Collect orphaned snapshots by name: the sidecar lives in tmpfs

A reboot mid-backup orphaned the entire tree, permanently. The sidecar is the record
of which snapshots a run pinned -- and /run is tmpfs. A reboot or crash between the
recursive snapshot and its cleanup destroyed that record, leaving one snapshot per
descendant dataset (250+ on a real pool) with nothing pointing at them. Nothing would
ever have found them.

gc_stale_snapshots() identifies leftovers by NAME, so it works when the record is
gone. It runs after the sidecar reclaim -- the recorded path stays authoritative and
the collector only mops up what the record lost.

It deletes data on a name match, which is a weaker claim than a recorded fact, so the
selection is a pure function with the harshest tests here. A snapshot is collected
only if the name is exactly <dataset>@<task>-<YYYYMMDDHHMMSS>, it is not the current
run's, NOTHING IS MOUNTED FROM IT (this, not the age guard, is what protects a
concurrent backup), and it is over an hour old.

Checked against the real pool: of 4728 snapshots including 2341 periodic ones, it
selects exactly the orphans of the task being run and nothing else.

CHANGELOG.md

@@ -207,7 +207,35 @@ worse than no alert, because one day it carries a security fix.
   warning, not a refusal: declining to install over a string we failed to read would
   be a worse failure than the one being prevented.
 
-- **A stable release may not leave work stranded under `## Unreleased`.** Either it
+- **A stable release may not leave work stranded under `## Unreleased
+
+### Fixed
+
+- **A reboot mid-backup orphaned the entire snapshot tree, permanently.** The sidecar
+  is the record of which snapshots a run pinned — and it lives in `/run`, which is
+  **tmpfs**. A reboot (or a crash) between taking the recursive snapshot and cleaning
+  it up destroyed that record, leaving one snapshot per descendant dataset — **250+ on
+  a real pool** — with nothing left pointing at them. Nothing would ever have found
+  them again.
+
+  `gc_stale_snapshots()` is the backstop: it identifies leftovers **by name**, so it
+  works when the record is gone. It runs at the start of every backup, after the
+  sidecar reclaim — the recorded path stays authoritative, and the collector only ever
+  mops up what the record lost.
+
+  Because it deletes data on a *name match* — a weaker claim than a recorded fact — the
+  selection is a **pure function** with the harshest tests in the suite. A snapshot is
+  collected only if **all** of these hold:
+
+  | | |
+  | --- | --- |
+  | name is exactly `<dataset>@<task>-<YYYYMMDDHHMMSS>` | so `cloud_backup-5` never matches `cloud_backup-50`, an `auto-*` periodic snapshot, or anything a human made |
+  | it is not the current run's | parent *and* children are excluded |
+  | **nothing is mounted from it** | an in-flight run pins its own snapshots — this, not the age guard, is what protects a concurrent backup |
+  | it is **over an hour old** | covers the seconds-long window where a live run has snapshotted but not yet mounted |
+
+  Verified against the real pool: of **4,728** snapshots — including **2,341** periodic
+  ones — it selects exactly the orphans of the task being run, and nothing else.`.** Either it
   is finished and belongs in the release, or the release is premature. Candidates
   are exempt: an rc may legitimately have work queued behind it.
 

docs/nested-snapshots.md

@@ -95,6 +95,21 @@ re-scan each time.
 
 ### Snapshot lifecycle
 
+> **Two mechanisms clean up, and the second exists because the first can be destroyed.**
+>
+> 1. **The sidecar** records exactly which snapshots a run pinned, and is removed only
+>    on a confirmed-clean sweep. Precise, and it survives a middlewared restart.
+> 2. **The garbage collector** finds leftovers by *name*, so it still works when the
+>    sidecar is gone — and it can be: **the sidecar lives in `/run`, which is tmpfs.** A
+>    reboot mid-backup takes it, and with it the only record of a 250-snapshot tree.
+>
+> The collector runs at the start of every backup, after the sidecar reclaim. It will
+> only touch a snapshot named `<dataset>@<task>-<timestamp>` that is not the current
+> run's, has **nothing mounted from it** (which is what protects a concurrently-running
+> backup), and is **over an hour old**. Periodic `auto-*` snapshots, other tasks'
+> snapshots, and anything you made by hand are structurally out of reach.
+
+
 > **A snapshot may survive a run, and that is expected.** ZFS **automounts**
 > `<dataset>/.zfs/snapshot/<snap>` the moment it is read, and holds it for
 > `zfs_expire_snapshot` seconds (**300** by default) after the last access. So

patch/truecloud_nested.py

@@ -55,6 +55,7 @@ Therefore this module owns the whole lifecycle:
 from __future__ import annotations
 
 import contextlib
+import datetime
 import os
 import stat
 import subprocess
@@ -68,10 +69,13 @@ __all__ = [
     "cleanup_task",
     "current_mounts_under",
     "delete_snapshot_tree",
+    "gc_stale_snapshots",
+    "mounted_snapshots",
     "plan_staging",
     "sidecar_for",
     "snapshot_tree_names",
     "stage_nested",
+    "stale_snapshot_names",
     "staging_root_for",
     "teardown",
     "verify_staged",
@@ -176,6 +180,76 @@ def snapshot_tree_names(snapshot: str, all_names) -> list[str]:
     ]
 
 
+#: A snapshot must be at least this old before the garbage collector will touch it.
+#:
+#: The GC identifies our leftovers by NAME, so its only real risk is deleting a
+#: snapshot belonging to a run that is still starting up -- the window between
+#: `zfs snapshot -r` and the bind mounts appearing, which is seconds. An hour is three
+#: orders of magnitude more slack than that window needs, and still reclaims a lost
+#: tree on the very next daily run.
+GC_MIN_AGE_SECONDS = 3600
+
+
+def stale_snapshot_names(task_name, current_snapshot, all_names, now,
+                         in_use=(), min_age=GC_MIN_AGE_SECONDS):
+    """Snapshots THIS task created in an earlier run and never cleaned up.
+
+    Pure, because this is the one function here that DELETES DATA on a name match, and
+    a name match is a weaker claim than a recorded fact. Everything it relies on is an
+    argument, so every way it could be wrong is a test.
+
+    Why a garbage collector exists at all, when there is already a sidecar: **the
+    sidecar lives in /run, which is tmpfs.** A reboot mid-backup destroys it, and with
+    it the only record of a 250-snapshot tree. The sidecar handles the normal case
+    precisely; this handles the case where the record itself is gone.
+
+    A snapshot is ours to collect only if ALL of these hold:
+
+      * its name is exactly ``<dataset>@<task_name>-<YYYYMMDDHHMMSS>`` -- so
+        ``cloud_backup-5`` never matches ``cloud_backup-50``'s snapshots, and never
+        matches a periodic ``auto-2026-…`` or anything a human made;
+      * it is not the snapshot the current run is using;
+      * nothing is mounted from it (`in_use`) -- an in-flight run pins its own
+        snapshots, so this alone protects a concurrent one-time backup;
+      * it is older than `min_age` -- which covers the seconds-long window in which a
+        run has taken its snapshot but not yet mounted it.
+
+    `now` is a timezone-aware datetime; timestamps in the name are UTC (stock builds
+    them with `utc_now()`).
+    """
+    prefix = task_name + "-"
+    stale = []
+
+    for name in all_names:
+        _dataset, _, snapname = name.partition("@")
+        if not snapname or not snapname.startswith(prefix):
+            continue
+        if name == current_snapshot or snapname == _snapname_of(current_snapshot):
+            continue
+        if name in in_use:
+            continue
+
+        stamp = snapname[len(prefix):]
+        try:
+            when = datetime.datetime.strptime(stamp, "%Y%m%d%H%M%S").replace(
+                tzinfo=datetime.UTC
+            )
+        except ValueError:
+            # Not our timestamp format. Something else owns this name; leave it alone.
+            continue
+
+        if (now - when).total_seconds() < min_age:
+            continue
+
+        stale.append(name)
+
+    return stale
+
+
+def _snapname_of(snapshot):
+    return snapshot.partition("@")[2] if snapshot else ""
+
+
 def _probe_snapdir(path):
     """Classify a snapshot directory: ``ok``, ``missing``, or why it is unusable.
 
@@ -598,6 +672,80 @@ def delete_snapshot_tree(middleware, snapshot, logger=None, attempts=4,
     return remaining
 
 
+def mounted_snapshots(mounts_file="/proc/self/mounts"):
+    """Every ZFS snapshot something is currently mounted from.
+
+    The device field of a snapshot mount IS the snapshot name (`Tap/apps/x@snap`), for
+    both our staging bind mounts and ZFS's own .zfs automounts. So this is a direct,
+    factual answer to "is anything using this snapshot right now" -- which is what
+    protects a concurrently-running backup from the garbage collector, rather than
+    trusting an age heuristic to be generous enough.
+    """
+    live = set()
+    try:
+        with open(mounts_file, encoding="utf-8") as fh:
+            for line in fh:
+                dev = line.split(" ", 1)[0]
+                if "@" in dev:
+                    live.add(dev.replace("\\040", " "))
+    except OSError:
+        return set()
+    return live
+
+
+def gc_stale_snapshots(middleware, task_name, current_snapshot, logger=None,
+                       now=None, mounts_file="/proc/self/mounts"):
+    """Delete snapshots this task left behind in an earlier run. Returns what remains.
+
+    The backstop for when the RECORD is gone, not just the snapshots: the sidecar lives
+    in /run (tmpfs), so a reboot mid-backup takes it with them. Without this, that tree
+    -- one snapshot per descendant dataset, 250+ on a real pool -- is orphaned with
+    nothing left pointing at it.
+
+    Selection is `stale_snapshot_names()`, which is pure and heavily tested, because a
+    name match is a weaker claim than a recorded fact and this deletes data on one.
+    """
+    dataset = current_snapshot.partition("@")[0]
+    now = now or datetime.datetime.now(datetime.UTC)
+
+    try:
+        snaps = middleware.call_sync(
+            "zfs.snapshot.query", [["name", "^", dataset]], {"select": ["name"]}
+        )
+    except Exception as e:  # noqa: BLE001 - cannot enumerate; collect nothing
+        if logger:
+            logger.warning(
+                "truecloud-patch: could not enumerate snapshots for GC: %r", e
+            )
+        return []
+
+    stale = stale_snapshot_names(
+        task_name, current_snapshot, [s["name"] for s in snaps], now,
+        in_use=mounted_snapshots(mounts_file),
+    )
+    if not stale:
+        return []
+
+    if logger:
+        logger.warning(
+            "truecloud-patch: %d snapshot(s) from an earlier run of %s were never "
+            "cleaned up (a lost record, e.g. a reboot mid-backup); collecting them",
+            len(stale), task_name,
+        )
+
+    remaining = []
+    for name in stale:
+        try:
+            middleware.call_sync("zfs.snapshot.delete", name)
+        except Exception as e:  # noqa: BLE001 - busy, or gone; either way, next run
+            remaining.append(name)
+            if logger:
+                logger.debug(
+                    "truecloud-patch: could not collect %s: %r", name, e
+                )
+    return remaining
+
+
 def stage_nested(middleware, path, snapshot, base_dataset, base_mountpoint,
                  task_name, datasets, logger=None):
     """Build a complete staging tree for `path` from the already-taken `snapshot`.
@@ -650,6 +798,18 @@ def stage_nested(middleware, path, snapshot, base_dataset, base_mountpoint,
             "carrying them forward to the next run", len(pending),
         )
 
+    # ...and collect anything from an earlier run that has NO record at all.
+    #
+    # The sidecar above is precise but lives in /run, which is tmpfs -- a reboot
+    # mid-backup destroys it and orphans the whole tree with nothing pointing at it.
+    # This finds those by name and is the only thing that ever will.
+    #
+    # It runs AFTER the sidecar reclaim on purpose: the recorded path is authoritative
+    # and cheap, and the GC should only ever be mopping up what the record lost.
+    pending.extend(
+        gc_stale_snapshots(middleware, task_name, snapshot, logger=logger)
+    )
+
     # Record the snapshot BEFORE mounting anything, not after. middlewared can
     # die at any point (this patch even schedules a restart at boot), and the
     # sidecar is the only thing that survives it -- an in-process dict would take

tests/test_truecloud_nested.py

@@ -893,3 +893,159 @@ class TestTheSidecarCarriesEveryPendingTree:
         assert len(notes) == 2
         assert "'Tap@a'" in notes[0] and "'Tap@b'" in notes[1]
         assert "[" not in "".join(notes)
+
+
+class TestGarbageCollectorSelection:
+    """`stale_snapshot_names` DELETES DATA on a name match.
+
+    A name match is a weaker claim than a recorded fact, so every way it could be wrong
+    is a test. It exists because the sidecar — which IS a recorded fact — lives in /run,
+    which is tmpfs: a reboot mid-backup destroys it and orphans a 250-snapshot tree with
+    nothing left pointing at it. This is the only thing that would ever find those.
+    """
+
+    import datetime as _dt
+    NOW = _dt.datetime(2026, 7, 14, 12, 0, 0, tzinfo=_dt.UTC)
+    CURRENT = "Tap@cloud_backup-5-20260714115900"        # 1 minute ago
+    OLD = "Tap/apps/x@cloud_backup-5-20260713030000"     # ~33 hours ago
+
+    def collect(self, names, **kw):
+        import truecloud_nested as tn
+        return tn.stale_snapshot_names(
+            "cloud_backup-5", self.CURRENT, names, self.NOW, **kw
+        )
+
+    def test_it_collects_our_own_leftovers(self):
+        assert self.collect([self.OLD]) == [self.OLD]
+
+    def test_it_NEVER_touches_the_current_run(self):
+        # Both the parent and its children share the current snapname.
+        names = [self.CURRENT, "Tap/apps/x@cloud_backup-5-20260714115900"]
+        assert self.collect(names) == []
+
+    def test_it_NEVER_touches_a_periodic_snapshot(self):
+        assert self.collect(["Tap/apps/x@auto-2026-07-13_03-00"]) == []
+
+    def test_it_NEVER_touches_a_human_made_snapshot(self):
+        assert self.collect(["Tap@before-i-broke-everything"]) == []
+
+    def test_it_NEVER_touches_another_TASK(self):
+        # cloud_backup-5 must not match cloud_backup-50. This is why the prefix
+        # carries the trailing dash.
+        assert self.collect(["Tap/apps/x@cloud_backup-50-20260713030000"]) == []
+        assert self.collect(["Tap/apps/x@cloud_backup-7-20260713030000"]) == []
+
+    def test_it_NEVER_touches_a_one_time_backup(self):
+        assert self.collect(["Tap@cloud_backup-onetime-20260713030000"]) == []
+
+    def test_it_NEVER_touches_a_snapshot_that_is_MOUNTED(self):
+        # An in-flight run pins its own snapshots. This — not the age heuristic — is
+        # what actually protects a concurrent backup.
+        assert self.collect([self.OLD], in_use={self.OLD}) == []
+
+    def test_it_NEVER_touches_a_snapshot_younger_than_the_minimum_age(self):
+        # Covers the seconds-long window between `zfs snapshot -r` and the mounts
+        # appearing, when a live run's snapshots look exactly like garbage.
+        young = "Tap/apps/x@cloud_backup-5-20260714113000"   # 30 minutes ago
+        assert self.collect([young]) == []
+        assert self.collect([young], min_age=60) == [young]
+
+    def test_a_name_it_cannot_parse_is_left_alone(self):
+        assert self.collect(["Tap@cloud_backup-5-not-a-timestamp"]) == []
+        assert self.collect(["Tap@cloud_backup-5-"]) == []
+
+    def test_a_realistic_mixed_pool(self):
+        names = [
+            self.CURRENT,                                       # ours, running
+            "Tap/apps/x@cloud_backup-5-20260714115900",         # ours, running (child)
+            self.OLD,                                           # ours, orphaned  <-
+            "Tap/apps/y@cloud_backup-5-20260712030000",         # ours, orphaned  <-
+            "Tap/apps/x@auto-2026-07-13_03-00",                 # periodic
+            "Tap/apps/x@cloud_backup-7-20260713030000",         # another task
+            "Tap@manual-keepme",                                # human
+        ]
+        assert sorted(self.collect(names)) == sorted(
+            [self.OLD, "Tap/apps/y@cloud_backup-5-20260712030000"]
+        )
+
+
+class TestMountedSnapshots:
+    def test_it_reads_snapshot_names_out_of_the_mount_table(self, tmp_path):
+        import truecloud_nested as tn
+        mounts = tmp_path / "mounts"
+        mounts.write_text(
+            "tmpfs /run tmpfs rw 0 0\n"
+            "Tap/apps/x@snap1 /run/truecloud-nested/t/apps/x zfs ro 0 0\n"
+            "Tap/apps/y@snap1 /mnt/Tap/apps/y/.zfs/snapshot/snap1 zfs ro 0 0\n"
+            "Tap/live /mnt/Tap/live zfs rw 0 0\n"
+        )
+        live = tn.mounted_snapshots(str(mounts))
+        assert live == {"Tap/apps/x@snap1", "Tap/apps/y@snap1"}
+        assert "Tap/live" not in live      # a live dataset is not a snapshot
+
+
+class TestGarbageCollectorExecution:
+    def test_it_deletes_the_stale_ones_and_nothing_else(self, monkeypatch, tmp_path):
+        import datetime as dt
+        import truecloud_nested as tn
+
+        mounts = tmp_path / "mounts"
+        mounts.write_text("")
+        now = dt.datetime(2026, 7, 14, 12, 0, 0, tzinfo=dt.UTC)
+
+        mw = FakeMiddleware([
+            "Tap@cloud_backup-5-20260714115900",             # current run
+            "Tap/apps/x@cloud_backup-5-20260713030000",      # orphan   <-
+            "Tap/apps/x@auto-2026-07-13_03-00",              # periodic
+            "Tap/apps/x@cloud_backup-7-20260713030000",      # other task
+        ])
+        remaining = tn.gc_stale_snapshots(
+            mw, "cloud_backup-5", "Tap@cloud_backup-5-20260714115900",
+            now=now, mounts_file=str(mounts),
+        )
+        assert remaining == []
+        assert mw.snapshots == [
+            "Tap@cloud_backup-5-20260714115900",
+            "Tap/apps/x@auto-2026-07-13_03-00",
+            "Tap/apps/x@cloud_backup-7-20260713030000",
+        ]
+
+    def test_a_busy_orphan_is_reported_not_swallowed(self, monkeypatch, tmp_path):
+        import datetime as dt
+        import truecloud_nested as tn
+
+        mounts = tmp_path / "mounts"
+        mounts.write_text("")
+        now = dt.datetime(2026, 7, 14, 12, 0, 0, tzinfo=dt.UTC)
+        orphan = "Tap/apps/x@cloud_backup-5-20260713030000"
+
+        mw = BusyMiddleware(
+            ["Tap@cloud_backup-5-20260714115900", orphan],
+            busy=[orphan], busy_for=99,
+        )
+        remaining = tn.gc_stale_snapshots(
+            mw, "cloud_backup-5", "Tap@cloud_backup-5-20260714115900",
+            now=now, mounts_file=str(mounts),
+        )
+        assert remaining == [orphan]
+
+    def test_it_collects_NOTHING_when_the_query_fails(self, tmp_path):
+        # Cannot enumerate => cannot know what is ours => delete nothing.
+        import datetime as dt
+        import truecloud_nested as tn
+
+        mounts = tmp_path / "mounts"
+        mounts.write_text("")
+
+        class Broken(FakeMiddleware):
+            def call_sync(self, method, *args):
+                if method == "zfs.snapshot.query":
+                    raise RuntimeError("middleware is having a day")
+                return super().call_sync(method, *args)
+
+        assert tn.gc_stale_snapshots(
+            Broken(["Tap/apps/x@cloud_backup-5-20260713030000"]),
+            "cloud_backup-5", "Tap@cloud_backup-5-20260714115900",
+            now=dt.datetime(2026, 7, 14, 12, 0, 0, tzinfo=dt.UTC),
+            mounts_file=str(mounts),
+        ) == []