[dependabot]

Bumps the uv group with 1 update in the /docker/pytorch/2.14/cuda directory: tornado.
Bumps the uv group with 1 update in the /docker/ray-train directory: filelock.
Bumps the uv group with 2 updates in the /docker/tensorflow/training/2.21/cpu directory: tornado and onnx.
Bumps the uv group with 2 updates in the /docker/tensorflow/training/2.21/cuda directory: tornado and onnx.
Bumps the uv group with 1 update in the /docker/xgboost directory: tornado.

Updates tornado from 6.5.10 to 6.6

Changelog

Sourced from tornado's changelog.

Release notes

.. toctree:: :maxdepth: 2

releases/v6.6.0 releases/v6.5.10 releases/v6.5.9 releases/v6.5.8 releases/v6.5.7 releases/v6.5.6 releases/v6.5.5 releases/v6.5.4 releases/v6.5.3 releases/v6.5.2 releases/v6.5.1 releases/v6.5.0 releases/v6.4.2 releases/v6.4.1 releases/v6.4.0 releases/v6.3.3 releases/v6.3.2 releases/v6.3.1 releases/v6.3.0 releases/v6.2.0 releases/v6.1.0 releases/v6.0.4 releases/v6.0.3 releases/v6.0.2 releases/v6.0.1 releases/v6.0.0 releases/v5.1.1 releases/v5.1.0 releases/v5.0.2 releases/v5.0.1 releases/v5.0.0 releases/v4.5.3 releases/v4.5.2 releases/v4.5.1 releases/v4.5.0 releases/v4.4.3 releases/v4.4.2 releases/v4.4.1 releases/v4.4.0 releases/v4.3.0 releases/v4.2.1 releases/v4.2.0 releases/v4.1.0 releases/v4.0.2

... (truncated)

Commits

• e2b0f45 Merge pull request #3831 from bdarnell/final-6.6
• 90c4e3f Version bump for 6.6 final release
• ec17353 Merge pull request #3828 from bdarnell/downstream-ci-66
• 15f66b0 ci: Add explicit read permission to CI jobs
• 4446441 ci: Run the downstream testbed on a wheel built in the same run
• c2e24d2 Merge pull request #3823 from bdarnell/claude/project-thread-uzh2s8
• c2d8884 test: Add pre-6.3 security fixes to the corpus
• de32687 test: Add a regression corpus of past security fixes
• 7848790 maint: Add pre-tag checks and release instructions
• f14321e ci: Compare benchmarks with the base branch on every PR
• Additional commits viewable in compare view

Updates tornado from 6.5.10 to 6.6

Changelog

Sourced from tornado's changelog.

Release notes

.. toctree:: :maxdepth: 2

releases/v6.6.0 releases/v6.5.10 releases/v6.5.9 releases/v6.5.8 releases/v6.5.7 releases/v6.5.6 releases/v6.5.5 releases/v6.5.4 releases/v6.5.3 releases/v6.5.2 releases/v6.5.1 releases/v6.5.0 releases/v6.4.2 releases/v6.4.1 releases/v6.4.0 releases/v6.3.3 releases/v6.3.2 releases/v6.3.1 releases/v6.3.0 releases/v6.2.0 releases/v6.1.0 releases/v6.0.4 releases/v6.0.3 releases/v6.0.2 releases/v6.0.1 releases/v6.0.0 releases/v5.1.1 releases/v5.1.0 releases/v5.0.2 releases/v5.0.1 releases/v5.0.0 releases/v4.5.3 releases/v4.5.2 releases/v4.5.1 releases/v4.5.0 releases/v4.4.3 releases/v4.4.2 releases/v4.4.1 releases/v4.4.0 releases/v4.3.0 releases/v4.2.1 releases/v4.2.0 releases/v4.1.0 releases/v4.0.2

... (truncated)

Commits

• e2b0f45 Merge pull request #3831 from bdarnell/final-6.6
• 90c4e3f Version bump for 6.6 final release
• ec17353 Merge pull request #3828 from bdarnell/downstream-ci-66
• 15f66b0 ci: Add explicit read permission to CI jobs
• 4446441 ci: Run the downstream testbed on a wheel built in the same run
• c2e24d2 Merge pull request #3823 from bdarnell/claude/project-thread-uzh2s8
• c2d8884 test: Add pre-6.3 security fixes to the corpus
• de32687 test: Add a regression corpus of past security fixes
• 7848790 maint: Add pre-tag checks and release instructions
• f14321e ci: Compare benchmarks with the base branch on every PR
• Additional commits viewable in compare view

Updates tornado from 6.5.10 to 6.6

Changelog

Sourced from tornado's changelog.

Release notes

.. toctree:: :maxdepth: 2

releases/v6.6.0 releases/v6.5.10 releases/v6.5.9 releases/v6.5.8 releases/v6.5.7 releases/v6.5.6 releases/v6.5.5 releases/v6.5.4 releases/v6.5.3 releases/v6.5.2 releases/v6.5.1 releases/v6.5.0 releases/v6.4.2 releases/v6.4.1 releases/v6.4.0 releases/v6.3.3 releases/v6.3.2 releases/v6.3.1 releases/v6.3.0 releases/v6.2.0 releases/v6.1.0 releases/v6.0.4 releases/v6.0.3 releases/v6.0.2 releases/v6.0.1 releases/v6.0.0 releases/v5.1.1 releases/v5.1.0 releases/v5.0.2 releases/v5.0.1 releases/v5.0.0 releases/v4.5.3 releases/v4.5.2 releases/v4.5.1 releases/v4.5.0 releases/v4.4.3 releases/v4.4.2 releases/v4.4.1 releases/v4.4.0 releases/v4.3.0 releases/v4.2.1 releases/v4.2.0 releases/v4.1.0 releases/v4.0.2

... (truncated)

Commits

• e2b0f45 Merge pull request #3831 from bdarnell/final-6.6
• 90c4e3f Version bump for 6.6 final release
• ec17353 Merge pull request #3828 from bdarnell/downstream-ci-66
• 15f66b0 ci: Add explicit read permission to CI jobs
• 4446441 ci: Run the downstream testbed on a wheel built in the same run
• c2e24d2 Merge pull request #3823 from bdarnell/claude/project-thread-uzh2s8
• c2d8884 test: Add pre-6.3 security fixes to the corpus
• de32687 test: Add a regression corpus of past security fixes
• 7848790 maint: Add pre-tag checks and release instructions
• f14321e ci: Compare benchmarks with the base branch on every PR
• Additional commits viewable in compare view

Updates tornado from 6.5.10 to 6.6

Changelog

Sourced from tornado's changelog.

Release notes

.. toctree:: :maxdepth: 2

releases/v6.6.0 releases/v6.5.10 releases/v6.5.9 releases/v6.5.8 releases/v6.5.7 releases/v6.5.6 releases/v6.5.5 releases/v6.5.4 releases/v6.5.3 releases/v6.5.2 releases/v6.5.1 releases/v6.5.0 releases/v6.4.2 releases/v6.4.1 releases/v6.4.0 releases/v6.3.3 releases/v6.3.2 releases/v6.3.1 releases/v6.3.0 releases/v6.2.0 releases/v6.1.0 releases/v6.0.4 releases/v6.0.3 releases/v6.0.2 releases/v6.0.1 releases/v6.0.0 releases/v5.1.1 releases/v5.1.0 releases/v5.0.2 releases/v5.0.1 releases/v5.0.0 releases/v4.5.3 releases/v4.5.2 releases/v4.5.1 releases/v4.5.0 releases/v4.4.3 releases/v4.4.2 releases/v4.4.1 releases/v4.4.0 releases/v4.3.0 releases/v4.2.1 releases/v4.2.0 releases/v4.1.0 releases/v4.0.2

... (truncated)

Commits

• e2b0f45 Merge pull request #3831 from bdarnell/final-6.6
• 90c4e3f Version bump for 6.6 final release
• ec17353 Merge pull request #3828 from bdarnell/downstream-ci-66
• 15f66b0 ci: Add explicit read permission to CI jobs
• 4446441 ci: Run the downstream testbed on a wheel built in the same run
• c2e24d2 Merge pull request #3823 from bdarnell/claude/project-thread-uzh2s8
• c2d8884 test: Add pre-6.3 security fixes to the corpus
• de32687 test: Add a regression corpus of past security fixes
• 7848790 maint: Add pre-tag checks and release instructions
• f14321e ci: Compare benchmarks with the base branch on every PR
• Additional commits viewable in compare view

Updates filelock from 3.29.7 to 4.2.1

Release notes

Sourced from filelock's releases.

4.2.1

Bug fixes

• On a singleton wrapper, AsyncReadWriteLock.close() and AsyncSoftReadWriteLock.close() end only the holds that tasks acquired through that wrapper. A reader that arrives during close() gets the closed error instead of joining the ending hold. After synchronous code closes or force-releases the shared lock, the wrappers refuse to admit readers into the ended hold, and close() on such a wrapper returns. Closing one wrapper no longer ends a task's nested hold taken through another wrapper on the same path. #832 by @​gaborbernat.
• ReadWriteLock locks again after a release() that an exception from a signal handler interrupted. The async read-write wrappers stop admitting readers into such a level. The async read-write locks forget a hold whose backend release failed after the backend let go, and they free the memory of each acquire that times out or is canceled. AsyncFileLock returns from release() after a few hundred acquires gave up behind a waiting one. An exception from a signal handler delivered while an AsyncFileLock call joins its queue no longer blocks every later call on that lock. #833 by @​gaborbernat.
• Fix ReadWriteLock acting on -journal, -wal or -shm files another process puts beside its database: a write could land in another file, a reader could get in beside a writer, and an acquire could fail. SQLite now opens the database by name only in a directory no other account can write, and through a private hard link elsewhere. On Windows the directory's access control list decides. Where hard links are unavailable, keep the database in a directory other accounts cannot write, such as one under %LOCALAPPDATA%. The lock removes link directories through a checked descriptor rather than by name, reads a link directory's name only when the locking account owns it, and leaves alone a directory with other content renamed onto a link directory's name. A database whose schema takes more than a few kilobytes to load raises ValueError instead of running past the acquire timeout. #834 by @​gaborbernat.
• Fix five bugs in opening the lock file. UnixFileLock no longer truncates a file it did not create, such as one renamed to the lock path, and no longer applies mode to it. An on_acquired hook that writes holder metadata now truncates the file itself. A lease (Linux) or an oplock (Windows) on the lock file or a soft lock's marker no longer holds acquire() past its timeout or blocking=False, and an oplock no longer holds release(); ReadWriteLock on Windows waits for its timeout and raises Timeout. On Windows, a mode without group and other bits, such as 0o600, now creates the lock file with an owner-only permission list; the ReadWriteLock database gets the same. The soft locks' writability check no longer follows a symlink swapped in at the lock path. On a filesystem without flock, a contender from another account now waits on a marker it cannot write instead of raising PermissionError. #835 by @​gaborbernat.
• Fix SoftFileLock, MarkerSoftFileLock and SoftFileLease removing a live holder's marker in three cases: the holder and the contender run in Linux time namespaces with different boottime offsets; only one of them can read the boot id; a process other than the owner rewrote the marker. Soft markers no longer carry group or other write bits. A process name containing a carriage return or another line separator no longer makes the stale check raise ValueError. When /proc/self/ns/time exists but cannot be read, the stale check no longer compares start times and keeps the marker. #836 by @​gaborbernat.
• The async read-write locks record their hold under a task owner, so synchronous code on the executor thread that took the hold can no longer re-enter or release it. AsyncSoftReadWriteLock gives back one level on release, so synchronous readers of a shared singleton keep theirs. SoftReadWriteLock tracks read levels per thread: release() from a thread that holds none raises RuntimeError. After its heartbeat recorded a lost hold, it refuses new acquires until that hold is released. #838 by @​gaborbernat.
• StrictSoftFileLock creates <path>.filelock and its claims directory with the lock's mode and sets the sticky bit (a directory flag that lets only a file's owner remove it) on both. On the next acquisition it narrows a directory it owns that an earlier release created wider; it leaves a directory of another account as it is, so remove such leftovers by hand. It reaches claims through a descriptor opened without following symlinks; on Windows it holds both directories open and refuses a junction (a directory that points at another directory). Under a sticky parent such as /tmp it refuses a coordination directory of another account, as SoftReadWriteLock does for <path>.rw. One read grace now covers a whole claim scan, a failed claim cleanup no longer leaves the sentinel descriptor open, and a non-regular node at a sentinel private-record name raises SoftFileLockProtocolError. On Windows, contenders read claims without blocking their owner's unlink. #839 by @​gaborbernat.
• Fix a lock reporting a hold the OS had dropped after a signal handler raised during acquire or release. A release that a signal handler cuts short after it has started now completes before the error propagates. A handler that acquires the lock its own thread is releasing now gets RuntimeError instead of a hold that stays held. The singleton cache no longer hands out an is_singleton=True lock or a ReadWriteLock whose finalizer is running, which released the new holder's lock. A signal during os.fork() no longer leaves the other threads blocked in their next lock call, and a fork from another thread no longer waits for a ReadWriteLock acquire on a busy database. #840 by @​gaborbernat.
• Fix a lock on a relative path locking a file in the directory the process changed to while the lock waited, and leaving its marker behind on release after such a change. The path resolved when an acquisition starts stays in use until the hold ends. An is_singleton=True lookup no longer returns a cached lock whose relative path now resolves to another file; a held lock still matches the file it holds. A rename of the lock's directory during a hold no longer leaves the marker behind or reports a false lease compromise, and a waiter no longer creates a second lock file in a directory that was renamed away. #841 by @​gaborbernat.
• SoftReadWriteLock reaches its records through descriptors of <path>.rw and gen or holders opened without following symlinks. On Windows it holds those directories open and refuses a junction (a directory that points at another directory). A tree swapped in after an earlier acquire therefore cannot redirect them. Its heartbeat rewrites a holder record only when the name still leads to the file it created, and does not block on a FIFO (a named pipe). A heartbeat that finds <path>.rw replaced by a symlink or junction reports the hold lost instead of ending the heartbeat thread. A live lock object repeats the ownership check whenever the tree it opens is not the one it prepared. #842 by @​gaborbernat.
• Fix SoftFileLock, MarkerSoftFileLock and SoftFileLease removing or displacing a live holder's marker in five cases. ext4 gave a freed inode number to the next marker. A stalled FileLock removed its ENOSYS placeholder after a peer replaced it. A failed acquire cleaned up after a successor took the path. A stale break left the lock path empty while it checked the marker it had moved. A lifetime break judged a marker by a reused inode number. On Linux and macOS a stale marker now changes hands in one atomic exchange. On other platforms without one, except Windows, including NFS and SMB, the marker stays in place and waits for timeout or break_lock(). Restoring a marker no longer follows a symlink found under the break name. A holder that releases while a peer's exchange has its marker aside leaves a record no lock parses, so the put-back lands a marker that ages out. A lease heartbeat that finds another file at the lock path once, while a peer's break has its marker aside, waits for the next refresh before it reports a compromise. A contender that finds a dead breaker's marker puts back the marker that breaker left under its break name. A marker is the breaker's own only when it names the breaker's process on this host; the same process number on another host is a holder. A contender takes a stranded marker under a name of its own before it puts the marker back, so two contenders no longer reverse each other's exchange and remove the live holder's marker. A claim left behind by a failed put-back is retried on the next poll. #843 by @​gaborbernat.
• Canceling an AsyncReadWriteLock or AsyncSoftReadWriteLock acquire that waits for another process now stops the wait and raises the cancellation. It used to wait until that process released, or raise Timeout. A thread_local=True AsyncFileLock that threads with their own event loops share no longer deadlocks when one thread waits for the lock while another holds it. #844 by @​gaborbernat.
• SoftReadWriteLock refuses a commit that lands a whole retention window behind the log, so two committers stalled at adjacent generations can no longer admit a second writer. Its stale check samples the clock before it reads a holder record, so a pause between the two no longer counts as time the record stayed unchanged. #845 by @​gaborbernat.
• Fix a SoftFileLock contender evicting an empty marker after two seconds while its owner held the lock. The owner is a SoftFileLock before 3.22, which writes nothing into its marker, or a current one still writing its record. An empty marker now reads as held, and timeout, lifetime or break_lock() clears it. A contender that finds an empty marker a dead breaker left aside puts it back instead of removing it. #846 by @​gaborbernat.
• Fix ReadWriteLock on Windows acting on -journal and -wal files another account places beside its database: an acquire could fail with OperationalError, or SQLite could replay the journal and delete the file it names. The lock now reads the database directory's access control list. Where an account other than the locking account, SYSTEM and Administrators can add or remove names, SQLite opens a hard link inside a directory only those accounts can write. A filesystem without hard links raises OSError. #854 by @​gaborbernat.
• Fix SoftReadWriteLock subclasses sharing singleton instances with the base class or sibling subclasses. Each concrete class now keeps its own cache and construction guard, including after fork, while preserving cooperative class-definition keywords. #855 by @​AryanRG.

New Contributors

• @​AryanRG made their first contribution in #855

🐍 PyPI release

📜 Full Changelog

4.2.0

Bug fixes

• Fix StrictSoftFileLock raising SoftFileLockProtocolError on Windows when a contender read a claim whose owner was still finishing its unlink. #829 by @​gaborbernat.

Performance

• import filelock is about twice as fast: lock types other than FileLock load on first use. #831 by @​gaborbernat.

🐍 PyPI release

📜 Full Changelog

4.1.2

... (truncated)

Changelog

Sourced from filelock's changelog.

########### Changelog ###########

.. towncrier-draft-entries:: Unreleased

.. towncrier release notes start

.. _release-4-2-4:

---

4.2.4 (2026-10-19)

---

.. _release-4-2-3:

---

4.2.3 (2026-10-19)

---

.. _release-4-2-3-fixed:

Fixed

• Creating a lock on a Python 3.15 interpreter without types.LazyImportType no longer triggers AttributeError at exit. :pr:861 - by :user:gaborbernat.
• Handle EFTYPE when a no-follow open refuses a symlink in a soft reader/writer lock's protocol tree, so heartbeat reports lost ownership rather than a transient refresh error. :pr:859 - by :user:edsonsims.

.. _release-4-2-2:

---

4.2.2 (2026-10-18)

---

.. _release-4-2-2-fixed:

Fixed

• Fix a lock on a relative path that stayed held after a later with lock: ended, when an earlier acquire() had failed because the working directory no longer exists. That call still raises :class:FileNotFoundError, and it now leaves the lock's hold count unchanged. :pr:856 - by :user:spicysharkco.

.. _release-4-2-1:

---

4.2.1 (2026-10-17)

---

... (truncated)

Commits

• 60331bb Release 4.2.1
• 834bf13 🧪 test(soft): skip Unix pin test on Windows
• 9d15e38 fix(soft-rw): isolate subclass singleton registries (#855)
• e1e825b 🐛 fix(open): keep lock files intact, honor deadlines and mode (#835)
• 5e3f2ec 🐛 fix(read-write): keep SQLite off shared names, bound schema (#834)
• bfa066d 🐛 fix(strict): keep claim directories private to the lock (#839)
• 977d2c3 🐛 fix(soft): wait on an empty marker instead of evicting (#846)
• 740ff66 🐛 fix(soft): break stale markers without emptying the lock path (#843)
• 86c6d5f 🐛 fix(async): honor cancellation and per-thread gates (#844)
• eeceaa8 🐛 fix(soft): keep live holders when owner identity differs (#836)
• Additional commits viewable in compare view

Updates filelock from 3.29.7 to 4.2.1

Release notes

Sourced from filelock's releases.

4.2.1

Bug fixes

• On a singleton wrapper, AsyncReadWriteLock.close() and AsyncSoftReadWriteLock.close() end only the holds that tasks acquired through that wrapper. A reader that arrives during close() gets the closed error instead of joining the ending hold. After synchronous code closes or force-releases the shared lock, the wrappers refuse to admit readers into the ended hold, and close() on such a wrapper returns. Closing one wrapper no longer ends a task's nested hold taken through another wrapper on the same path. #832 by @​gaborbernat.
• ReadWriteLock locks again after a release() that an exception from a signal handler interrupted. The async read-write wrappers stop admitting readers into such a level. The async read-write locks forget a hold whose backend release failed after the backend let go, and they free the memory of each acquire that times out or is canceled. AsyncFileLock returns from release() after a few hundred acquires gave up behind a waiting one. An exception from a signal handler delivered while an AsyncFileLock call joins its queue no longer blocks every later call on that lock. #833 by @​gaborbernat.
• Fix ReadWriteLock acting on -journal, -wal or -shm files another process puts beside its database: a write could land in another file, a reader could get in beside a writer, and an acquire could fail. SQLite now opens the database by name only in a directory no other account can write, and through a private hard link elsewhere. On Windows the directory's access control list decides. Where hard links are unavailable, keep the database in a directory other accounts cannot write, such as one under %LOCALAPPDATA%. The lock removes link directories through a checked descriptor rather than by name, reads a link directory's name only when the locking account owns it, and leaves alone a directory with other content renamed onto a link directory's name. A database whose schema takes more than a few kilobytes to load raises ValueError instead of running past the acquire timeout. #834 by @​gaborbernat.
• Fix five bugs in opening the lock file. UnixFileLock no longer truncates a file it did not create, such as one renamed to the lock path, and no longer applies mode to it. An on_acquired hook that writes holder metadata now truncates the file itself. A lease (Linux) or an oplock (Windows) on the lock file or a soft lock's marker no longer holds acquire() past its timeout or blocking=False, and an oplock no longer holds release(); ReadWriteLock on Windows waits for its timeout and raises Timeout. On Windows, a mode without group and other bits, such as 0o600, now creates the lock file with an owner-only permission list; the ReadWriteLock database gets the same. The soft locks' writability check no longer follows a symlink swapped in at the lock path. On a filesystem without flock, a contender from another account now waits on a marker it cannot write instead of raising PermissionError. #835 by @​gaborbernat.
• Fix SoftFileLock, MarkerSoftFileLock and SoftFileLease removing a live holder's marker in three cases: the holder and the contender run in Linux time namespaces with different boottime offsets; only one of them can read the boot id; a process other than the owner rewrote the marker. Soft markers no longer carry group or other write bits. A process name containing a carriage return or another line separator no longer makes the stale check raise ValueError. When /proc/self/ns/time exists but cannot be read, the stale check no longer compares start times and keeps the marker. #836 by @​gaborbernat.
• The async read-write locks record their hold under a task owner, so synchronous code on the executor thread that took the hold can no longer re-enter or release it. AsyncSoftReadWriteLock gives back one level on release, so synchronous readers of a shared singleton keep theirs. SoftReadWriteLock tracks read levels per thread: release() from a thread that holds none raises RuntimeError. After its heartbeat recorded a lost hold, it refuses new acquires until that hold is released. #838 by @​gaborbernat.
• StrictSoftFileLock creates <path>.filelock and its claims directory with the lock's mode and sets the sticky bit (a directory flag that lets only a file's owner remove it) on both. On the next acquisition it narrows a directory it owns that an earlier release created wider; it leaves a directory of another account as it is, so remove such leftovers by hand. It reaches claims through a descriptor opened without following symlinks; on Windows it holds both directories open and refuses a junction (a directory that points at another directory). Under a sticky parent such as /tmp it refuses a coordination directory of another account, as SoftReadWriteLock does for <path>.rw. One read grace now covers a whole claim scan, a failed claim cleanup no longer leaves the sentinel descriptor open, and a non-regular node at a sentinel private-record name raises SoftFileLockProtocolError. On Windows, contenders read claims without blocking their owner's unlink. #839 by @​gaborbernat.
• Fix a lock reporting a hold the OS had dropped after a signal handler raised during acquire or release. A release that a signal handler cuts short after it has started now completes before the error propagates. A handler that acquires the lock its own thread is releasing now gets RuntimeError instead of a hold that stays held. The singleton cache no longer hands out an is_singleton=True lock or a ReadWriteLock whose finalizer is running, which released the new holder's lock. A signal during os.fork() no longer leaves the other threads blocked in their next lock call, and a fork from another thread no longer waits for a ReadWriteLock acquire on a busy database. #840 by @​gaborbernat.
• Fix a lock on a relative path locking a file in the directory the process changed to while the lock waited, and leaving its marker behind on release after such a change. The path resolved when an acquisition starts stays in use until the hold ends. An is_singleton=True lookup no longer returns a cached lock whose relative path now resolves to another file; a held lock still matches the file it holds. A rename of the lock's directory during a hold no longer leaves the marker behind or reports a false lease compromise, and a waiter no longer creates a second lock file in a directory that was renamed away. #841 by @​gaborbernat.
• SoftReadWriteLock reaches its records through descriptors of <path>.rw and gen or holders opened without following symlinks. On Windows it holds those directories open and refuses a junction (a directory that points at another directory). A tree swapped in after an earlier acquire therefore cannot redirect them. Its heartbeat rewrites a holder record only when the name still leads to the file it created, and does not block on a FIFO (a named pipe). A heartbeat that finds <path>.rw replaced by a symlink or junction reports the hold lost instead of ending the heartbeat thread. A live lock object repeats the ownership check whenever the tree it opens is not the one it prepared. #842 by @​gaborbernat.
• Fix SoftFileLock, MarkerSoftFileLock and SoftFileLease removing or displacing a live holder's marker in five cases. ext4 gave a freed inode number to the next marker. A stalled FileLock removed its ENOSYS placeholder after a peer replaced it. A failed acquire cleaned up after a successor took the path. A stale break left the lock path empty while it checked the marker it had moved. A lifetime break judged a marker by a reused inode number. On Linux and macOS a stale marker now changes hands in one atomic exchange. On other platforms without one, except Windows, including NFS and SMB, the marker stays in place and waits for timeout or break_lock(). Restoring a marker no longer follows a symlink found under the break name. A holder that releases while a peer's exchange has its marker aside leaves a record no lock parses, so the put-back lands a marker that ages out. A lease heartbeat that finds another file at the lock path once, while a peer's break has its marker aside, waits for the next refresh before it reports a compromise. A contender that finds a dead breaker's marker puts back the marker that breaker left under its break name. A marker is the breaker's own only when it names the breaker's process on this host; the same process number on another host is a holder. A contender takes a stranded marker under a name of its own before it puts the marker back, so two contenders no longer reverse each other's exchange and remove the live holder's marker. A claim left behind by a failed put-back is retried on the next poll. #843 by @​gaborbernat.
• Canceling an AsyncReadWriteLock or AsyncSoftReadWriteLock acquire that waits for another process now stops the wait and raises the cancellation. It used to wait until that process released, or raise Timeout. A thread_local=True AsyncFileLock that threads with their own event loops share no longer deadlocks when one thread waits for the lock while another holds it. #844 by @​gaborbernat.
• SoftReadWriteLock refuses a commit that lands a whole retention window behind the log, so two committers stalled at adjacent generations can no longer admit a second writer. Its stale check samples the clock before it reads a holder record, so a pause between the two no longer counts as time the record stayed unchanged. #845 by @​gaborbernat.
• Fix a SoftFileLock contender evicting an empty marker after two seconds while its owner held the lock. The owner is a SoftFileLock before 3.22, which writes nothing into its marker, or a current one still writing its record. An empty marker now reads as held, and timeout, lifetime or break_lock() clears it. A contender that finds an empty marker a dead breaker left aside puts it back instead of removing it. #846 by @​gaborbernat.
• Fix ReadWriteLock on Windows acting on -journal and -wal files another account places beside its database: an acquire could fail with OperationalError, or SQLite could replay the journal and delete the file it names. The lock now reads the database directory's access control list. Where an account other than the locking account, SYSTEM and Administrators can add or remove names, SQLite opens a hard link inside a directory only those accounts can write. A filesystem without hard links raises OSError. #854 by @​gaborbernat.
• Fix SoftReadWriteLock subclasses sharing singleton instances with the base class or sibling subclasses. Each concrete class now keeps its own cache and construction guard, including after fork, while preserving cooperative class-definition keywords. #855 by @​AryanRG.

New Contributors

• @​AryanRG made their first contribution in #855

🐍 PyPI release

📜 Full Changelog

4.2.0

Bug fixes

• Fix StrictSoftFileLock raising SoftFileLockProtocolError on Windows when a contender read a claim whose owner was still finishing its unlink. #829 by @​gaborbernat.

Performance

• import filelock is about twice as fast: lock types other than FileLock load on first use. #831 by @​gaborbernat.

🐍 PyPI release

📜 Full Changelog

4.1.2

... (truncated)

Changelog

Sourced from filelock's changelog.

########### Changelog ###########

.. towncrier-draft-entries:: Unreleased

.. towncrier release notes start

.. _release-4-2-4:

---

4.2.4 (2026-10-19)

---

.. _release-4-2-3:

---

4.2.3 (2026-10-19)

---

.. _release-4-2-3-fixed:

Fixed

• Creating a lock on a Python 3.15 interpreter without types.LazyImportType no longer triggers AttributeError at exit. :pr:861 - by :user:gaborbernat.
• Handle EFTYPE when a no-follow open refuses a symlink in a soft reader/writer lock's protocol tree, so heartbeat reports lost ownership rather than a transient refresh error. :pr:859 - by :user:edsonsims.

.. _release-4-2-2:

---

4.2.2 (2026-10-18)

---

.. _release-4-2-2-fixed:

Fixed

• Fix a lock on a relative path that stayed held after a later with lock: ended, when an earlier acquire() had failed because the working directory no longer exists. That call still raises :class:FileNotFoundError, and it now leaves the lock's hold count unchanged. :pr:856 - by :user:spicysharkco.

.. _release-4-2-1:

---

4.2.1 (2026-10-17)

---

... (truncated)

Commits

• 60331bb Release 4.2.1
• 834bf13 🧪 test(soft): skip Unix pin test on Windows
• 9d15e38 fix(soft-rw): isolate subclass singleton registries (#855)
• e1e825b 🐛 fix(open): keep lock files intact, honor deadlines and mode (#835)
• 5e3f2ec 🐛 fix(read-write): keep SQLite off shared names, bound schema (#834)
• bfa066d 🐛 fix(strict): keep claim directories private to the lock (#839)
• 977d2c3 🐛 fix(soft): wait on an empty marker instead of evicting (#846)
• 740ff66 🐛 fix(soft): break stale markers without emptying the lock path (#843)
• 86c6d5f 🐛 fix(async): honor cancellation and per-thread gates (#844)
• eeceaa8 🐛 fix(soft): keep live holders when owner identity differs (#836)
• Additional commits viewable in compare view

Updates filelock from 3.29.7 to 4.2.1

Release notes

Sourced from filelock's releases.

4.2.1

Bug fixes

• On a singleton wrapper, AsyncReadWriteLock.close() and AsyncSoftReadWriteLock.close() end only the holds that tasks acquired through that wrapper. A reader that arrives during close() gets the closed error instead of joining the ending hold. After synchronous code closes or force-releases the shared lock, the wrappers refuse to admit readers into the ended hold, and close() on such a wrapper returns. Closing one wrapper no longer ends a task's nested hold taken through another wrapper on the same path. #832 by @​gaborbernat.
• ReadWriteLock locks again after a release() that an exception from a signal handler interrupted. The async read-write wrappers stop admitting readers into such a level. The async read-write locks forget a hold whose backend release failed after the backend let go, and they free the memory of each acquire that times out or is canceled. AsyncFileLock returns from release() after a few hundred acquires gave up behind a waiting one. An exception from a signal handler delivered while an AsyncFileLock call joins its queue no longer blocks every later call on that lock. #833 by @​gaborbernat.
• Fix ReadWriteLock acting on -journal, -wal or -shm files another process puts beside its database: a write could land in another file, a reader could get in beside a writer, and an acquire could fail. SQLite now opens the database by name only in a directory no other account can write, and through a private hard link elsewhere. On Windows the directory's access control list decides. Where hard links are unavailable, keep the database in a directory other accounts cannot write, such as one under %LOCALAPPDATA%. The lock removes link directories through a checked descriptor rather than by name, reads a link directory's name only when the locking account owns it, and leaves alone a directory with other content renamed onto a link directory's name. A database whose schema takes more than a few kilobytes to load raises ValueError instead of running past the acquire timeout. #834 by @​gaborbernat.
• Fix five bugs in opening the lock file. UnixFileLock no longer truncates a file it did not create, such as one renamed to the lock path, and no longer applies mode to it. An on_acquired hook that writes holder metadata now truncates the file itself. A lease (Linux) or an oplock (Windows) on the lock file or a soft lock's marker no longer holds acquire() past its timeout or blocking=False, and an oplock no longer holds release(); ReadWriteLock on Windows waits for its timeout and raises Timeout. On Windows, a mode without group and other bits, such as 0o600, now creates the lock file with an owner-only permission list; the ReadWriteLock database gets the same. The soft locks' writability check no longer follows a symlink swapped in at the lock path. On a filesystem without flock, a contender from another account now waits on a marker it cannot write instead of raising PermissionError. #835 by @​gaborbernat.
• Fix SoftFileLock, MarkerSoftFileLock and SoftFileLease removing a live holder's marker in three cases: the holder and the contender run in Linux time namespaces with different boottime offsets; only one of them can read the boot id; a process other than the owner rewrote the marker. Soft markers no longer carry group or other write bits. A process name containing a carriage return or another line separator no longer makes the stale check raise ValueError. When /proc/self/ns/time exists but cannot be read, the stale check no longer compares start times and keeps the marker. #836 by @​gaborbernat.
• The async read-write locks record their hold under a task owner, so synchronous code on the executor thread that took the hold can no longer re-enter or release it. AsyncSoftReadWriteLock gives back one level on release, so synchronous readers of a shared singleton keep theirs. SoftReadWriteLock tracks read levels per thread: release() from a thread that holds none raises RuntimeError. After its heartbeat recorded a lost hold, it refuses new acquires until that hold is released. #838 by @​gaborbernat.
• StrictSoftFileLock creates <path>.filelock and its claims directory with the lock's mode and sets the sticky bit (a directory flag that lets only a file's owner remove it) on both. On the next acquisition it narrows a directory it owns that an earlier release created wider; it leaves a directory of another account as it is, so remove such leftovers by hand. It reaches claims through a descriptor opened without following symlinks; on Windows it holds both directories open and refuses a junction (a directory that points at another directory). Under a sticky parent such as /tmp it refuses a coordination directory of another account, as SoftReadWriteLock does for <path>.rw. One read grace now covers a whole claim scan, a failed claim cleanup no longer leaves the sentinel descriptor open, and a non-regular node at a sentinel private-record name raises SoftFileLockProtocolError. On Windows, contenders read claims without blocking their owner's unlink. #839 by @​gaborbernat.
• Fix a lock reporting a hold the OS had dropped after a signal handler raised during acquire or release. A release that a signal handler cuts short after it has started now completes before the error propagates. A handler that acquires the lock its own thread is releasing now gets RuntimeError instead of a hold that stays held. The singleton cache no longer hands out an is_singleton=True lock or a ReadWriteLock whose finalizer is running, which released the new holder's lock. A signal during os.fork() no longer leaves the other threads blocked in their next lock call, and a fork from another thread no longer waits for a ReadWriteLock acquire on a busy database. #840 by @​gaborbernat.
• Fix a lock on a relative path locking a file in the directory the process changed to while the lock waited, and leaving its marker behind on release after such a change. The path resolved when an acquisition starts stays in use until the hold ends. An is_singleton=True lookup no longer returns a cached lock whose relative path now resolves to another file; a held lock still matches the file it holds. A rename of the lock's directory during a hold no longer leaves the marker behind or reports a false lease compromise, and a waiter no longer creates a second lock file in a directory that was renamed away. #841 by @​gaborbernat.
• SoftReadWriteLock reaches its records through descriptors of <path>.rw and gen or holders opened without following symlinks. On Windows it holds those directories open and refuses a junction (a directory that points at another directory). A tree swapped in after an earlier acquire therefore cannot redirect them. Its heartbeat rewrites a holder record only when the name still leads to the file it created, and does not block on a FIFO (a named pipe). A heartbeat that finds <path>.rw replaced by a symlink or junction reports the hold lost instead of ending the heartbeat thread. A live lock object repeats the ownership check whenever the tree it opens is not the one it prepared. #842 by @​gaborbernat.
• Fix SoftFileLock, MarkerSoftFileLock and SoftFileLease removing or displacing a live holder's marker in five cases. ext4 gave a freed inode number to the next marker. A stalled FileLock removed its ENOSYS placeholder after a peer replaced it. A failed acquire cleaned up after a successor took the path. A stale break left the lock path empty while it checked the marker it had moved. A lifetime break judged a marker by a reused inode number. On Linux and macOS a stale marker now changes hands in one atomic exchange. On other platforms without one, except Windows, including NFS and SMB, the marker stays in place and waits for timeout or break_lock(). Restoring a marker no longer follows a symlink found under the break name. A holder that releases while a peer's exchange has its marker aside leaves a record no lock parses, so the put-back lands a marker that ages out. A lease heartbeat that finds another file at the lock path once, while a peer's break has its marker aside, waits for the next refresh before it reports a compromise. A contender that finds a dead breaker's marker puts back the marker that breaker left under its break name. A marker is the breaker's own only when it names the breaker's process on this host; the same process number on another host is a holder. A contender takes a stranded marker under a name of its own before it puts the marker back, so two contenders no longer reverse each other's exchange and remove the live holder's marker. A claim left behind by a failed put-back is retried on the next poll. #843 by @​gaborbernat.
• Canceling an AsyncReadWriteLock or AsyncSoftReadWriteLock acquire that waits for another process now stops the wait and raises the cancellation. It used to wait until that process released, or raise Description has been truncated