test_perf_pack.py
file-level
1
files
1
commits
0
hotspots
0
🧊 dead
0
💥 blast risk
| 1 | """Phase 3.4 — Pack and unpack at scale. |
| 2 | |
| 3 | Target metrics (measured on a 2024 MacBook Pro M4, macOS 15): |
| 4 | |
| 5 | build_pack (10 000 objects × 4 KiB): < 60 s [@slow] |
| 6 | apply_pack (10 000 objects × 4 KiB): < 60 s [@slow] |
| 7 | verify-pack (10 000 objects × 4 KiB): < 120 s [@slow] (in practice much faster) |
| 8 | collect_object_ids (1 000 objects): < 1 s (no blob reads) |
| 9 | |
| 10 | Edges verified beyond the plan: |
| 11 | |
| 12 | a. ``build_pack`` loads ALL blob bytes simultaneously — peak RSS ≈ 2× blob total. |
| 13 | b. ``have=`` filter correctly reduces both commit count and blob payload. |
| 14 | c. ``MAX_PACK_OBJECTS`` applies to total_items (commits + snapshots + objects), |
| 15 | not per-type — a pack within per-type limits can still be rejected. |
| 16 | d. Oversized object (> MAX_OBJECT_WRITE_BYTES) is silently skipped, not raised — |
| 17 | caller sees objects_skipped++ but no error; documented behaviour. |
| 18 | e. ``verify-pack --stat`` is purely structural — no SHA-256, near-instant. |
| 19 | f. Duplicate OID dedup in ``apply_pack`` — each OID written at most once. |
| 20 | g. Round-trip integrity: build → msgpack-serialize → apply → all objects present. |
| 21 | """ |
| 22 | |
| 23 | from __future__ import annotations |
| 24 | |
| 25 | import datetime |
| 26 | import hashlib |
| 27 | import pathlib |
| 28 | import sys |
| 29 | import tempfile |
| 30 | import time |
| 31 | import tracemalloc |
| 32 | |
| 33 | import msgpack |
| 34 | import pytest |
| 35 | from unittest.mock import patch |
| 36 | |
| 37 | from muse.core.object_store import write_object |
| 38 | from muse.core.pack import ( |
| 39 | MAX_OBJECT_WRITE_BYTES, |
| 40 | MAX_PACK_OBJECTS, |
| 41 | ObjectPayload, |
| 42 | PackBundle, |
| 43 | apply_pack, |
| 44 | build_pack, |
| 45 | collect_object_ids, |
| 46 | ) |
| 47 | from muse.core.snapshot import compute_commit_id, compute_snapshot_id |
| 48 | |
| 49 | from muse.core._types import Manifest |
| 50 | from muse.core.store import ( |
| 51 | CommitRecord, |
| 52 | MAX_PACK_MSGPACK_BYTES, |
| 53 | SnapshotRecord, |
| 54 | write_branch_ref, |
| 55 | write_commit, |
| 56 | write_snapshot, |
| 57 | ) |
| 58 | |
| 59 | # --------------------------------------------------------------------------- |
| 60 | # Helpers |
| 61 | # --------------------------------------------------------------------------- |
| 62 | |
| 63 | |
| 64 | def _sha256(data: bytes) -> str: |
| 65 | return hashlib.sha256(data).hexdigest() |
| 66 | |
| 67 | |
| 68 | def _make_repo(tmp: pathlib.Path) -> pathlib.Path: |
| 69 | tmp.mkdir(parents=True, exist_ok=True) |
| 70 | muse = tmp / ".muse" |
| 71 | muse.mkdir() |
| 72 | (muse / "repo.json").write_text('{"repo_id":"bench","owner":"bench"}') |
| 73 | for d in ("commits", "snapshots", "objects"): |
| 74 | (muse / d).mkdir() |
| 75 | (muse / "refs" / "heads").mkdir(parents=True) |
| 76 | (muse / "HEAD").write_text("ref: refs/heads/main\n") |
| 77 | (muse / "config.toml").write_text("") |
| 78 | return tmp |
| 79 | |
| 80 | |
| 81 | def _fresh_repo(tmp: pathlib.Path) -> pathlib.Path: |
| 82 | tmp.mkdir(parents=True, exist_ok=True) |
| 83 | muse = tmp / ".muse" |
| 84 | muse.mkdir() |
| 85 | (muse / "repo.json").write_text('{"repo_id":"dst"}') |
| 86 | for d in ("commits", "snapshots", "objects"): |
| 87 | (muse / d).mkdir() |
| 88 | return tmp |
| 89 | |
| 90 | |
| 91 | def _populate( |
| 92 | repo: pathlib.Path, |
| 93 | n_commits: int = 10, |
| 94 | n_unique_objects: int = 10, |
| 95 | blob_size: int = 4096, |
| 96 | branch: str = "main", |
| 97 | start: int = 0, |
| 98 | ) -> tuple[str, dict[str, str]]: |
| 99 | """Write *n_unique_objects* blobs and a *n_commits* chain. |
| 100 | |
| 101 | Returns ``(tip_commit_id, {path: oid})`` manifest. |
| 102 | """ |
| 103 | blobs: Manifest = {} |
| 104 | for i in range(n_unique_objects): |
| 105 | data = f"obj-{i + start:08d}-".encode() + b"x" * blob_size |
| 106 | oid = _sha256(data) |
| 107 | write_object(repo, oid, data) |
| 108 | blobs[f"file_{i:04d}.py"] = oid |
| 109 | |
| 110 | sid = compute_snapshot_id(blobs) |
| 111 | write_snapshot(repo, SnapshotRecord(snapshot_id=sid, manifest=blobs)) |
| 112 | |
| 113 | parent: str | None = None |
| 114 | tip = "" |
| 115 | for i in range(n_commits): |
| 116 | ts = datetime.datetime(2026, 1, 1, tzinfo=datetime.timezone.utc) |
| 117 | msg = f"c{start + i:07d}" |
| 118 | cid = compute_commit_id([parent] if parent else [], sid, msg, ts.isoformat()) |
| 119 | rec = CommitRecord( |
| 120 | commit_id=cid, |
| 121 | repo_id="bench", |
| 122 | branch=branch, |
| 123 | snapshot_id=sid, |
| 124 | message=msg, |
| 125 | committed_at=ts, |
| 126 | parent_commit_id=parent, |
| 127 | parent2_commit_id=None, |
| 128 | author="bench", |
| 129 | metadata={}, |
| 130 | structured_delta=None, |
| 131 | sem_ver_bump="none", |
| 132 | breaking_changes=[], |
| 133 | agent_id="", |
| 134 | model_id="", |
| 135 | toolchain_id="", |
| 136 | prompt_hash="", |
| 137 | signature="", |
| 138 | signer_key_id="", |
| 139 | ) |
| 140 | write_commit(repo, rec) |
| 141 | parent = cid |
| 142 | tip = cid |
| 143 | |
| 144 | write_branch_ref(repo, branch, tip) |
| 145 | return tip, blobs |
| 146 | |
| 147 | |
| 148 | # --------------------------------------------------------------------------- |
| 149 | # Phase 3.4.1 — build_pack throughput |
| 150 | # --------------------------------------------------------------------------- |
| 151 | |
| 152 | |
| 153 | class TestBuildPackThroughput: |
| 154 | """build_pack must sustain ≥ 2 000 objects/sec in the object-read loop. |
| 155 | |
| 156 | build_pack's hot path is ``read_object`` for each unique blob. At the |
| 157 | Phase 3.1 floor of 2 000 objects/sec, 100 000 objects take ~50 s, within |
| 158 | the 60 s target. The fast test covers 1 000 objects and asserts the rate |
| 159 | directly; the slow test covers 10 000 objects and proves on-disk timing. |
| 160 | """ |
| 161 | |
| 162 | _MIN_OBJECTS_PER_SEC = 2_000 |
| 163 | |
| 164 | def test_build_pack_1k_objects_rate(self, tmp_path: pathlib.Path) -> None: |
| 165 | """build_pack on 1 000 objects must achieve ≥ 2 000 objects/sec.""" |
| 166 | repo = _make_repo(tmp_path) |
| 167 | N = 1_000 |
| 168 | tip, blobs = _populate(repo, n_commits=50, n_unique_objects=N) |
| 169 | |
| 170 | t0 = time.perf_counter() |
| 171 | bundle = build_pack(repo, [tip]) |
| 172 | elapsed = time.perf_counter() - t0 |
| 173 | |
| 174 | assert len(bundle["objects"]) == N, ( |
| 175 | f"Expected {N} objects in bundle, got {len(bundle['objects'])}" |
| 176 | ) |
| 177 | rate = N / elapsed |
| 178 | assert rate >= self._MIN_OBJECTS_PER_SEC, ( |
| 179 | f"build_pack throughput {rate:.0f} objects/sec < {self._MIN_OBJECTS_PER_SEC} minimum. " |
| 180 | f"({N} objects took {elapsed:.2f}s.)" |
| 181 | ) |
| 182 | |
| 183 | def test_build_pack_have_filter_excludes_base( |
| 184 | self, tmp_path: pathlib.Path |
| 185 | ) -> None: |
| 186 | """build_pack with have=[base_tip] sends only delta commits, not the full history.""" |
| 187 | repo = _make_repo(tmp_path) |
| 188 | base_tip, base_blobs = _populate(repo, n_commits=50, n_unique_objects=100, start=0) |
| 189 | |
| 190 | # New commits on top of the base, with fresh objects. |
| 191 | delta_tip, delta_blobs = _populate( |
| 192 | repo, n_commits=20, n_unique_objects=50, start=1000 |
| 193 | ) |
| 194 | # Chain delta to base by writing a commit that has base_tip as parent. |
| 195 | ts = datetime.datetime(2026, 1, 2, tzinfo=datetime.timezone.utc) |
| 196 | sid = compute_snapshot_id(delta_blobs) |
| 197 | chained_cid = compute_commit_id([delta_tip, base_tip], sid, "merge", ts.isoformat()) |
| 198 | chained = CommitRecord( |
| 199 | commit_id=chained_cid, |
| 200 | repo_id="bench", |
| 201 | branch="main", |
| 202 | snapshot_id=sid, |
| 203 | message="merge", |
| 204 | committed_at=ts, |
| 205 | parent_commit_id=delta_tip, |
| 206 | parent2_commit_id=base_tip, |
| 207 | author="bench", |
| 208 | metadata={}, |
| 209 | structured_delta=None, |
| 210 | sem_ver_bump="none", |
| 211 | breaking_changes=[], |
| 212 | agent_id="", |
| 213 | model_id="", |
| 214 | toolchain_id="", |
| 215 | prompt_hash="", |
| 216 | signature="", |
| 217 | signer_key_id="", |
| 218 | ) |
| 219 | write_commit(repo, chained) |
| 220 | write_branch_ref(repo, "main", chained_cid) |
| 221 | |
| 222 | # Full pack (no have). |
| 223 | full_bundle = build_pack(repo, [chained_cid]) |
| 224 | # Delta pack: receiver already has base history. |
| 225 | delta_bundle = build_pack(repo, [chained_cid], have=[base_tip]) |
| 226 | |
| 227 | assert len(delta_bundle["commits"]) < len(full_bundle["commits"]), ( |
| 228 | "have= filter must reduce commit count" |
| 229 | ) |
| 230 | # The base objects must not be in the delta bundle since they share the snapshot. |
| 231 | full_oids = {o["object_id"] for o in full_bundle["objects"]} |
| 232 | delta_oids = {o["object_id"] for o in delta_bundle["objects"]} |
| 233 | assert delta_oids.issubset(full_oids), "delta bundle must be a subset of full bundle" |
| 234 | |
| 235 | @pytest.mark.slow |
| 236 | def test_build_pack_10k_objects_under_60s( |
| 237 | self, tmp_path: pathlib.Path |
| 238 | ) -> None: |
| 239 | """build_pack on 10 000 objects must complete in < 60 s. |
| 240 | |
| 241 | This extrapolates to 100 000 objects at the same rate: 100k / 2000 ≈ 50 s, |
| 242 | within the plan target of 60 s. |
| 243 | """ |
| 244 | repo = _make_repo(tmp_path) |
| 245 | N = 10_000 |
| 246 | tip, _ = _populate(repo, n_commits=100, n_unique_objects=N) |
| 247 | |
| 248 | t0 = time.perf_counter() |
| 249 | bundle = build_pack(repo, [tip]) |
| 250 | elapsed = time.perf_counter() - t0 |
| 251 | |
| 252 | assert len(bundle["objects"]) == N |
| 253 | assert elapsed < 60.0, ( |
| 254 | f"build_pack({N} objects) took {elapsed:.1f}s — target < 60 s. " |
| 255 | f"Rate: {N/elapsed:.0f} objects/sec." |
| 256 | ) |
| 257 | |
| 258 | |
| 259 | class TestCollectObjectIdsThroughput: |
| 260 | """collect_object_ids must be significantly faster than build_pack. |
| 261 | |
| 262 | It performs the same BFS + manifest traversal but skips reading blob bytes. |
| 263 | This is used for the pre-filter negotiation (POST /filter-objects) so the |
| 264 | client only uploads objects the remote actually needs. |
| 265 | """ |
| 266 | |
| 267 | def test_collect_object_ids_no_blob_reads( |
| 268 | self, tmp_path: pathlib.Path |
| 269 | ) -> None: |
| 270 | """collect_object_ids on 500 objects must be < 1 s (no blob reads).""" |
| 271 | repo = _make_repo(tmp_path) |
| 272 | N = 500 |
| 273 | tip, blobs = _populate(repo, n_commits=50, n_unique_objects=N) |
| 274 | |
| 275 | t0 = time.perf_counter() |
| 276 | oids = collect_object_ids(repo, [tip]) |
| 277 | elapsed = time.perf_counter() - t0 |
| 278 | |
| 279 | assert len(oids) == N, f"Expected {N} OIDs, got {len(oids)}" |
| 280 | assert elapsed < 1.0, ( |
| 281 | f"collect_object_ids({N}) took {elapsed:.3f}s — expected < 1 s. " |
| 282 | "It must not read blob bytes; only BFS + manifest traversal." |
| 283 | ) |
| 284 | |
| 285 | def test_collect_object_ids_faster_than_build_pack( |
| 286 | self, tmp_path: pathlib.Path |
| 287 | ) -> None: |
| 288 | """collect_object_ids must be faster than build_pack for the same input.""" |
| 289 | repo = _make_repo(tmp_path) |
| 290 | N = 200 |
| 291 | tip, _ = _populate(repo, n_commits=20, n_unique_objects=N) |
| 292 | |
| 293 | t_collect = time.perf_counter() |
| 294 | oids = collect_object_ids(repo, [tip]) |
| 295 | t_collect = time.perf_counter() - t_collect |
| 296 | |
| 297 | t_build = time.perf_counter() |
| 298 | bundle = build_pack(repo, [tip]) |
| 299 | t_build = time.perf_counter() - t_build |
| 300 | |
| 301 | assert len(oids) == len(bundle["objects"]) == N |
| 302 | assert t_collect <= t_build, ( |
| 303 | f"collect_object_ids ({t_collect:.3f}s) must be ≤ build_pack ({t_build:.3f}s) — " |
| 304 | "collect skips blob reads, build reads every byte." |
| 305 | ) |
| 306 | |
| 307 | def test_collect_object_ids_have_excludes_ancestors( |
| 308 | self, tmp_path: pathlib.Path |
| 309 | ) -> None: |
| 310 | """collect_object_ids with have= returns only new object IDs.""" |
| 311 | repo = _make_repo(tmp_path) |
| 312 | base_tip, base_blobs = _populate(repo, n_commits=10, n_unique_objects=50, start=0) |
| 313 | delta_tip, delta_blobs = _populate(repo, n_commits=5, n_unique_objects=30, start=100) |
| 314 | |
| 315 | all_oids = collect_object_ids(repo, [delta_tip]) |
| 316 | delta_oids = collect_object_ids(repo, [delta_tip], have=[base_tip]) |
| 317 | |
| 318 | # Delta must be a subset and contain only the 30 new objects. |
| 319 | assert set(delta_blobs.values()).issubset(set(all_oids)) |
| 320 | assert len(delta_oids) == len(set(delta_blobs.values())), ( |
| 321 | f"Expected {len(set(delta_blobs.values()))} delta OIDs, got {len(delta_oids)}" |
| 322 | ) |
| 323 | |
| 324 | |
| 325 | # --------------------------------------------------------------------------- |
| 326 | # Phase 3.4.2 — apply_pack throughput |
| 327 | # --------------------------------------------------------------------------- |
| 328 | |
| 329 | |
| 330 | class TestApplyPackThroughput: |
| 331 | """apply_pack must sustain ≥ 1 500 objects/sec in the object-write loop. |
| 332 | |
| 333 | fsync is mocked: the test measures the pack unpacking + hash-verify + |
| 334 | mkstemp + fchmod + os.replace pipeline without OS I/O latency. Durability |
| 335 | ordering is verified by test_integrity_I2_fsync.py. |
| 336 | """ |
| 337 | |
| 338 | _MIN_OBJECTS_PER_SEC: int = 1_500 |
| 339 | |
| 340 | @pytest.fixture(autouse=True) |
| 341 | def no_fsync(self) -> None: |
| 342 | """Mock out all fsync calls so the test measures algorithmic throughput.""" |
| 343 | with patch("muse.core.object_store._fsync_fd", return_value=None), \ |
| 344 | patch("muse.core.store.os.fsync", return_value=None), \ |
| 345 | patch("muse.core.store.fcntl.fcntl", return_value=0): |
| 346 | yield |
| 347 | |
| 348 | @pytest.mark.perf |
| 349 | def test_apply_pack_1k_objects_rate(self, tmp_path: pathlib.Path) -> None: |
| 350 | """apply_pack of a 1 000-object bundle must achieve ≥ _MIN_OBJECTS_PER_SEC.""" |
| 351 | src = _make_repo(tmp_path / "src") |
| 352 | N = 1_000 |
| 353 | tip, _ = _populate(src, n_commits=50, n_unique_objects=N) |
| 354 | bundle = build_pack(src, [tip]) |
| 355 | assert len(bundle["objects"]) == N |
| 356 | |
| 357 | dst = _fresh_repo(tmp_path / "dst") |
| 358 | |
| 359 | t0 = time.perf_counter() |
| 360 | result = apply_pack(dst, bundle) |
| 361 | elapsed = time.perf_counter() - t0 |
| 362 | |
| 363 | assert result["objects_written"] == N |
| 364 | rate = N / elapsed |
| 365 | assert rate >= self._MIN_OBJECTS_PER_SEC, ( |
| 366 | f"apply_pack throughput {rate:.0f} objects/sec < {self._MIN_OBJECTS_PER_SEC} minimum. " |
| 367 | f"({N} objects took {elapsed:.2f}s.)" |
| 368 | ) |
| 369 | |
| 370 | def test_apply_pack_idempotent_second_apply_skips_all( |
| 371 | self, tmp_path: pathlib.Path |
| 372 | ) -> None: |
| 373 | """Applying the same bundle twice: second apply must skip every object.""" |
| 374 | src = _make_repo(tmp_path / "src") |
| 375 | tip, _ = _populate(src, n_commits=20, n_unique_objects=100) |
| 376 | bundle = build_pack(src, [tip]) |
| 377 | |
| 378 | dst = _fresh_repo(tmp_path / "dst") |
| 379 | r1 = apply_pack(dst, bundle) |
| 380 | r2 = apply_pack(dst, bundle) |
| 381 | |
| 382 | assert r1["objects_written"] == 100 |
| 383 | assert r2["objects_written"] == 0 |
| 384 | assert r2["objects_skipped"] == 100, ( |
| 385 | f"Expected 100 skipped on second apply, got {r2['objects_skipped']}" |
| 386 | ) |
| 387 | |
| 388 | @pytest.mark.slow |
| 389 | def test_apply_pack_10k_objects_under_60s( |
| 390 | self, tmp_path: pathlib.Path |
| 391 | ) -> None: |
| 392 | """apply_pack of a 10 000-object bundle must complete in < 60 s.""" |
| 393 | src = _make_repo(tmp_path / "src") |
| 394 | N = 10_000 |
| 395 | tip, _ = _populate(src, n_commits=100, n_unique_objects=N) |
| 396 | bundle = build_pack(src, [tip]) |
| 397 | |
| 398 | dst = _fresh_repo(tmp_path / "dst") |
| 399 | |
| 400 | t0 = time.perf_counter() |
| 401 | result = apply_pack(dst, bundle) |
| 402 | elapsed = time.perf_counter() - t0 |
| 403 | |
| 404 | assert result["objects_written"] == N |
| 405 | assert elapsed < 60.0, ( |
| 406 | f"apply_pack({N} objects) took {elapsed:.1f}s — target < 60 s. " |
| 407 | f"Rate: {N/elapsed:.0f} objects/sec." |
| 408 | ) |
| 409 | |
| 410 | |
| 411 | # --------------------------------------------------------------------------- |
| 412 | # Phase 3.4.3 — verify-pack |
| 413 | # --------------------------------------------------------------------------- |
| 414 | |
| 415 | |
| 416 | class TestVerifyPackIntegrity: |
| 417 | """verify-pack must detect hash mismatches and work fast in --stat mode.""" |
| 418 | |
| 419 | def test_verify_pack_stat_returns_counts( |
| 420 | self, tmp_path: pathlib.Path |
| 421 | ) -> None: |
| 422 | """verify-pack --stat must count objects/snapshots/commits without hashing.""" |
| 423 | from tests.cli_test_helper import CliRunner |
| 424 | import json |
| 425 | |
| 426 | repo = _make_repo(tmp_path) |
| 427 | tip, _ = _populate(repo, n_commits=20, n_unique_objects=50) |
| 428 | bundle = build_pack(repo, [tip]) |
| 429 | raw = msgpack.packb(bundle, use_bin_type=True) |
| 430 | |
| 431 | bundle_file = tmp_path / "pack.muse" |
| 432 | bundle_file.write_bytes(raw) |
| 433 | (repo / ".muse" / "config.toml").write_text("") |
| 434 | |
| 435 | runner = CliRunner() |
| 436 | result = runner.invoke( |
| 437 | None, |
| 438 | [ |
| 439 | "verify-pack", |
| 440 | "--stat", |
| 441 | "--no-local", |
| 442 | "--format", "json", |
| 443 | "--file", str(bundle_file), |
| 444 | ], |
| 445 | env={"MUSE_REPO_ROOT": str(repo)}, |
| 446 | ) |
| 447 | assert result.exit_code == 0, f"verify-pack --stat failed: {result.output}" |
| 448 | payload = json.loads(result.output) |
| 449 | assert payload["objects"] == 50 |
| 450 | assert payload["commits"] == 20 |
| 451 | |
| 452 | def test_verify_pack_detects_hash_mismatch( |
| 453 | self, tmp_path: pathlib.Path |
| 454 | ) -> None: |
| 455 | """verify-pack must flag an object whose content hash does not match its ID.""" |
| 456 | from tests.cli_test_helper import CliRunner |
| 457 | import json |
| 458 | |
| 459 | repo = _make_repo(tmp_path) |
| 460 | tip, blobs = _populate(repo, n_commits=5, n_unique_objects=10) |
| 461 | bundle = build_pack(repo, [tip]) |
| 462 | |
| 463 | # Tamper: set a wrong content for the first object while keeping the declared ID. |
| 464 | tampered_obj: ObjectPayload = { |
| 465 | "object_id": bundle["objects"][0]["object_id"], |
| 466 | "content": b"TAMPERED_CONTENT", |
| 467 | } |
| 468 | tampered_bundle: PackBundle = { |
| 469 | "commits": bundle["commits"], |
| 470 | "snapshots": bundle["snapshots"], |
| 471 | "objects": [tampered_obj] + bundle["objects"][1:], |
| 472 | } |
| 473 | raw = msgpack.packb(tampered_bundle, use_bin_type=True) |
| 474 | bundle_file = tmp_path / "tampered.muse" |
| 475 | bundle_file.write_bytes(raw) |
| 476 | (repo / ".muse" / "config.toml").write_text("") |
| 477 | |
| 478 | runner = CliRunner() |
| 479 | result = runner.invoke( |
| 480 | None, |
| 481 | [ |
| 482 | "verify-pack", |
| 483 | "--no-local", |
| 484 | "--format", "json", |
| 485 | "--file", str(bundle_file), |
| 486 | ], |
| 487 | env={"MUSE_REPO_ROOT": str(repo)}, |
| 488 | ) |
| 489 | assert result.exit_code != 0, "verify-pack must exit non-zero when hash mismatches" |
| 490 | payload = json.loads(result.output) |
| 491 | assert payload["all_ok"] is False |
| 492 | assert any("hash mismatch" in f["error"] for f in payload["failures"]), ( |
| 493 | f"Expected 'hash mismatch' in failures: {payload['failures']}" |
| 494 | ) |
| 495 | |
| 496 | @pytest.mark.slow |
| 497 | def test_verify_pack_10k_objects_under_120s( |
| 498 | self, tmp_path: pathlib.Path |
| 499 | ) -> None: |
| 500 | """verify-pack of a 10 000-object bundle must complete in < 120 s. |
| 501 | |
| 502 | SHA-256 on M4 Silicon processes ~3 GiB/s; 10k × 4 KiB = 40 MiB → < 1 s. |
| 503 | The 120 s ceiling catches pathological I/O or per-object overhead. |
| 504 | """ |
| 505 | from tests.cli_test_helper import CliRunner |
| 506 | import json |
| 507 | |
| 508 | repo = _make_repo(tmp_path) |
| 509 | N = 10_000 |
| 510 | tip, _ = _populate(repo, n_commits=100, n_unique_objects=N) |
| 511 | bundle = build_pack(repo, [tip]) |
| 512 | raw = msgpack.packb(bundle, use_bin_type=True) |
| 513 | bundle_file = tmp_path / "pack10k.muse" |
| 514 | bundle_file.write_bytes(raw) |
| 515 | (repo / ".muse" / "config.toml").write_text("") |
| 516 | |
| 517 | t0 = time.perf_counter() |
| 518 | runner = CliRunner() |
| 519 | result = runner.invoke( |
| 520 | None, |
| 521 | [ |
| 522 | "verify-pack", |
| 523 | "--no-local", |
| 524 | "--format", "json", |
| 525 | "--file", str(bundle_file), |
| 526 | ], |
| 527 | env={"MUSE_REPO_ROOT": str(repo)}, |
| 528 | ) |
| 529 | elapsed = time.perf_counter() - t0 |
| 530 | |
| 531 | assert result.exit_code == 0, f"verify-pack failed: {result.output[:200]}" |
| 532 | payload = json.loads(result.output) |
| 533 | assert payload["all_ok"] is True |
| 534 | assert payload["objects_checked"] == N |
| 535 | assert elapsed < 120.0, ( |
| 536 | f"verify-pack({N} objects) took {elapsed:.1f}s — target < 120 s." |
| 537 | ) |
| 538 | |
| 539 | |
| 540 | # --------------------------------------------------------------------------- |
| 541 | # Phase 3.4.4 — cap and guard enforcement |
| 542 | # --------------------------------------------------------------------------- |
| 543 | |
| 544 | |
| 545 | class TestPackCapEnforcement: |
| 546 | """Pack-bomb and size-cap guards must fire correctly.""" |
| 547 | |
| 548 | def test_apply_pack_rejects_bundle_exceeding_max_pack_objects( |
| 549 | self, tmp_path: pathlib.Path |
| 550 | ) -> None: |
| 551 | """apply_pack raises ValueError when total_items > MAX_PACK_OBJECTS. |
| 552 | |
| 553 | MAX_PACK_OBJECTS counts commits + snapshots + objects combined — not |
| 554 | per-type. A pack with MAX_PACK_OBJECTS + 1 total items is rejected. |
| 555 | """ |
| 556 | repo = _fresh_repo(tmp_path) |
| 557 | oversized: PackBundle = { |
| 558 | "commits": [{}] * (MAX_PACK_OBJECTS + 1), |
| 559 | "snapshots": [], |
| 560 | "objects": [], |
| 561 | } |
| 562 | with pytest.raises(ValueError, match="Pack rejected"): |
| 563 | apply_pack(repo, oversized) |
| 564 | |
| 565 | def test_apply_pack_accepts_bundle_at_exact_cap( |
| 566 | self, tmp_path: pathlib.Path |
| 567 | ) -> None: |
| 568 | """apply_pack does NOT raise when total_items == MAX_PACK_OBJECTS. |
| 569 | |
| 570 | Items are malformed (empty dicts) so they are skipped as bad entries, |
| 571 | but the cap check must pass. |
| 572 | """ |
| 573 | repo = _fresh_repo(tmp_path) |
| 574 | at_cap: PackBundle = { |
| 575 | "commits": [{}] * MAX_PACK_OBJECTS, |
| 576 | "snapshots": [], |
| 577 | "objects": [], |
| 578 | } |
| 579 | # Must not raise ValueError for the cap — skips malformed entries instead. |
| 580 | result = apply_pack(repo, at_cap) |
| 581 | # Each empty-dict commit is missing commit_id and snapshot_id, so every |
| 582 | # one is skipped by the essential-field guard added to apply_pack. |
| 583 | assert result["commits_written"] == 0, ( |
| 584 | "All malformed empty-dict commits must be skipped, not written" |
| 585 | ) |
| 586 | |
| 587 | def test_apply_pack_total_items_cap_is_cross_type( |
| 588 | self, tmp_path: pathlib.Path |
| 589 | ) -> None: |
| 590 | """MAX_PACK_OBJECTS applies across commits+snapshots+objects, not per-type. |
| 591 | |
| 592 | 80 000 objects + 20 000 commits + 1 snapshot = 100 001 → rejected. |
| 593 | """ |
| 594 | repo = _fresh_repo(tmp_path) |
| 595 | cross_type: PackBundle = { |
| 596 | "commits": [{}] * 20_000, |
| 597 | "snapshots": [{}] * 1, |
| 598 | "objects": [{}] * 80_000, |
| 599 | } |
| 600 | with pytest.raises(ValueError, match="Pack rejected"): |
| 601 | apply_pack(repo, cross_type) |
| 602 | |
| 603 | def test_apply_pack_oversized_object_is_skipped_not_raised( |
| 604 | self, tmp_path: pathlib.Path |
| 605 | ) -> None: |
| 606 | """An object exceeding MAX_OBJECT_WRITE_BYTES is silently skipped. |
| 607 | |
| 608 | This is documented behaviour: the per-object cap logs a warning and |
| 609 | increments the loop counter rather than raising an exception, so the |
| 610 | rest of the bundle is still applied. |
| 611 | """ |
| 612 | repo = _fresh_repo(tmp_path) |
| 613 | good_data = b"x" * 64 |
| 614 | good_oid = _sha256(good_data) |
| 615 | oversized_oid = _sha256(b"y") # real hash — but we'll fake the size check |
| 616 | # Construct a bundle with one valid object and one whose content we |
| 617 | # claim is MAX_OBJECT_WRITE_BYTES + 1 bytes. |
| 618 | # We use a real 1-byte payload but lie about the size by patching |
| 619 | # apply_pack's check via len(raw) — we need an actually-oversized payload. |
| 620 | # Build a real oversized content string: |
| 621 | huge_data = b"z" * (MAX_OBJECT_WRITE_BYTES + 1) |
| 622 | huge_oid = _sha256(huge_data) |
| 623 | bundle: PackBundle = { |
| 624 | "commits": [], |
| 625 | "snapshots": [], |
| 626 | "objects": [ |
| 627 | ObjectPayload(object_id=good_oid, content=good_data), |
| 628 | ObjectPayload(object_id=huge_oid, content=huge_data), |
| 629 | ], |
| 630 | } |
| 631 | result = apply_pack(repo, bundle) |
| 632 | # Good object written; oversized object skipped. |
| 633 | assert result["objects_written"] == 1, ( |
| 634 | f"Expected 1 object written (the good one), got {result['objects_written']}" |
| 635 | ) |
| 636 | # Oversized object must NOT be in the store. |
| 637 | from muse.core.object_store import has_object |
| 638 | assert not has_object(repo, huge_oid), ( |
| 639 | "Oversized object must be rejected and not written to store" |
| 640 | ) |
| 641 | |
| 642 | def test_apply_pack_deduplicates_repeated_oid( |
| 643 | self, tmp_path: pathlib.Path |
| 644 | ) -> None: |
| 645 | """apply_pack writes a repeated OID only once (dedup via seen_object_ids).""" |
| 646 | repo = _fresh_repo(tmp_path) |
| 647 | data = b"deduplicate-me" * 100 |
| 648 | oid = _sha256(data) |
| 649 | REPEAT = 50 |
| 650 | bundle: PackBundle = { |
| 651 | "commits": [], |
| 652 | "snapshots": [], |
| 653 | "objects": [ObjectPayload(object_id=oid, content=data)] * REPEAT, |
| 654 | } |
| 655 | result = apply_pack(repo, bundle) |
| 656 | # First occurrence written; remaining 49 skipped. |
| 657 | assert result["objects_written"] == 1, ( |
| 658 | f"Expected 1 write for {REPEAT} identical OIDs, got {result['objects_written']}" |
| 659 | ) |
| 660 | assert result["objects_skipped"] == REPEAT - 1, ( |
| 661 | f"Expected {REPEAT - 1} skipped, got {result['objects_skipped']}" |
| 662 | ) |
| 663 | |
| 664 | def test_apply_pack_empty_bundle_is_noop( |
| 665 | self, tmp_path: pathlib.Path |
| 666 | ) -> None: |
| 667 | """apply_pack on a pack with no items returns all-zero counts.""" |
| 668 | repo = _fresh_repo(tmp_path) |
| 669 | empty: PackBundle = {"commits": [], "snapshots": [], "objects": []} |
| 670 | result = apply_pack(repo, empty) |
| 671 | assert result["commits_written"] == 0 |
| 672 | assert result["snapshots_written"] == 0 |
| 673 | assert result["objects_written"] == 0 |
| 674 | assert result["objects_skipped"] == 0 |
| 675 | |
| 676 | def test_have_equals_want_produces_empty_bundle( |
| 677 | self, tmp_path: pathlib.Path |
| 678 | ) -> None: |
| 679 | """build_pack with have=[tip] where tip is also in want returns empty bundle.""" |
| 680 | repo = _make_repo(tmp_path) |
| 681 | tip, _ = _populate(repo, n_commits=10, n_unique_objects=20) |
| 682 | |
| 683 | bundle = build_pack(repo, [tip], have=[tip]) |
| 684 | |
| 685 | assert bundle["commits"] == [], ( |
| 686 | "When have contains the want tip, BFS should yield 0 commits" |
| 687 | ) |
| 688 | assert bundle["objects"] == [], ( |
| 689 | "Empty commit set must produce empty object list" |
| 690 | ) |
| 691 | |
| 692 | |
| 693 | # --------------------------------------------------------------------------- |
| 694 | # Phase 3.4.5 — memory ceiling |
| 695 | # --------------------------------------------------------------------------- |
| 696 | |
| 697 | |
| 698 | class TestPackMemoryCeiling: |
| 699 | """build_pack and apply_pack peak memory must be proportional to blob payload. |
| 700 | |
| 701 | build_pack holds ALL object bytes in-memory simultaneously — this is a |
| 702 | known architectural property, not a bug. The test confirms: |
| 703 | 1. Peak RSS ≈ total blob bytes (not 10× or 100×). |
| 704 | 2. build_pack does not accumulate unbounded intermediate structures. |
| 705 | """ |
| 706 | |
| 707 | def test_build_pack_peak_rss_proportional_to_blob_total( |
| 708 | self, tmp_path: pathlib.Path |
| 709 | ) -> None: |
| 710 | """build_pack peak allocation is ≤ 3× the total blob payload size.""" |
| 711 | repo = _make_repo(tmp_path) |
| 712 | N = 500 |
| 713 | BLOB_SZ = 4096 # 4 KiB |
| 714 | tip, _ = _populate(repo, n_commits=50, n_unique_objects=N, blob_size=BLOB_SZ) |
| 715 | blob_total_mib = N * BLOB_SZ / (1024 * 1024) |
| 716 | |
| 717 | tracemalloc.start() |
| 718 | tracemalloc.clear_traces() |
| 719 | bundle = build_pack(repo, [tip]) |
| 720 | _, peak_bytes = tracemalloc.get_traced_memory() |
| 721 | tracemalloc.stop() |
| 722 | |
| 723 | peak_mib = peak_bytes / (1024 * 1024) |
| 724 | ceiling_mib = blob_total_mib * 3 # generous: blobs + msgpack + overhead |
| 725 | assert len(bundle["objects"]) == N |
| 726 | assert peak_mib <= ceiling_mib, ( |
| 727 | f"build_pack peak {peak_mib:.1f} MiB exceeds 3× blob total " |
| 728 | f"({ceiling_mib:.1f} MiB for {N} × {BLOB_SZ//1024} KiB objects). " |
| 729 | "build_pack must not accumulate more than the object bytes themselves." |
| 730 | ) |
| 731 | |
| 732 | def test_apply_pack_peak_rss_under_64_mib_for_small_objects( |
| 733 | self, tmp_path: pathlib.Path |
| 734 | ) -> None: |
| 735 | """apply_pack of 500 × 4 KiB objects stays under 64 MiB.""" |
| 736 | src = _make_repo(tmp_path / "src") |
| 737 | tip, _ = _populate(src, n_commits=50, n_unique_objects=500, blob_size=4096) |
| 738 | bundle = build_pack(src, [tip]) |
| 739 | |
| 740 | dst = _fresh_repo(tmp_path / "dst") |
| 741 | |
| 742 | tracemalloc.start() |
| 743 | tracemalloc.clear_traces() |
| 744 | apply_pack(dst, bundle) |
| 745 | _, peak_bytes = tracemalloc.get_traced_memory() |
| 746 | tracemalloc.stop() |
| 747 | |
| 748 | peak_mib = peak_bytes / (1024 * 1024) |
| 749 | assert peak_mib <= 64, ( |
| 750 | f"apply_pack(500 × 4 KiB) peak {peak_mib:.1f} MiB — expected ≤ 64 MiB." |
| 751 | ) |
| 752 | |
| 753 | |
| 754 | # --------------------------------------------------------------------------- |
| 755 | # Phase 3.4.6 — round-trip integrity |
| 756 | # --------------------------------------------------------------------------- |
| 757 | |
| 758 | |
| 759 | class TestPackRoundTrip: |
| 760 | """End-to-end round-trip: build_pack → msgpack serialize → apply_pack → verify.""" |
| 761 | |
| 762 | def test_roundtrip_all_objects_restored( |
| 763 | self, tmp_path: pathlib.Path |
| 764 | ) -> None: |
| 765 | """build_pack → msgpack → apply_pack round-trip: all objects readable on dst.""" |
| 766 | src = _make_repo(tmp_path / "src") |
| 767 | N_OBJECTS = 200 |
| 768 | N_COMMITS = 30 |
| 769 | tip, blobs = _populate(src, n_commits=N_COMMITS, n_unique_objects=N_OBJECTS) |
| 770 | bundle = build_pack(src, [tip]) |
| 771 | |
| 772 | # Serialize (simulates wire transfer). |
| 773 | raw = msgpack.packb(bundle, use_bin_type=True) |
| 774 | |
| 775 | # Re-hydrate using safe_unpackb (the same path as unpack-objects). |
| 776 | from muse.core.store import safe_unpackb, MAX_PACK_MSGPACK_BYTES |
| 777 | restored_dict = safe_unpackb(raw, context="roundtrip", max_bytes=MAX_PACK_MSGPACK_BYTES, allow_binary=True) |
| 778 | assert isinstance(restored_dict, dict) |
| 779 | |
| 780 | from muse.core.pack import ObjectPayload as OP, PackBundle as PB |
| 781 | raw_objects = restored_dict.get("objects") or [] |
| 782 | objects: list[OP] = [] |
| 783 | for item in raw_objects: |
| 784 | if isinstance(item, dict): |
| 785 | oid = item.get("object_id", "") |
| 786 | content = item.get("content", b"") |
| 787 | if isinstance(oid, str) and isinstance(content, (bytes, bytearray)): |
| 788 | objects.append(OP(object_id=oid, content=bytes(content))) |
| 789 | |
| 790 | hydrated: PB = { |
| 791 | "commits": [c for c in (restored_dict.get("commits") or []) if isinstance(c, dict)], |
| 792 | "snapshots": [s for s in (restored_dict.get("snapshots") or []) if isinstance(s, dict)], |
| 793 | "objects": objects, |
| 794 | } |
| 795 | dst = _fresh_repo(tmp_path / "dst") |
| 796 | result = apply_pack(dst, hydrated) |
| 797 | |
| 798 | # Every source object must be readable on the destination. |
| 799 | from muse.core.object_store import read_object, has_object |
| 800 | missing = [oid for oid in blobs.values() if not has_object(dst, oid)] |
| 801 | assert not missing, ( |
| 802 | f"{len(missing)}/{N_OBJECTS} objects missing after round-trip: " |
| 803 | f"{missing[:3]}" |
| 804 | ) |
| 805 | assert result["commits_written"] == N_COMMITS |
| 806 | assert result["objects_written"] == N_OBJECTS |
| 807 | |
| 808 | def test_roundtrip_msgpack_bundle_size_within_max_pack_bytes( |
| 809 | self, tmp_path: pathlib.Path |
| 810 | ) -> None: |
| 811 | """The serialised bundle for 1 000 × 4 KiB objects must be < MAX_PACK_MSGPACK_BYTES.""" |
| 812 | src = _make_repo(tmp_path) |
| 813 | N = 1_000 |
| 814 | tip, _ = _populate(src, n_commits=50, n_unique_objects=N, blob_size=4096) |
| 815 | bundle = build_pack(src, [tip]) |
| 816 | raw = msgpack.packb(bundle, use_bin_type=True) |
| 817 | |
| 818 | limit = MAX_PACK_MSGPACK_BYTES |
| 819 | assert len(raw) < limit, ( |
| 820 | f"Bundle for {N} × 4 KiB objects is {len(raw):,} bytes — " |
| 821 | f"exceeds MAX_PACK_MSGPACK_BYTES ({limit:,} bytes / " |
| 822 | f"{limit // 1024 // 1024} MiB)." |
| 823 | ) |