set_key and unset_key rewrite the .env file through rewrite() (src/dotenv/main.py): they write a temporary file next to the target, then os.replace it over the target. That makes the swap atomic for other processes, but nothing forces the new contents to disk before the rename. The temporary file is closed (which flushes Python's buffer to the OS) and then renamed, with no fsync.
What can go wrong
If the machine loses power or the kernel crashes shortly after set_key returns, the rename can reach the disk before the file's data does. After a reboot, .env can then be empty (zero length) or contain garbage, instead of holding either the old or the new contents. This applies to set_key, unset_key, and the dotenv set / dotenv unset commands.
How likely this is depends on the filesystem. ext4 with its default auto_da_alloc option flushes data before a rename that replaces an existing file, which covers this exact pattern. Other filesystems (XFS, and ext4 mounted with noauto_da_alloc) make no such promise, and a .env that has never existed before isn't covered by that heuristic at all. I haven't reproduced data loss; this is based on the documented semantics of rename(2) and fsync(2), and is the reason other write-then-rename implementations call fsync.
Since .env files often hold credentials, losing one silently is a bad outcome.
Proposed change
In rewrite(), before os.replace:
dest.flush()
os.fsync(dest.fileno())
This has to happen while the temporary file is still open, i.e. inside the with temp_file as dest: block, after the caller's writes have finished.
Optionally, after os.replace on POSIX, fsync the containing directory so the rename itself is durable:
if os.name == "posix":
dir_fd = os.open(os.path.dirname(os.path.abspath(path)), os.O_RDONLY)
try:
os.fsync(dir_fd)
finally:
os.close(dir_fd)Directory fsync isn't available on Windows, and some filesystems reject it with EINVAL. Since the file contents are already safe at that point, an OSError from it could reasonably be ignored.
Cost
.env files are small and set_key is not called in hot loops, so one or two fsync calls per write should not be noticeable. A test can check that os.fsync is called on the temporary file before os.replace, for example by patching both in dotenv.main and recording the call order.
Out of scope
Concurrent writers (two set_key calls at once can lose one update) and file-replacement side effects such as single-file Docker bind mounts are separate problems and not addressed here.
set_keyandunset_keyrewrite the.envfile throughrewrite()(src/dotenv/main.py): they write a temporary file next to the target, thenos.replaceit over the target. That makes the swap atomic for other processes, but nothing forces the new contents to disk before the rename. The temporary file is closed (which flushes Python's buffer to the OS) and then renamed, with nofsync.What can go wrong
If the machine loses power or the kernel crashes shortly after
set_keyreturns, the rename can reach the disk before the file's data does. After a reboot,.envcan then be empty (zero length) or contain garbage, instead of holding either the old or the new contents. This applies toset_key,unset_key, and thedotenv set/dotenv unsetcommands.How likely this is depends on the filesystem. ext4 with its default
auto_da_allocoption flushes data before a rename that replaces an existing file, which covers this exact pattern. Other filesystems (XFS, and ext4 mounted withnoauto_da_alloc) make no such promise, and a.envthat has never existed before isn't covered by that heuristic at all. I haven't reproduced data loss; this is based on the documented semantics ofrename(2)andfsync(2), and is the reason other write-then-rename implementations callfsync.Since
.envfiles often hold credentials, losing one silently is a bad outcome.Proposed change
In
rewrite(), beforeos.replace:This has to happen while the temporary file is still open, i.e. inside the
with temp_file as dest:block, after the caller's writes have finished.Optionally, after
os.replaceon POSIX,fsyncthe containing directory so the rename itself is durable:Directory
fsyncisn't available on Windows, and some filesystems reject it withEINVAL. Since the file contents are already safe at that point, anOSErrorfrom it could reasonably be ignored.Cost
.envfiles are small andset_keyis not called in hot loops, so one or twofsynccalls per write should not be noticeable. A test can check thatos.fsyncis called on the temporary file beforeos.replace, for example by patching both indotenv.mainand recording the call order.Out of scope
Concurrent writers (two
set_keycalls at once can lose one update) and file-replacement side effects such as single-file Docker bind mounts are separate problems and not addressed here.