Stage 07 · Manifest publication crash matrix

Goal

Pin the manifest rename as the linearization point by crashing immediately before and after it.

Hands-on task

Starting from stage-06, Exercise DiskStorage.persist_bucket through CrashOnce at every publication boundary. Keep all behavior inside the listed source-like boundaries; do not copy the patch first.

Deliverable files / 交付文件

• tests/test_storage.py

Self-check

1. Where is this stage's visibility or state transition owned?

   Answer

   Before rename, recovery sees the old state; after rename, it sees the complete new state.

2. Which test would fail first if the new boundary were bypassed?

   Answer

   Read tests.txt, identify the narrowest new node, and name the public call it exercises.

Pass command

uv run pytest -q $(cat journey/stages/07-publication-crash-matrix/tests.txt)

The real S3 lesson

Before rename, recovery sees the old state; after rename, it sees the complete new state.

Textbook

Chapter 5

Compare this stage on GitHub

After finishing, use git checkout stage-07 to compare your result.

Try first, then peek: stage.patch

diff --git a/tests/test_storage.py b/tests/test_storage.py
index c96a7a6..5faad97 100644
--- a/tests/test_storage.py
+++ b/tests/test_storage.py
@@ -30,3 +30,69 @@ def test_restart_restores_versions_bodies_and_counter(tmp_path: Path) -> None:

     assert reopened.get_object("b", "k", version_id=first.version_id).body == b"one"
     assert second.version_id != first.version_id
+
+
+def test_crash_before_manifest_publish_leaves_old_state(tmp_path: Path) -> None:
+    MiniS3(tmp_path).create_bucket("b")
+    crashing = MiniS3(
+        tmp_path,
+        counter=SequenceCounter(10),
+        crash_injector=CrashOnce("before_manifest_publish"),
+    )
+
+    with pytest.raises(InjectedCrash):
+        crashing.put_object("b", "new", b"value")
+
+    reopened = MiniS3(tmp_path)
+    with pytest.raises(NoSuchKey):
+        reopened.get_object("b", "new")
+    assert not list(tmp_path.rglob("*.tmp-*"))
+    assert not list(tmp_path.rglob("e00000010.*"))
+
+
+@pytest.mark.parametrize(
+    "crash_point",
+    [
+        "after_object_directory_create",
+        "after_object_directory_parent_fsync",
+        "before_manifest_publish",
+    ],
+)
+def test_crash_before_manifest_publication_matrix_leaves_old_state(
+    tmp_path: Path,
+    crash_point: str,
+) -> None:
+    MiniS3(tmp_path).create_bucket("b")
+    crashing = MiniS3(
+        tmp_path,
+        counter=SequenceCounter(10),
+        crash_injector=CrashOnce(crash_point),
+    )
+
+    with pytest.raises(InjectedCrash):
+        crashing.put_object("b", "new", b"value")
+
+    reopened = MiniS3(tmp_path)
+    with pytest.raises(NoSuchKey):
+        reopened.get_object("b", "new")
+    assert not list(tmp_path.rglob("*.tmp-*"))
+    assert not list(tmp_path.rglob("e00000010.*"))
+
+
+def test_crash_after_manifest_publish_exposes_complete_new_state(
+    tmp_path: Path,
+) -> None:
+    MiniS3(tmp_path).create_bucket("b")
+    crashing = MiniS3(
+        tmp_path,
+        counter=SequenceCounter(10),
+        crash_injector=CrashOnce("after_manifest_publish"),
+    )
+
+    with pytest.raises(InjectedCrash):
+        crashing.put_object("b", "new", b"value")
+
+    reopened = MiniS3(tmp_path)
+    visible = reopened.get_object("b", "new")
+    assert visible.body == b"value"
+    assert visible.etag == '"2063c1608d6e0baf80249c42e2be5804"'