2494}
2495
2496pub(crate) fn idx_topic_key_from_frame(frame: &Frame) -> Result<Vec<u8>, crate::error::Error> {
2497    validate_topic(&frame.topic)?;
2498    let mut v = idx_topic_key_prefix(&frame.topic);
2499    v.extend(frame.id.as_bytes());
2500    Ok(v)
2501}
2502
2503fn idx_topic_frame_id_from_key(key: &[u8]) -> Scru128Id {
2504    let frame_id_bytes = &key[key.len() - 16..];
2505    Scru128Id::from_bytes(frame_id_bytes.try_into().unwrap())
2506}
2507
2508fn deserialize_frame<B1: AsRef<[u8]> + std::fmt::Debug, B2: AsRef<[u8]>>(
2509    record: (B1, B2),
2510) -> Frame {
2511    serde_json::from_slice(record.1.as_ref()).unwrap_or_else(|e| {
2512        // Try to convert the key to a Scru128Id and print in a format that can be copied for deletion
2513        let key_bytes = record.0.as_ref();
2514        if key_bytes.len() == 16 {
2515            if let Ok(bytes) = key_bytes.try_into() {
2516                let id = Scru128Id::from_bytes(bytes);
2517                eprintln!("CORRUPTED_RECORD_ID: {id}");
2518            }
2519        }
2520        let key = std::str::from_utf8(record.0.as_ref()).unwrap();
2521        let value = std::str::from_utf8(record.1.as_ref()).unwrap();
2522        panic!("Failed to deserialize frame: {e} {key} {value}")
2523    })
2524}