kevy-persist 3.8.0

kevy persistence — RDB-style snapshots + AOF, pure Rust, zero deps.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
//! AOF-rewrite serialization: render the live keyspace as the minimal set of
//! RESP write commands that reconstruct it (`BGREWRITEAOF`'s output), plus the
//! shared multi-bulk frame writer / size estimator the live append path uses.
//!
//! Split out of `lib.rs` (the binary-snapshot format) to keep both files under
//! the 500-LOC house cap. TTL is emitted as an absolute `PEXPIREAT` deadline
//! so a replay reconstructs the original instant (INC-2026-06-09) rather than
//! re-anchoring to replay-time.

use crate::SNAPSHOT_BUF_CAP;
use kevy_resp::{Argv, ArgvView};
use kevy_store::{StreamData, StreamId, Value};
use std::fs::File;
use std::io::{self, BufWriter, Write};
use std::path::Path;

/// Write `src`'s state (a live `Store` or a frozen
/// [`kevy_store::SnapshotView`]) to `path` as a sequence of mutating RESP
/// commands prefixed with `crate::aof::AOF_MAGIC`; flush + fsync before
/// returning. Returns `(keys, bytes)`. The magic header is consistent with
/// `Aof::open`'s fresh-file behavior so BGREWRITEAOF-produced files replay
/// the same way live-appended ones do.
///
/// `pub` (not just crate-internal) because the COW rewrite path calls it
/// from a background thread: [`crate::Aof::begin_view_rewrite`] starts the
/// tee, this serializes the frozen view to the temp file off-thread, and
/// `finish_concurrent_rewrite` swaps it in.
pub fn dump_aof<S: crate::SnapshotSource>(path: &Path, src: &S) -> io::Result<(u64, u64)> {
    let f = File::create(path)?;
    let mut w = BufWriter::with_capacity(SNAPSHOT_BUF_CAP, f);
    w.write_all(crate::aof::AOF_MAGIC)?;
    let mut keys = 0u64;
    let mut err: Option<io::Error> = None;
    src.for_each_entry(|key, value, ttl_ms| {
        if err.is_some() {
            return;
        }
        if let Err(e) = write_value_as_commands(&mut w, key, value, ttl_ms) {
            err = Some(e);
        } else {
            keys += 1;
        }
    });
    if let Some(e) = err {
        return Err(e);
    }
    // v2.4: hash field TTLs re-emitted as absolute HPEXPIREAT frames
    // (after the HSETs that recreate their fields).
    let mut ferr: Option<io::Error> = None;
    src.for_each_hash_ttl(|key, field, deadline_ms| {
        if ferr.is_some() {
            return;
        }
        let ms = deadline_ms.to_string();
        let mut argv = kevy_resp::Argv::with_capacity(6, 0);
        argv.push(b"HPEXPIREAT");
        argv.push(key);
        argv.push(ms.as_bytes());
        argv.push(b"FIELDS");
        argv.push(b"1");
        argv.push(field);
        if let Err(e) = write_multibulk(&mut w, &argv) {
            ferr = Some(e);
        }
    });
    if let Some(e) = ferr {
        return Err(e);
    }
    w.flush()?;
    let inner = w
        .into_inner()
        .map_err(|e| io::Error::other(e.to_string()))?;
    let bytes = inner.metadata().map_or(0, |m| m.len());
    inner.sync_all()?;
    Ok((keys, bytes))
}

/// Serialize `src`'s state into an in-memory AOF image (magic + the same
/// RESP command stream [`dump_aof`] writes). Returns the bytes and the key
/// count. Used by the non-blocking rewrite: the caller produces this buffer
/// under the store lock, then spills it to disk *off* the lock. `Vec<u8>`
/// is an infallible `Write`, so no error path exists.
pub(crate) fn dump_store_to_buf<S: crate::SnapshotSource>(src: &S) -> (Vec<u8>, u64) {
    let mut buf = Vec::with_capacity(crate::SNAPSHOT_BUF_CAP);
    buf.extend_from_slice(crate::aof::AOF_MAGIC);
    let mut keys = 0u64;
    src.for_each_entry(|key, value, ttl_ms| {
        let _ = write_value_as_commands(&mut buf, key, value, ttl_ms);
        keys += 1;
    });
    (buf, keys)
}

/// Emit one (or two, if TTL'd) RESP write commands that, when replayed,
/// reconstruct `key`'s `value` and TTL exactly.
///
/// C6: marked `#[cold]` — AOF rewrite only runs on `BGREWRITEAOF`
/// or `auto-aof-rewrite-percentage`, never on the live write path.
/// The full type-switch over `Value` variants is ~9 KB; pushing it
/// off the hot iTLB pages around `start_command` is the point.
#[cold]
fn write_value_as_commands<W: Write>(
    w: &mut W,
    key: &[u8],
    value: &Value,
    ttl_ms: Option<u64>,
) -> io::Result<()> {
    match value {
        Value::Str(s) => {
            let argv = Argv::from(vec![b"SET".to_vec(), key.to_vec(), s.to_vec()]);
            write_multibulk(w, &argv)?;
        }
        // L2: persist Int as the canonical ASCII bytes; the replay path's
        // SET will auto-detect it back to Int via parse_canonical_i64.
        Value::Int(n) => {
            let argv = Argv::from(vec![b"SET".to_vec(), key.to_vec(), n.to_string().into_bytes()]);
            write_multibulk(w, &argv)?;
        }
        // L1: Arc-bulk serialises via the same SET argv path; replay's
        // SET routing picks ArcBulk again for > BULK_THRESHOLD bytes.
        Value::ArcBulk(a) => {
            let argv = Argv::from(vec![b"SET".to_vec(), key.to_vec(), a.as_ref().to_vec()]);
            write_multibulk(w, &argv)?;
        }
        Value::Hash(h) => {
            let mut argv: Vec<Vec<u8>> = Vec::with_capacity(2 + h.len() * 2);
            argv.push(b"HSET".to_vec());
            argv.push(key.to_vec());
            for (f, v) in h.iter() {
                argv.push(f.to_vec());
                argv.push(v.clone());
            }
            write_multibulk(w, &Argv::from(argv))?;
        }
        // A.8: inline hash rewrites to the same HSET command form as the
        // heap-backed `Value::Hash`. Replay routes the pairs back through
        // the encoding switch so small hashes land inline again.
        Value::SmallHashInline(h) => {
            let mut argv: Vec<Vec<u8>> = Vec::with_capacity(2 + h.len() * 2);
            argv.push(b"HSET".to_vec());
            argv.push(key.to_vec());
            for (f, v) in h.iter() {
                argv.push(f.to_vec());
                argv.push(v.to_vec());
            }
            write_multibulk(w, &Argv::from(argv))?;
        }
        Value::List(l) => {
            let mut argv: Vec<Vec<u8>> = Vec::with_capacity(2 + l.len());
            argv.push(b"RPUSH".to_vec());
            argv.push(key.to_vec());
            for v in l.iter() {
                argv.push(v.clone());
            }
            write_multibulk(w, &Argv::from(argv))?;
        }
        // A.8: inline list rewrites to the same RPUSH command form as
        // the heap-backed `Value::List`.
        Value::SmallListInline(l) => {
            let mut argv: Vec<Vec<u8>> = Vec::with_capacity(2 + l.len());
            argv.push(b"RPUSH".to_vec());
            argv.push(key.to_vec());
            for v in l.iter() {
                argv.push(v.to_vec());
            }
            write_multibulk(w, &Argv::from(argv))?;
        }
        Value::Set(s) => {
            let mut argv: Vec<Vec<u8>> = Vec::with_capacity(2 + s.len());
            argv.push(b"SADD".to_vec());
            argv.push(key.to_vec());
            for m in s.iter() {
                argv.push(m.to_vec());
            }
            write_multibulk(w, &Argv::from(argv))?;
        }
        // A.7 O5: inline-encoded set rewrites to the same SADD command
        // form as the heap-backed `Value::Set`. Replay through the live
        // SADD handler routes the members back through the encoding
        // switch — small sets land in `SmallSetInline` again, oversized
        // ones promote to `KevySet` naturally.
        Value::SmallSetInline(s) => {
            let mut argv: Vec<Vec<u8>> = Vec::with_capacity(2 + s.len());
            argv.push(b"SADD".to_vec());
            argv.push(key.to_vec());
            for m in s.iter() {
                argv.push(m.to_vec());
            }
            write_multibulk(w, &Argv::from(argv))?;
        }
        Value::ZSet(z) => {
            let mut argv: Vec<Vec<u8>> = Vec::with_capacity(2 + z.ordered().count() * 2);
            argv.push(b"ZADD".to_vec());
            argv.push(key.to_vec());
            for (m, sc) in z.ordered() {
                argv.push(fmt_zset_score(sc));
                argv.push(m.to_vec());
            }
            write_multibulk(w, &Argv::from(argv))?;
        }
        // A.8: inline zset rewrites to the same ZADD command form as
        // the heap-backed `Value::ZSet`.
        Value::SmallZSetInline(z) => {
            let mut argv: Vec<Vec<u8>> = Vec::with_capacity(2 + z.len() * 2);
            argv.push(b"ZADD".to_vec());
            argv.push(key.to_vec());
            for (m, sc) in z.iter() {
                argv.push(fmt_zset_score(sc));
                argv.push(m.to_vec());
            }
            write_multibulk(w, &Argv::from(argv))?;
        }
        Value::Stream(s) => write_stream_as_commands(w, key, s)?,
    }
    if let Some(ms) = ttl_ms {
        // `ms` is remaining; emit an absolute `PEXPIREAT` deadline so a replay
        // of this rewritten AOF reconstructs the original instant instead of
        // re-anchoring to replay-time (INC-2026-06-09).
        let deadline = kevy_store::now_unix_ms().saturating_add(ms);
        let argv = Argv::from(vec![
            b"PEXPIREAT".to_vec(),
            key.to_vec(),
            deadline.to_string().into_bytes(),
        ]);
        write_multibulk(w, &argv)?;
    }
    Ok(())
}

/// Render one stream as commands: one XADD per entry (slow on huge
/// streams but correct — a multi-entry XADD batch is a future parser
/// feature), then `XSETID` whenever a bare replay of those XADDs would
/// not reproduce the scalar state (deleted tail, deleted-only stream,
/// non-zero `entries_added` drift), then the consumer-group section.
fn write_stream_as_commands<W: Write>(w: &mut W, key: &[u8], s: &StreamData) -> io::Result<()> {
    for (id, fv) in s.iter_entries() {
        let mut argv: Vec<Vec<u8>> = Vec::with_capacity(3 + fv.len() * 2);
        argv.push(b"XADD".to_vec());
        argv.push(key.to_vec());
        argv.push(id.encode());
        for (f, v) in fv {
            argv.push(f.to_vec());
            argv.push(v.to_vec());
        }
        write_multibulk(w, &Argv::from(argv))?;
    }
    let (len, last, mxd, added) =
        (s.length(), s.last_id(), s.max_deleted_id(), s.entries_added());
    if len == 0 && last != StreamId::MIN {
        // Empty stream whose ID clock advanced (all entries deleted):
        // re-create the key with the right `last_id` via the
        // `XADD MAXLEN 0` trick — the inline trim wipes the dummy row.
        let argv = vec![
            b"XADD".to_vec(), key.to_vec(), b"MAXLEN".to_vec(), b"0".to_vec(),
            last.encode(), b"x".to_vec(), b"x".to_vec(),
        ];
        write_multibulk(w, &Argv::from(argv))?;
    }
    // What replaying the commands emitted so far yields. The only no-key
    // case left is the virgin empty stream (groups-only) — its scalars
    // are all zero by construction, so skipping XSETID there is exact.
    let natural = if len > 0 {
        (s.last_entry().map_or(StreamId::MIN, |(id, _)| id), len, StreamId::MIN)
    } else {
        (last, u64::from(last != StreamId::MIN), last)
    };
    if natural != (last, added, mxd) {
        let argv = vec![
            b"XSETID".to_vec(), key.to_vec(), last.encode(),
            b"ENTRIESADDED".to_vec(), added.to_string().into_bytes(),
            b"MAXDELETEDID".to_vec(), mxd.encode(),
        ];
        write_multibulk(w, &Argv::from(argv))?;
    }
    write_stream_group_commands(w, key, s)
}

/// Consumer-group section of a stream rewrite: `XGROUP CREATE … MKSTREAM`
/// (MKSTREAM covers groups on a virgin empty stream), one CREATECONSUMER
/// per known consumer, then one `XCLAIM … TIME t RETRYCOUNT n FORCE JUSTID`
/// per live PEL row — full delivery_time/count fidelity, the same technique
/// Redis's own AOF rewrite uses. Tombstone PEL rows (entry XDEL'd while
/// pending) are skipped: XCLAIM purges rather than re-creates those, so
/// only the snapshot path preserves them (RFC 2026-06-11 trade-off).
fn write_stream_group_commands<W: Write>(
    w: &mut W,
    key: &[u8],
    s: &StreamData,
) -> io::Result<()> {
    for g in s.export_groups() {
        let last_delivered =
            StreamId { ms: g.last_delivered.0, seq: g.last_delivered.1 };
        let argv = vec![
            b"XGROUP".to_vec(), b"CREATE".to_vec(), key.to_vec(), g.name.clone(),
            last_delivered.encode(), b"MKSTREAM".to_vec(),
        ];
        write_multibulk(w, &Argv::from(argv))?;
        for (consumer, _last_seen_ms) in &g.consumers {
            let argv = vec![
                b"XGROUP".to_vec(), b"CREATECONSUMER".to_vec(), key.to_vec(),
                g.name.clone(), consumer.clone(),
            ];
            write_multibulk(w, &Argv::from(argv))?;
        }
        for (ms, seq, consumer, delivery_time_ms, delivery_count) in &g.pel {
            let id = StreamId { ms: *ms, seq: *seq };
            if !s.contains_entry(id) {
                continue;
            }
            let argv = vec![
                b"XCLAIM".to_vec(), key.to_vec(), g.name.clone(), consumer.clone(),
                b"0".to_vec(), id.encode(),
                b"TIME".to_vec(), delivery_time_ms.to_string().into_bytes(),
                b"RETRYCOUNT".to_vec(), delivery_count.to_string().into_bytes(),
                b"FORCE".to_vec(), b"JUSTID".to_vec(),
            ];
            write_multibulk(w, &Argv::from(argv))?;
        }
    }
    Ok(())
}

/// Format a sorted-set score the way Redis does (no trailing `.0` for
/// integers; up to 17 sig figs for non-integer doubles). Tests want the
/// replay-roundtrip to compare byte-equal, so don't introduce locale
/// differences (`format!` is locale-free here).
fn fmt_zset_score(s: f64) -> Vec<u8> {
    // Bit-exact compare is the contract — "no fractional bits in the f64",
    // not "approximately integer". An epsilon would mis-classify near-int
    // values as integers and change wire bytes.
    #[allow(clippy::float_cmp)]
    let is_integer_valued = s.is_finite() && s == s.trunc();
    if is_integer_valued && s.abs() < 1e17 {
        format!("{}", s as i64).into_bytes()
    } else {
        format!("{s:.17}").into_bytes()
    }
}

/// Cheap byte-count estimator for a single multi-bulk frame:
/// `*<n>\r\n` + per-arg `$<len>\r\n<bytes>\r\n`. No allocation, no
/// double-pass — accurate to within a couple of bytes per arg.
pub(crate) fn estimate_multibulk_bytes<A: ArgvView + ?Sized>(args: &A) -> u64 {
    let mut n: u64 = 3 + u64::from(decimal_digits(args.len() as u64));
    for i in 0..args.len() {
        let a = &args[i];
        n += 3 + u64::from(decimal_digits(a.len() as u64)) + a.len() as u64 + 2;
    }
    n
}

#[inline]
fn decimal_digits(mut x: u64) -> u32 {
    if x == 0 {
        return 1;
    }
    let mut d = 0;
    while x > 0 {
        d += 1;
        x /= 10;
    }
    d
}

pub(crate) fn write_multibulk<W: Write, A: ArgvView + ?Sized>(
    w: &mut W,
    args: &A,
) -> io::Result<()> {
    write!(w, "*{}\r\n", args.len())?;
    for i in 0..args.len() {
        let a = &args[i];
        write!(w, "${}\r\n", a.len())?;
        w.write_all(a)?;
        w.write_all(b"\r\n")?;
    }
    Ok(())
}

/// Parity manifest (v2.1): every verb the AOF rewrite emits.
/// Cross-checked against `kevy_resp::ops_table` below.
#[cfg_attr(not(test), allow(dead_code))]
pub(crate) const REWRITE_EMIT_VERBS: &[&str] = &[
    "SET", "HSET", "RPUSH", "SADD", "ZADD", "PEXPIREAT", "HPEXPIREAT",
    "XADD", "XSETID", "XGROUP", "XCLAIM",
];

#[cfg(test)]
mod op_table_parity {
    use super::REWRITE_EMIT_VERBS;
    use kevy_resp::ops_table::{ops_with, surface};
    use std::collections::BTreeSet;

    #[test]
    fn rewrite_manifest_matches_table() {
        let m: BTreeSet<&str> = REWRITE_EMIT_VERBS.iter().copied().collect();
        let t: BTreeSet<&str> = ops_with(surface::REWRITE).into_iter().collect();
        assert_eq!(
            m, t,
            "rewrite emit set != OP_TABLE REWRITE flags — update both when changing rewrite_fmt"
        );
    }

    #[test]
    fn rewrite_manifest_verbs_have_source_literals() {
        let src = include_str!("rewrite_fmt.rs");
        for v in REWRITE_EMIT_VERBS {
            let lit = format!("\"{v}");
            assert!(
                src.contains(&lit) || src.contains(&format!("b\"{v}\"")),
                "REWRITE_EMIT_VERBS lists {v} but rewrite_fmt.rs has no literal for it"
            );
        }
    }
}