diff --git a/src/backy/backends/chunked/chunk.py b/src/backy/backends/chunked/chunk.py
index 474902c2..7939a336 100644
--- a/src/backy/backends/chunked/chunk.py
+++ b/src/backy/backends/chunked/chunk.py
@@ -51,7 +51,7 @@ def __init__(
         self.data = None
 
     def _read_existing(self) -> None:
-        if self.data is not None:
+        if self.data:
             return
         # Prepare working with the chunk. We keep the data in RAM for
         # easier random access combined with transparent compression.
@@ -82,7 +82,7 @@ def read(self, offset: int, size: int = -1) -> Tuple[bytes, int]:
         Return the data and the remaining size that should be read.
         """
         self._read_existing()
-        assert self.data is not None
+        assert self.data
 
         self.data.seek(offset)
         data = self.data.read(size)
@@ -107,7 +107,7 @@ def write(self, offset: int, data: bytes) -> Tuple[int, bytes]:
             chunk_stats["write_full"] += 1
         else:
             self._read_existing()
-            assert self.data is not None
+            assert self.data
             self.data.seek(offset)
             self.data.write(data)
             chunk_stats["write_partial"] += 1
@@ -121,7 +121,7 @@ def flush(self) -> Optional[Hash]:
         """
         if self.clean:
             return None
-        assert self.data is not None
+        assert self.data
         # I'm not using read() here to a) avoid cache accounting and b)
         # use a faster path to get the data.
         self.hash = hash(self.data.getvalue())
@@ -129,21 +129,22 @@ def flush(self) -> Optional[Hash]:
         needs_forced_write = (
             self.store.force_writes and self.hash not in self.store.seen_forced
         )
-        if self.hash not in self.store.known or needs_forced_write:
-            # Create the tempfile in the right directory to increase locality
-            # of our change - avoid renaming between multiple directories to
-            # reduce traffic on the directory nodes.
-            fd, tmpfile_name = tempfile.mkstemp(dir=target.parent)
-            posix_fadvise(fd, 0, 0, os.POSIX_FADV_DONTNEED)  # type: ignore
-            with os.fdopen(fd, mode="wb") as f:
-                data = lzo.compress(self.data.getvalue())
-                f.write(data)
-            # Micro-optimization: chmod before rename to help against
-            # metadata flushes and then changing metadata again.
-            os.chmod(tmpfile_name, 0o440)
-            os.rename(tmpfile_name, target)
-            self.store.seen_forced.add(self.hash)
-            self.store.known.add(self.hash)
+        if self.hash not in self.store.seen:
+            if needs_forced_write or not target.exists():
+                # Create the tempfile in the right directory to increase locality
+                # of our change - avoid renaming between multiple directories to
+                # reduce traffic on the directory nodes.
+                fd, tmpfile_name = tempfile.mkstemp(dir=target.parent)
+                posix_fadvise(fd, 0, 0, os.POSIX_FADV_DONTNEED)  # type: ignore
+                with os.fdopen(fd, mode="wb") as f:
+                    data = lzo.compress(self.data.getvalue())
+                    f.write(data)
+                # Micro-optimization: chmod before rename to help against
+                # metadata flushes and then changing metadata again.
+                os.chmod(tmpfile_name, 0o440)
+                os.rename(tmpfile_name, target)
+                self.store.seen_forced.add(self.hash)
+            self.store.seen.add(self.hash)
         self.clean = True
         return self.hash
 
diff --git a/src/backy/backends/chunked/store.py b/src/backy/backends/chunked/store.py
index 241fd2a5..d60151a0 100644
--- a/src/backy/backends/chunked/store.py
+++ b/src/backy/backends/chunked/store.py
@@ -23,7 +23,7 @@ class Store(object):
 
     path: Path
     seen_forced: set[Hash]
-    known: set[Hash]
+    seen: set[Hash]
     log: BoundLogger
 
     def __init__(self, path: Path, log: BoundLogger):
@@ -36,8 +36,7 @@ def __init__(self, path: Path, log: BoundLogger):
         if not self.path.joinpath("store").exists():
             self.convert_to_v2()
 
-        self.known = set(self.ls())
-        self.log.debug("init", known_chunks=len(self.known))
+        self.seen = set()
 
     def convert_to_v2(self) -> None:
         self.log.info("to-v2")
@@ -61,11 +60,12 @@ def ls(self) -> Iterable[Hash]:
     def purge(self, used_chunks: Set[Hash]) -> None:
         # This assumes exclusive lock on the store. This is guaranteed by
         # backy's main locking.
-        to_delete = self.known - used_chunks
-        self.log.info("purge", purging=len(to_delete))
-        for file_hash in sorted(to_delete):
+        self.log.info("purge")
+        for file_hash in self.ls():
+            if file_hash in used_chunks:
+                continue
             self.chunk_path(file_hash).unlink(missing_ok=True)
-        self.known -= to_delete
+            self.seen.discard(file_hash)
 
     def chunk_path(self, hash: Hash) -> Path:
         dir1 = hash[:2]
diff --git a/src/backy/daemon.py b/src/backy/daemon.py
index 40a91882..377f98c5 100644
--- a/src/backy/daemon.py
+++ b/src/backy/daemon.py
@@ -299,13 +299,13 @@ async def purge_old_files(self):
         # thread.
         while True:
             self.log.info("purge-scanning")
-            for candidate in self.base_dir.iterdir():
-                if not candidate.is_dir() or candidate.is_symlink():
+            for candidate in os.scandir(self.base_dir):
+                if not candidate.is_dir(follow_symlinks=False):
                     continue
-                self.log.debug("purge-candidate", candidate=candidate)
+                self.log.debug("purge-candidate", candidate=candidate.path)
                 reference_time = time.time() - 3 * 31 * 24 * 60 * 60
                 if not has_recent_changes(candidate, reference_time):
-                    self.log.info("purging", candidate=candidate)
+                    self.log.info("purging", candidate=candidate.path)
                     shutil.rmtree(candidate)
             self.log.info("purge-finished")
             await asyncio.sleep(24 * 60 * 60)
diff --git a/src/backy/sources/ceph/source.py b/src/backy/sources/ceph/source.py
index 97a0d4fa..82145138 100644
--- a/src/backy/sources/ceph/source.py
+++ b/src/backy/sources/ceph/source.py
@@ -90,14 +90,13 @@ def backup(self, target: BackyBackend) -> None:
             break
         self.diff(target, parent)
 
-    def diff(self, target_: BackyBackend, parent: Revision) -> None:
+    def diff(self, target: BackyBackend, parent: Revision) -> None:
         self.log.info("diff")
         snap_from = "backy-" + parent.uuid
         snap_to = "backy-" + self.revision.uuid
         s = self.rbd.export_diff(self._image_name + "@" + snap_to, snap_from)
-        t = target_.open("r+b", parent)
-        with s as source, t as target:
-            bytes = source.integrate(target, snap_from, snap_to)
+        with s as source, target.open("r+b", parent) as target_:
+            bytes = source.integrate(target_, snap_from, snap_to)
         self.log.info("diff-integration-finished")
 
         self.revision.stats["bytes_written"] = bytes
@@ -107,19 +106,18 @@ def diff(self, target_: BackyBackend, parent: Revision) -> None:
 
         self.revision.stats["chunk_stats"] = chunk_stats
 
-    def full(self, target_: BackyBackend) -> None:
+    def full(self, target: BackyBackend) -> None:
         self.log.info("full")
         s = self.rbd.export(
             "{}/{}@backy-{}".format(self.pool, self.image, self.revision.uuid)
         )
-        t = target_.open("r+b")
         copied = 0
-        with s as source, t as target:
+        with s as source, target.open("r+b") as target_:
             while True:
                 buf = source.read(4 * backy.utils.MiB)
                 if not buf:
                     break
-                target.write(buf)
+                target_.write(buf)
                 copied += len(buf)
         self.revision.stats["bytes_written"] = copied
 
@@ -128,19 +126,17 @@ def full(self, target_: BackyBackend) -> None:
 
         self.revision.stats["chunk_stats"] = chunk_stats
 
-    def verify(self, target_: BackyBackend) -> bool:
+    def verify(self, target: BackyBackend) -> bool:
         s = self.rbd.image_reader(
             "{}/{}@backy-{}".format(self.pool, self.image, self.revision.uuid)
         )
-        t = target_.open("rb")
-
         self.revision.stats["ceph-verification"] = "partial"
 
-        with s as source, t as target:
+        with s as source, target.open("rb") as target_:
             self.log.info("verify")
             return backy.utils.files_are_roughly_equal(
                 source,
-                target,
+                target_,
                 report=lambda s, t, o: self.revision.backup.quarantine.add_report(
                     QuarantineReport(s, t, o)
                 ),
diff --git a/src/backy/sources/ceph/tests/test_ceph_source.py b/src/backy/sources/ceph/tests/test_ceph_source.py
index 7e7455cb..71c34923 100644
--- a/src/backy/sources/ceph/tests/test_ceph_source.py
+++ b/src/backy/sources/ceph/tests/test_ceph_source.py
@@ -112,8 +112,8 @@ def test_choose_full_without_parent(ceph_rbd_imagesource, backup, log):
 
     revision = Revision.create(backup, set(), log)
 
-    with source(revision):
-        source.backup(revision.backend)
+    with source(revision) as s:
+        s.backup(revision.backend)
 
     assert not source.diff.called
     assert source.full.called
diff --git a/src/backy/sources/file.py b/src/backy/sources/file.py
index ed2a5a59..2312943b 100644
--- a/src/backy/sources/file.py
+++ b/src/backy/sources/file.py
@@ -41,29 +41,27 @@ def ready(self) -> bool:
             return False
         return True
 
-    def backup(self, target_: "backy.backends.BackyBackend") -> None:
+    def backup(self, target: "backy.backends.BackyBackend") -> None:
         self.log.debug("backup")
         s = open(self.filename, "rb")
         parent = self.revision.get_parent()
-        t = target_.open("r+b", parent)
-        with s as source, t as target:
+        with s as source, target.open("r+b", parent) as target_:
             if self.cow and parent:
                 self.log.info("backup-sparse")
-                bytes = copy_overwrite(source, target)
+                bytes = copy_overwrite(source, target_)
             else:
                 self.log.info("backup-full")
-                bytes = copy(source, target)
+                bytes = copy(source, target_)
 
         self.revision.stats["bytes_written"] = bytes
 
-    def verify(self, target_: "backy.backends.BackyBackend") -> bool:
+    def verify(self, target: "backy.backends.BackyBackend") -> bool:
         self.log.info("verify")
         s = open(self.filename, "rb")
-        t = target_.open("rb")
-        with s as source, t as target:
+        with s as source, target.open("rb") as target_:
             return files_are_equal(
                 source,
-                target,
+                target_,
                 report=lambda s, t, o: self.revision.backup.quarantine.add_report(
                     QuarantineReport(s, t, o)
                 ),
diff --git a/src/backy/tests/test_backup.py b/src/backy/tests/test_backup.py
index 3943aab3..e2f68b21 100644
--- a/src/backy/tests/test_backup.py
+++ b/src/backy/tests/test_backup.py
@@ -83,7 +83,7 @@ def test_backup_corrupted(simple_file_config):
     backup.backup({"daily"})
 
     store = backup.history[0].backend.store
-    chunk_path = store.chunk_path(next(iter(store.known)))
+    chunk_path = store.chunk_path(next(iter(store.seen)))
     os.chmod(chunk_path, 0o664)
     with open(chunk_path, "wb") as f:
         f.write(b"invalid")
diff --git a/src/backy/utils.py b/src/backy/utils.py
index 4b851989..ad72f406 100644
--- a/src/backy/utils.py
+++ b/src/backy/utils.py
@@ -10,7 +10,7 @@
 import tempfile
 import time
 import typing
-from pathlib import Path
+from os import DirEntry
 from typing import IO, Callable, Iterable, List, Optional, TypeVar
 from zoneinfo import ZoneInfo
 
@@ -146,7 +146,7 @@ def open_new(self, mode):
     def open_copy(self, mode):
         """Open an existing file, make a copy first, rename on close."""
         self.open_new("wb")
-        assert self.f is not None
+        assert self.f
         if os.path.exists(self.filename):
             cp_reflink(self.filename, self.f.name)
         self.f.close()
@@ -176,36 +176,36 @@ def use_write_protection(self):
 
     @property
     def name(self):
-        assert self.f is not None
+        assert self.f
         return self.f.name
 
     def read(self, *args, **kw):
-        assert self.f is not None
+        assert self.f
         data = self.f.read(*args, **kw)
         if self.encoding:
             data = data.decode(self.encoding)
         return data
 
     def write(self, data):
-        assert self.f is not None
+        assert self.f
         if self.encoding:
             data = data.encode(self.encoding)
         self.f.write(data)
 
     def seek(self, offset, whence=0):
-        assert self.f is not None
+        assert self.f
         return self.f.seek(offset, whence)
 
     def tell(self):
-        assert self.f is not None
+        assert self.f
         return self.f.tell()
 
     def truncate(self, size=None):
-        assert self.f is not None
+        assert self.f
         return self.f.truncate(size)
 
     def fileno(self):
-        assert self.f is not None
+        assert self.f
         return self.f.fileno()
 
 
@@ -440,7 +440,7 @@ def min_date():
     return datetime.datetime.min.replace(tzinfo=ZoneInfo("UTC"))
 
 
-def has_recent_changes(entry: Path, reference_time: float):
+def has_recent_changes(entry: DirEntry, reference_time: float):
     # This is not efficient on a first look as we may stat things twice, but it
     # makes the recursion easier to read and the VFS will be caching this
     # anyway.
@@ -448,17 +448,17 @@ def has_recent_changes(entry: Path, reference_time: float):
     # higher levels will propagate changed mtimes do to new/deleted files
     # instead of just modified files in our case and looking at stats when
     # traversing a directory level is faster than going depth first.
-    st = entry.stat(follow_symlinks=False)
-    if st.st_mtime >= reference_time:
-        return True
-    if not entry.is_dir() or entry.is_symlink():
+    if not entry.is_dir(follow_symlinks=False):
         return False
+    if entry.stat(follow_symlinks=False).st_mtime >= reference_time:
+        return True
+    candidates = list(os.scandir(entry.path))
     # First pass: stat all direct entries
-    for candidate in entry.iterdir():
+    for candidate in candidates:
         if candidate.stat(follow_symlinks=False).st_mtime >= reference_time:
             return True
     # Second pass: start traversing
-    for candidate in entry.iterdir():
+    for candidate in candidates:
         if has_recent_changes(candidate, reference_time):
             return True
     return False