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
use super::*;
#[test]
fn t7_compact_boundary_durable_before_visible() {
// F2 ordering proof (durable-before-visible at the compact boundary), pinned
// by a flush-epoch watermark rather than only end-state. A synthetic
// FlushEpochDurable + FlushEpochObserver share a per-flush watermark; the
// durable records each commit's APPEND epoch (before its group flush) and the
// observer records each node's PUBLISH epoch (at fan-out). Because the group
// flush bumps the watermark between Stage-1 append and Stage-3 publish,
// publish_epoch(A,B) > append_epoch(A,B) proves A, B were FLUSHED before they
// were published. Compaction reclaims the in-memory holes (no real WAL needed
// — the dense layout comes from the graph, not the durable; real-WAL compact
// recovery is covered by T8). The compact is a hard flush boundary (F2):
// A, B flush+publish+ack BEFORE the dense Arc stores, and the final snapshot
// is dense AND contains A, B (the dense store is LAST).
let flush_epoch: FlushEpoch = Arc::new(AtomicU64::new(0));
let durable = FlushEpochDurable::new(b"FEP0", Arc::clone(&flush_epoch));
let observer = FlushEpochObserver::new(b"FOB0", Arc::clone(&flush_epoch));
let shared = Arc::new(
SharedGraph::from_graph_with_core_and_durables(
SeleneGraph::new(GraphId::new(70_010)),
vec![observer.clone() as Arc<dyn IndexProvider>],
vec![durable.clone()],
None,
None,
on(8, 8 * 1024 * 1024),
)
.unwrap(),
);
// Seed reclaimable holes so the compact actually densifies.
{
let mut txn = shared.begin_write();
let mut ids = Vec::new();
for _ in 0..20 {
ids.push(
txn.mutator()
.create_node(LabelSet::single(db_string("S")), PropertyMap::new())
.unwrap(),
);
}
txn.commit().unwrap();
let mut txn = shared.begin_write();
for id in &ids {
txn.mutator().delete_node(*id).unwrap();
}
txn.commit().unwrap();
}
// Seal A, B but do not submit.
let mut txn_a = shared.begin_write();
let a = txn_a
.mutator()
.create_node(LabelSet::single(db_string("A")), PropertyMap::new())
.unwrap();
let sealed_a = txn_a.seal(None, None).expect("A seals");
let mut txn_b = shared.begin_write();
let b = txn_b
.mutator()
.create_node(LabelSet::single(db_string("B")), PropertyMap::new())
.unwrap();
let sealed_b = txn_b.seal(None, None).expect("B seals");
// Run compact on a background thread: it allocates a seal_seq AFTER A, B
// under the lock, then submits its dense publish and blocks. Yield so the
// compact enqueues before A, B, exercising the reorder buffer.
let s_c = Arc::clone(&shared);
let compactor = thread::spawn(move || s_c.compact().expect("compaction ok"));
for _ in 0..2_000 {
thread::yield_now();
}
// Submit B then A (reverse) so the contiguous run [A,B] forms only after A.
let s_b = Arc::clone(&shared);
let b_thread = thread::spawn(move || s_b.submit_sealed_for_test(sealed_b).expect("B"));
for _ in 0..1_000 {
thread::yield_now();
}
let outcome_a = shared.submit_sealed_for_test(sealed_a).expect("A");
let outcome_b = b_thread.join().unwrap();
let report = compactor.join().expect("compactor ok");
// A, B durable (acked with a durable_at) and visible; report reclaimed.
let a_seq = outcome_a.durable_at.expect("A durable_at");
let b_seq = outcome_b.durable_at.expect("B durable_at");
assert!(report.reclaimed_nodes >= 20, "report: {report:?}");
// ORDERING PROOF (the F2 / R1 barrier): for both A and B the group flush bumped
// the shared watermark strictly between their append and their publish, so
// each was durable BEFORE it became visible.
let a_pub = observer.publish_epoch_of(a.get());
let b_pub = observer.publish_epoch_of(b.get());
assert!(
a_pub > durable.append_epoch_of(a_seq),
"A published (epoch {a_pub}) only after its group flush (append epoch {})",
durable.append_epoch_of(a_seq),
);
assert!(
b_pub > durable.append_epoch_of(b_seq),
"B published (epoch {b_pub}) only after its group flush (append epoch {})",
durable.append_epoch_of(b_seq),
);
// Final snapshot is the dense compacted one AND contains A, B (the dense
// store ran LAST — after A, B flushed + published).
let snap = shared.read();
assert!(snap.is_node_alive(a));
assert!(snap.is_node_alive(b));
assert_eq!(snap.node_count(), 2, "only A, B alive");
assert_eq!(
snap.node_store.len(),
2,
"published snapshot is dense (compaction not clobbered by A,B's stale snapshot)",
);
snap.assert_indexes_consistent()
.expect("structurally consistent");
}
#[test]
fn t7b_compact_at_head_publishes_with_zero_flush_calls() {
// A compact whose seal_seq is at next_publish_seq (no pending commit run)
// publishes the dense Arc with ZERO flush calls — all lower seqs already
// durable + visible.
let flush_epoch: FlushEpoch = Arc::new(AtomicU64::new(0));
let durable = FlushEpochDurable::new(b"FEP1", flush_epoch);
let shared = SharedGraph::from_graph_with_core_and_durables(
SeleneGraph::new(GraphId::new(70_011)),
Vec::new(),
vec![durable.clone()],
None,
None,
on(8, 8 * 1024 * 1024),
)
.unwrap();
// Two committed-then-deleted nodes, fully published + flushed.
let mut txn = shared.begin_write();
let ids: Vec<_> = (0..4)
.map(|_| {
txn.mutator()
.create_node(LabelSet::single(db_string("H")), PropertyMap::new())
.unwrap()
})
.collect();
txn.commit().unwrap();
let mut txn = shared.begin_write();
for id in &ids {
txn.mutator().delete_node(*id).unwrap();
}
txn.commit().unwrap();
let flushes_before = durable.flushes.load(Ordering::SeqCst);
// Compact is now strictly at head (nothing else pending) — publishes solo.
let report = shared.compact().expect("compact ok");
assert!(report.reclaimed_nodes >= 4);
assert_eq!(
durable.flushes.load(Ordering::SeqCst),
flushes_before,
"compact-at-head issues zero flush calls",
);
}
// ───────────────── T2b: within-batch commit flush-order (R1 barrier) ─────────
#[test]
fn t2b_within_batch_commits_flush_before_publish() {
// The headline durable-before-visible guarantee for GROUPED commits (not just
// at the compact boundary): the single group flush precedes EVERY commit's
// publish in the run. Fan in a contiguous run under On; the FlushEpochDurable
// records each commit's append epoch, the FlushEpochObserver records each
// node's publish epoch, and we assert publish_epoch > append_epoch for every
// commit — i.e. the R1 barrier sits strictly between Stage-1 append and
// Stage-3 publish for grouped commits. Buffer the later seqs behind a gap and
// release seq 0 last so a genuine multi-member batch forms (>= 2 in one run).
let flush_epoch: FlushEpoch = Arc::new(AtomicU64::new(0));
let durable = FlushEpochDurable::new(b"FEP2", Arc::clone(&flush_epoch));
let observer = FlushEpochObserver::new(b"FOB2", Arc::clone(&flush_epoch));
let shared = Arc::new(
SharedGraph::from_graph_with_core_and_durables(
SeleneGraph::new(GraphId::new(70_012)),
vec![observer.clone() as Arc<dyn IndexProvider>],
vec![durable.clone()],
None,
None,
on(8, 8 * 1024 * 1024),
)
.unwrap(),
);
// Seal 4 commits (seal_seq 0..3) each forking the prior.
let mut sealeds = Vec::new();
let mut ids = Vec::new();
for label in ["p", "q", "r", "s"] {
let mut txn = shared.begin_write();
let id = txn
.mutator()
.create_node(LabelSet::single(db_string(label)), PropertyMap::new())
.unwrap();
ids.push(id);
sealeds.push(txn.seal(None, None).expect("seals"));
}
// Withhold seq 0; submit seqs 3,2,1 first (buffer behind the gap), then seq 0
// last so the full [0,1,2,3] contiguous run drains as ONE batch with ONE flush.
let sealed_0 = sealeds.remove(0);
let mut handles = Vec::new();
while let Some(sealed) = sealeds.pop() {
let shared = Arc::clone(&shared);
handles.push(thread::spawn(move || {
shared
.submit_sealed_for_test(sealed)
.expect("buffered commit")
}));
for _ in 0..1_000 {
thread::yield_now();
}
}
let outcome_0 = shared.submit_sealed_for_test(sealed_0).expect("seq 0");
let mut durable_seqs = vec![outcome_0.durable_at.expect("durable_at")];
for handle in handles {
durable_seqs.push(handle.join().unwrap().durable_at.expect("durable_at"));
}
// The run grouped (fewer flushes than commits) — otherwise this would only be
// testing the degenerate cap-1 path.
assert!(
durable.flushes.load(Ordering::SeqCst) < 4,
"the 4 commits grouped into fewer than 4 flushes (got {})",
durable.flushes.load(Ordering::SeqCst),
);
// R1 barrier: every commit's append epoch is strictly below every commit's
// publish epoch — the group flush bumped the watermark strictly between
// Stage-1 append and Stage-3 publish for the whole run. Order-independent
// (max append epoch < min publish epoch) so the proof does not rely on which
// node id maps to which assigned seq.
let max_append = durable_seqs
.iter()
.map(|seq| durable.append_epoch_of(*seq))
.max()
.expect("at least one commit");
let min_publish = ids
.iter()
.map(|id| observer.publish_epoch_of(id.get()))
.min()
.expect("at least one node");
assert!(
min_publish > max_append,
"the group flush separated every append (max epoch {max_append}) from every \
publish (min epoch {min_publish}) — durable-before-visible for grouped commits",
);
assert_eq!(shared.read().node_count(), 4);
assert_eq!(shared.read().meta.generation, 4);
}