| 1 | """Robustness oracles: fault-fired accounting, torn worktrees, actionability, disk fill, row validation."""
|
| 2 |
|
| 3 | from __future__ import annotations
|
| 4 |
|
| 5 | import json
|
| 6 | import os
|
| 7 | import platform
|
| 8 | import shutil
|
| 9 | import subprocess
|
| 10 | import sys
|
| 11 | import tempfile
|
| 12 | import unittest
|
| 13 | from pathlib import Path
|
| 14 |
|
| 15 | ROOT = Path(__file__).resolve().parents[1]
|
| 16 | sys.path.insert(0, str(ROOT / "suites" / "robustness"))
|
| 17 | sys.path.insert(0, str(ROOT / "scripts"))
|
| 18 |
|
| 19 | import run_robustness # noqa: E402
|
| 20 | from robustness import FAULT_NOT_FIRED, HARNESS_ERROR, HELD, VIOLATED, crashpoints, known_bugs # noqa: E402
|
| 21 | from robustness.concurrency import reported_pushed_head # noqa: E402
|
| 22 | from robustness.crash import run_time_point, run_trial, time_fault # noqa: E402
|
| 23 | from robustness.engine import actionable, classify, torn_state, torn_verdict, torn_worktree # noqa: E402
|
| 24 | from robustness.fixtures import build_git # noqa: E402
|
| 25 | from robustness.procs import Result, run_killed_after # noqa: E402
|
| 26 | from robustness.readers import worktree_manifest # noqa: E402
|
| 27 |
|
| 28 | from oakbench.rows import SKIP_RETURNCODE, validate_rows # noqa: E402
|
| 29 | from oakbench.subjects import which_measured_git # noqa: E402
|
| 30 |
|
| 31 | GIT = which_measured_git()
|
| 32 |
|
| 33 |
|
| 34 | def result(rc, stdout="", stderr=""):
|
| 35 | return Result(["x"], rc, stdout, stderr, 1.0)
|
| 36 |
|
| 37 |
|
| 38 | def f(tag):
|
| 39 | return {"kind": "file", "size": len(tag), "sha256": tag, "executable": False}
|
| 40 |
|
| 41 |
|
| 42 | class FaultFiredTests(unittest.TestCase):
|
| 43 | def test_not_fired_is_never_held(self):
|
| 44 | self.assertEqual(classify({"a": True}, fault_fired=False), FAULT_NOT_FIRED)
|
| 45 | self.assertEqual(classify({"a": False}, fault_fired=False), FAULT_NOT_FIRED)
|
| 46 | self.assertEqual(classify({"a": True}, fault_fired=True), HELD)
|
| 47 | self.assertEqual(classify({"a": True}, fault_fired=None), HELD)
|
| 48 | self.assertEqual(classify({"a": True}, harness_error="x", fault_fired=False), HARNESS_ERROR)
|
| 49 |
|
| 50 | def test_reader_exception_is_harness_error_not_violation(self):
|
| 51 | self.assertEqual(classify({"a": None}, reader_errors=["reader error: boom"]), HARNESS_ERROR)
|
| 52 | # an independent subject failure still wins
|
| 53 | self.assertEqual(classify({"repo_opens": False, "a": None}, reader_errors=["x"]), VIOLATED)
|
| 54 |
|
| 55 | def test_process_that_exits_before_the_kill_is_not_fired(self):
|
| 56 | with tempfile.TemporaryDirectory() as temp:
|
| 57 | res = run_killed_after(["/bin/sh", "-c", "exit 0"], Path(temp), dict(os.environ), 2.0)
|
| 58 | self.assertEqual(res.returncode, 0)
|
| 59 |
|
| 60 | class Stub:
|
| 61 | def op_argv(self):
|
| 62 | return ["/bin/sh", "-c", "exit 0"]
|
| 63 |
|
| 64 | def op_cwd(self):
|
| 65 | return Path(temp)
|
| 66 |
|
| 67 | class adapter: # noqa: N801
|
| 68 | @staticmethod
|
| 69 | def env():
|
| 70 | return dict(os.environ)
|
| 71 |
|
| 72 | _, fired, point = time_fault(1.0)(Stub())
|
| 73 | self.assertFalse(fired)
|
| 74 | self.assertEqual(point, "t=1000.0ms")
|
| 75 |
|
| 76 | def test_process_killed_mid_run_is_fired(self):
|
| 77 | with tempfile.TemporaryDirectory() as temp:
|
| 78 | res = run_killed_after(["/bin/sleep", "5"], Path(temp), dict(os.environ), 0.05)
|
| 79 | self.assertEqual(res.returncode, -9)
|
| 80 | self.assertTrue(res.killed_by_harness)
|
| 81 |
|
| 82 |
|
| 83 | @unittest.skipUnless(GIT and shutil.which(GIT), "real git required")
|
| 84 | class NotFiredEndToEndTests(unittest.TestCase):
|
| 85 | def test_git_commit_that_finishes_before_the_kill_is_fault_not_fired(self):
|
| 86 | with tempfile.TemporaryDirectory() as temp:
|
| 87 | temp = Path(temp)
|
| 88 | fx = build_git(GIT, "commit", temp / "fx")
|
| 89 | from robustness.crash import calibrate
|
| 90 |
|
| 91 | calib = calibrate(fx, GIT, temp / "trials", None)
|
| 92 | self.assertTrue(calib.ok, calib.detail)
|
| 93 | row = run_trial(fx, GIT, temp / "trials", "late", calib, time_fault(60.0),
|
| 94 | lambda trial, c: shutil.rmtree(trial.dir, ignore_errors=True))
|
| 95 | self.assertIs(row["fault_fired"], False)
|
| 96 | self.assertEqual(row["classification"], FAULT_NOT_FIRED)
|
| 97 | self.assertIn("did not fire", row["skip_reason"])
|
| 98 | args = run_robustness.parse_args(["--subjects", "git", "--results", str(temp / "r"), "--workdir",
|
| 99 | str(temp / "w"), "--bench-id", "20260929T000000Z", "--git-bin", GIT])
|
| 100 | runner = run_robustness.Runner(args)
|
| 101 | runner.versions = {"git": "git test"}
|
| 102 | emitted = runner.emit("git", "crash_time_v1", "commit", "crash_time", row)
|
| 103 | self.assertEqual(emitted["returncode"], SKIP_RETURNCODE)
|
| 104 | self.assertTrue(emitted["skipped"])
|
| 105 | self.assertEqual(validate_rows([emitted], "robustness"), [])
|
| 106 |
|
| 107 | def test_time_point_re_aims_from_observed_duration(self):
|
| 108 | with tempfile.TemporaryDirectory() as temp:
|
| 109 | temp = Path(temp)
|
| 110 | fx = build_git(GIT, "commit", temp / "fx")
|
| 111 | from robustness.crash import calibrate
|
| 112 |
|
| 113 | calib = calibrate(fx, GIT, temp / "trials", None)
|
| 114 | calib.elapsed_ms = 120000.0 # a stale calibration: the first kill lands long after the op
|
| 115 | row = run_time_point(fx, GIT, temp / "trials", "reaim", calib, 0.5,
|
| 116 | lambda trial, c: shutil.rmtree(trial.dir, ignore_errors=True))
|
| 117 | attempts = row["observations"]["time_attempts"]
|
| 118 | self.assertIs(attempts[0]["fired"], False)
|
| 119 | self.assertGreater(len(attempts), 1)
|
| 120 | self.assertLess(attempts[1]["delay_ms"], attempts[0]["delay_ms"])
|
| 121 | if row["fault_fired"] is False:
|
| 122 | self.assertEqual(row["classification"], FAULT_NOT_FIRED)
|
| 123 | else:
|
| 124 | self.assertIn(row["classification"], (HELD, VIOLATED))
|
| 125 |
|
| 126 |
|
| 127 | class TornWorktreeTests(unittest.TestCase):
|
| 128 | PRE = {"a": f("a0"), "b": f("b0"), "c": f("c0")}
|
| 129 | POST = {"a": f("a1"), "b": f("b1"), "c": f("c0"), "n": f("n1")}
|
| 130 |
|
| 131 | def test_split(self):
|
| 132 | wt = {"a": f("a1"), "b": f("b0"), "c": f("c0"), "a.tmp-1": f("x")}
|
| 133 | split = torn_worktree(wt, self.PRE, self.POST)
|
| 134 | self.assertEqual(split["op_output"], ["a"])
|
| 135 | self.assertEqual(split["pre_side"], ["b", "n"])
|
| 136 | self.assertEqual(split["stray"], ["a.tmp-1"])
|
| 137 |
|
| 138 | def test_states(self):
|
| 139 | pre_split = torn_worktree(dict(self.PRE), self.PRE, self.POST)
|
| 140 | post_split = torn_worktree(dict(self.POST), self.PRE, self.POST)
|
| 141 | self.assertFalse(torn_state(pre_split, "pre"))
|
| 142 | self.assertFalse(torn_state(post_split, "post"))
|
| 143 | # every file already on the target side while the ref still names the old side
|
| 144 | self.assertTrue(torn_state(post_split, "pre"))
|
| 145 | # ref moved but the worktree lags
|
| 146 | self.assertTrue(torn_state(pre_split, "post"))
|
| 147 | mixed = torn_worktree({"a": f("a1"), "b": f("b0"), "c": f("c0")}, self.PRE, self.POST)
|
| 148 | self.assertTrue(torn_state(mixed, "pre"))
|
| 149 | self.assertTrue(torn_state(mixed, None))
|
| 150 | # user edits to files the op does not touch are not the op's output
|
| 151 | user = torn_worktree({**self.PRE, "c": f("mine")}, self.PRE, self.POST)
|
| 152 | self.assertFalse(torn_state(user, "pre"))
|
| 153 |
|
| 154 | def _verdict(self, changes, in_progress, head="old"):
|
| 155 | class Adapter:
|
| 156 | def head(self, repo):
|
| 157 | return head
|
| 158 |
|
| 159 | def status_state(self, repo):
|
| 160 | return changes, in_progress, result(0 if changes is not None else 4)
|
| 161 |
|
| 162 | class T:
|
| 163 | adapter = Adapter()
|
| 164 |
|
| 165 | from unittest import mock
|
| 166 | with mock.patch("robustness.readers.worktree_manifest",
|
| 167 | return_value={"a": f("a1"), "b": f("b0"), "c": f("c0")}):
|
| 168 | return torn_verdict(T(), Path("."), self.PRE, self.POST, "old")
|
| 169 |
|
| 170 | def test_partial_output_presented_as_user_changes_fails(self):
|
| 171 | ok, detail = self._verdict(True, False)
|
| 172 | self.assertIs(ok, False)
|
| 173 | self.assertEqual(detail["torn"], "PARTIAL_OP_OUTPUT_AS_USER_CHANGES")
|
| 174 | self.assertEqual(detail["ref_side"], "pre")
|
| 175 |
|
| 176 | def test_flagged_in_progress_or_refused_passes(self):
|
| 177 | self.assertIs(self._verdict(True, True)[0], True)
|
| 178 | self.assertIs(self._verdict(None, False)[0], True)
|
| 179 |
|
| 180 |
|
| 181 | @unittest.skipUnless(GIT and shutil.which(GIT), "real git required")
|
| 182 | class TornWorktreeRealGitTests(unittest.TestCase):
|
| 183 | def test_half_switched_git_worktree_is_torn_but_merge_head_is_not(self):
|
| 184 | with tempfile.TemporaryDirectory() as temp:
|
| 185 | temp = Path(temp)
|
| 186 | fx = build_git(GIT, "switch", temp / "fx")
|
| 187 | client = fx.pristine / "client"
|
| 188 | pre = worktree_manifest(client)
|
| 189 | from robustness.engine import Trial
|
| 190 |
|
| 191 | trial = Trial(fx, GIT, temp / "trials", "torn")
|
| 192 | trial.adapter.must([GIT, "switch", "-q", fx.meta["target"]], trial.client)
|
| 193 | post = worktree_manifest(trial.client)
|
| 194 | trial.adapter.must([GIT, "switch", "-q", fx.branch], trial.client)
|
| 195 | # simulate a crash mid-checkout: some files already hold the target side, HEAD is old
|
| 196 | for name in sorted(p for p in post if pre.get(p) != post.get(p))[:5]:
|
| 197 | src = subprocess.run([GIT, "show", f"{fx.meta['target']}:{name}"], cwd=trial.client,
|
| 198 | capture_output=True, check=True).stdout
|
| 199 | (trial.client / name).parent.mkdir(parents=True, exist_ok=True)
|
| 200 | (trial.client / name).write_bytes(src)
|
| 201 | ok, detail = torn_verdict(trial, trial.client, pre, post, fx.meta["pre_head"])
|
| 202 | self.assertIs(ok, False, detail)
|
| 203 | (trial.client / ".git" / "MERGE_HEAD").write_text(fx.meta["pre_head"] + "\n")
|
| 204 | ok, detail = torn_verdict(trial, trial.client, pre, post, fx.meta["pre_head"])
|
| 205 | self.assertIs(ok, True, detail)
|
| 206 | trial.cleanup()
|
| 207 |
|
| 208 |
|
| 209 | class ActionableTests(unittest.TestCase):
|
| 210 | def test_corruption_internal_and_panic_are_not_actionable(self):
|
| 211 | for text in ("error: failed to open oak database: file is not a database",
|
| 212 | "error: internal error in oak core",
|
| 213 | "error: push rejected: corrupted object",
|
| 214 | "fatal: not a git repository (or any of the parent directories): .git",
|
| 215 | "error: Database error: database disk image is malformed",
|
| 216 | "thread 'main' panicked at src/lib.rs:9:5",
|
| 217 | "error: internal error; please try again"):
|
| 218 | self.assertFalse(actionable(result(1, stderr=text)), text)
|
| 219 | internal = {"error": {"code": "internal", "message": "store malformed"}}
|
| 220 | self.assertFalse(actionable(result(1, stdout=json.dumps(internal))))
|
| 221 | generic = {"error": {"code": "error", "message": "something failed", "recommended_next_commands": []}}
|
| 222 | self.assertFalse(actionable(result(1, stdout=json.dumps(generic))))
|
| 223 |
|
| 224 | def test_recovery_commands_and_typed_codes_are_actionable(self):
|
| 225 | for text in ("error: working tree has uncommitted changes — run `oak commit` to keep them",
|
| 226 | "Remove this directory and re-run `oak clone`.",
|
| 227 | "The remote branch has commits this clone doesn't have.\nRun 'oak pull' to bring them in",
|
| 228 | "remove the file manually to continue.",
|
| 229 | "hint: Updates were rejected ... ! [rejected] main -> main (fetch first)",
|
| 230 | "error: repository locked by another process; try again"):
|
| 231 | self.assertTrue(actionable(result(4, stderr=text)), text)
|
| 232 | rec = {"error": {"code": "error", "message": "x", "recommended_next_commands": ["oak status --json"]}}
|
| 233 | self.assertTrue(actionable(result(6, stdout=json.dumps(rec))))
|
| 234 | self.assertTrue(actionable(result(5, stdout=json.dumps({"error": {"code": "merge_in_progress"}}))))
|
| 235 | self.assertFalse(actionable(result(0, stderr="run `oak pull`")))
|
| 236 |
|
| 237 |
|
| 238 | @unittest.skipUnless(platform.system() == "Darwin" and shutil.which("hdiutil"), "macOS RAM disk required")
|
| 239 | class DiskFillTests(unittest.TestCase):
|
| 240 | def test_fill_leaves_less_than_one_block_and_records_measured_free(self):
|
| 241 | from robustness.envfaults import RamDisk, fill, free_bytes
|
| 242 |
|
| 243 | with tempfile.TemporaryDirectory() as temp:
|
| 244 | mount = Path(temp) / "mnt"
|
| 245 | with RamDisk(mount, mib=8):
|
| 246 | _, free = fill(mount, 0)
|
| 247 | self.assertLess(free, 4096)
|
| 248 | self.assertEqual(free, free_bytes(mount))
|
| 249 | with self.assertRaises(OSError):
|
| 250 | with (mount / "more.bin").open("wb") as fh:
|
| 251 | fh.write(b"x" * 8192)
|
| 252 | fh.flush()
|
| 253 | os.fsync(fh.fileno())
|
| 254 | (mount / "more.bin").unlink(missing_ok=True)
|
| 255 | (mount / "p4-filler.bin").unlink()
|
| 256 | _, free = fill(mount, 64)
|
| 257 | self.assertGreaterEqual(free, 60 * 1024)
|
| 258 | self.assertLess(free, 72 * 1024)
|
| 259 |
|
| 260 |
|
| 261 | class RowValidationTests(unittest.TestCase):
|
| 262 | def _row(self, **kw):
|
| 263 | row = {"bench_id": "20260929T000000Z", "scenario": "crash_time_v1", "operation": "commit", "run": 0,
|
| 264 | "subject": "git", "subject_kind": "git", "elapsed_ms": 1.0, "returncode": 0, "classification": "held",
|
| 265 | "fault_family": "crash_time", "fault_point": "t=1ms", "fault_fired": True, "violations": []}
|
| 266 | row.update(kw)
|
| 267 | return row
|
| 268 |
|
| 269 | def test_rejects_unknown_and_inconsistent_classifications(self):
|
| 270 | self.assertEqual(validate_rows([self._row()], "robustness"), [])
|
| 271 | self.assertTrue(validate_rows([self._row(classification="bogus")], "robustness"))
|
| 272 | self.assertTrue(validate_rows([self._row(returncode=1)], "robustness"))
|
| 273 | self.assertTrue(validate_rows([self._row(classification="violated", returncode=0)], "robustness"))
|
| 274 | self.assertTrue(validate_rows([self._row(fault_fired=False)], "robustness"))
|
| 275 | self.assertTrue(validate_rows([self._row(violations=["x"])], "robustness"))
|
| 276 | not_fired = self._row(classification="fault_not_fired", returncode=77, fault_fired=False, skipped=True,
|
| 277 | skip_reason="op exited before the kill")
|
| 278 | self.assertEqual(validate_rows([not_fired], "robustness"), [])
|
| 279 | self.assertTrue(validate_rows([{**not_fired, "returncode": 0, "skipped": False}], "robustness"))
|
| 280 |
|
| 281 |
|
| 282 | class MiscTests(unittest.TestCase):
|
| 283 | def test_known_bug_tags(self):
|
| 284 | self.assertEqual(known_bugs("oak", "crash_time_v1", "pull_ff", ["worktree_not_torn"]), ["fb-535"])
|
| 285 | self.assertEqual(known_bugs("oak", "crash_syscall_v1", "switch", ["worktree_not_torn", "x"]), ["fb-535"])
|
| 286 | self.assertEqual(known_bugs("oak", "conc_same_branch_v1", "concurrent.push",
|
| 287 | ["no_lost_acknowledged_write", "linearizable"]), ["fb-534"])
|
| 288 | self.assertEqual(known_bugs("git", "crash_time_v1", "switch", ["worktree_not_torn"]), [])
|
| 289 | self.assertEqual(known_bugs("oak", "crash_time_v1", "commit", ["repo_opens"]), [])
|
| 290 |
|
| 291 | def test_no_boundary_reasons_are_subject_specific(self):
|
| 292 | self.assertNotIn("git", crashpoints.NO_BOUNDARY_REASON["oak"])
|
| 293 | self.assertIn("git", crashpoints.NO_BOUNDARY_REASON["git"])
|
| 294 | inst = crashpoints.CrashInstrument(Path("x"), None, Path("g"), None, {}, {"oak": "codesign failed"})
|
| 295 | self.assertEqual(crashpoints.unavailable_reason(inst, "", "oak"), "codesign failed")
|
| 296 | self.assertEqual(crashpoints.unavailable_reason(inst, "", "git"), "")
|
| 297 | self.assertEqual(crashpoints.unavailable_reason(None, "no clang", "git"), "no clang")
|
| 298 |
|
| 299 | def test_reported_pushed_head(self):
|
| 300 | self.assertEqual(reported_pushed_head(result(0, stdout=json.dumps({"pushed": True, "pushed_head": "abc"}))),
|
| 301 | "abc")
|
| 302 | self.assertIsNone(reported_pushed_head(result(0, stdout="plain text")))
|
| 303 |
|
| 304 |
|
| 305 | if __name__ == "__main__":
|
| 306 | unittest.main()
|