v0.10.6
Released 2026-11-17Changed
archil mount --conditional --forcenow takes over a disk from a client that disconnected or stopped responding while holding it, instead of waiting for that client to time out; writes that client had not yet uploaded are lost, even ones already reported as successful.
Fixed
- Creating a hard link to a file you cannot write, such as a read-only file you own, no longer fails with “Permission denied”.
- After a brief network disruption, a mount now reconnects to its server without first looking up the server’s address again, and no longer drops a healthy connection, along with every request waiting on it, when one request times out before it is sent.
- Removing an extended attribute that a file does not have now fails with “No such attribute” instead of succeeding silently.
- In a directory with the sticky bit set (such as a shared
/tmp-style directory), users can no longer delete, rename, or replace files owned by other users; this now fails with “Operation not permitted”, as on a local Linux filesystem.
v0.10.5
Released 2026-11-12Fixed
- On Linux, opening a file another client just resized could still use the old size and overwrite appended data. Opens now fail with “Stale file handle” while the size keeps changing.
v0.10.4
Released 2026-11-11Fixed
archil mountno longer hangs before the mount becomes usable on Linux hosts with audit rules loaded, which Daytona sandboxes began using on November 11, 2026.
v0.10.3
Released 2026-11-10Fixed
- Reads that bypass the page cache, such as with
O_DIRECT, now stop at the end of the file instead of returning zeros, so programs reading to the end no longer hang. - Reopening a file you wrote no longer returns old data when another client changed it while you had it open.
- On Linux, opening a file that another client just resized no longer sees the old size, which could cut reads short or crash writes through a memory mapping with SIGBUS.
v0.10.2
Released 2026-11-09Changed
- On
--sharedand--conditionalmounts and in the macOS app, changes from other clients now appear almost immediately instead of after cached data expires.
Fixed
- Disks now mount on IPv6-only hosts that reach IPv4 through NAT64, such as some sandbox providers. Previously file access hung after a successful mount.
- Mounts using a one-use disk token no longer fail to reconnect more than about five minutes after mounting.
- When clients share a disk, appending to a file another client just appended to no longer overwrites that client’s data.
- Opening or creating files in a directory another client moved no longer intermittently fails with “Stale file handle”.
- Writing to a disk that other clients just unmounted fails less often with “Resource temporarily unavailable”.
archil invalidate-cachenow reliably clears stale data, and neither it nor checkouts breakgetcwd()for processes working inside the mount on Linux 6.2 and newer.statandgit statusno longer slow down after writing many files on an exclusive mount or in a checked-out directory.- Renaming a file to its own name through the Node SDK no longer makes the disk read-only.
Performance
- Lower Linux client CPU usage for directory listings and for opening and closing files, about 30% less on repeated searches over a large source tree.
- Better caching: data read again after the cache expires is cached again, and files open in many processes keep their kernel-cached pages.
- Lower tail latency when a connection is saturated.
archil mountno longer waits on an extra DNS lookup that could add 5 seconds.
v0.10.1
Released 2026-10-28Changed
- The Linux
.deband.rpmpackages now install the system CA certificates the client needs to verify Archil servers.
Fixed
- The Linux install script now supports
microdnf-only minimal images (such as Amazon Linux 2023 minimal and UBI minimal) and openSUSE, and installs themounttool and FUSE library when a container image lacks them. - The install script no longer hangs at “Installing package…” during unattended installs, and Ctrl-C now stops it. Set
ARCHIL_AUTO_UPGRADE=trueorfalseto skip its upgrade prompt. - Fixed a rare race where concurrent opens of the same file in separate processes on the same mount could return stale data.
- When multiple clients share a disk, moving a directory that another client recently moved no longer fails with “Stale file handle”.
- Deleting, renaming or creating entries under the reserved
/.archildirectory now fails immediately with “Operation not permitted” instead of reaching the server, where it could turn the mount read-only, so mirror tools likerclone syncskip it and carry on. - Files created inside a setgid directory now inherit the directory’s group instead of the creating process’s group.
- Setting an extended attribute with a create-only or replace-only flag no longer makes later extended attribute writes fail for the rest of the mount session.
Performance
archil checkoutclears stale cached data from the kernel with less overhead.- Slightly lower CPU overhead for every filesystem request.
v0.10.0
Released 2026-10-22Changed
- The CLI, macOS app, and Node SDK now verify Archil server certificates against the operating system’s trust store, so Linux systems and containers must have CA certificates installed (for example, the
ca-certificatespackage) or setSSL_CERT_FILE. archil checkpoints createrefuses the namenull, which S3 tools use to mean the current version of a file.
Fixed
- The CLI no longer trusts HTTPS certificates from the eMudhra intermediate certificate authorities whose private keys were disclosed in October 2026.
- On disks shared between clients, a file’s size and permissions are now up to date immediately after opening it or changing it with chmod or truncate, instead of briefly showing older values.
- Linux mounts now support open-file-description locks and report file locks in
/proc/locksby using kernel-local locking.
Performance
- Prefetching many paths now resolves shared parent directories once per batch instead of once per path, reducing client CPU usage for large prefetch requests over deep directory trees.
- Improved performance of the client’s caches when many operations run at once.
v0.9.3
Released 2026-10-16Fixed
- Fixed a bug that could cause a client to report a stale file size when a disk is mounted by multiple clients.
- Prevents a rare condition in S3 native-format disks where a file create, hard link, or rename whose full path exceeds S3’s key length limit could be accepted locally and then fail to flush, leaving the disk read-only.
Performance
- Sequential writes to large files are faster.
- Creating, hard-linking, and renaming files deep in a directory tree is faster in cases where cached attributes have expired.
- The Linux client binary and its .deb/.rpm packages are about 80% smaller (roughly 52 MB instead of 267 MB).
v0.9.2
Released 2026-10-14Fixed
archil checkoutof a file or directory created less than about a second earlier no longer fails with “does not exist”.- On conditional mounts with a checked-out directory, deleting, renaming or hard-linking a file no longer fails with “Invalid cross-device link” after its other hard links outside the delegation have been removed.
- Closed a race in multi-writer scenarios that sometimes caused ENOENT failures when accessing a file immediately after creating it.
- Switching a mount to a read-only token with
archil update-tokennow makes it read-only instead of failing it.
macOS
- Disks mounted with a read-only token now mount read-only in the macOS app, so writes fail immediately instead of being accepted and rejected later.
v0.9.1
Released 2026-10-09Fixed
- Fixed the default data cache being sized from the host’s total RAM instead of the container’s memory limit in privileged containers and on cgroup v1 nodes (such as some GKE node pools).
- Fixed renaming or deleting a file in a checked-out directory that also has a hard link outside it turning the whole mount read-only; the operation now fails with an error instead.
- Fixed a stall when listing a directory while
archil invalidate-cacheran on a file in it on--conditionalmounts. - Fixed
archil set-cache-expiry --readdir-expirynot applying to names learned from a directory listing, so files another client deleted or renamed now disappear after that interval instead of after five minutes. - Fixed a rare crash or stale file view when a rename raced with other creates, renames, or deletes of the same paths on the same mount.
Performance
- Preserved data in the kernel page cache when re-opening a file that has not changed since it was last opened, instead of reading it from the server again.
- Sped up writing large directory trees, such as
npm install, by no longer asking the server about the contents of newly created directories.
v0.9.0
Released 2026-10-06Performance
- Improved write throughput in conditional mode by up to 256x when performing small writes.
Fixed
- Fixed kernel tuning reliability for CSI driver mounts on older kernels.
- Fixed a rare bug in which unmounting or closing a file emits an error.
- Fixed bugs related to stale timestamps and link counts for recently changed file attributes.
- Improved consistency of mount time.
- Various fixes around setuid/setgid during unprivileged fallocate or truncate when mounted in conditional mode.
- Fixed Linux musl (Alpine) client builds.
- Fixed a bug where the file data cache was underutilized, improving read performance for large files.