2018-03-04, S3QL 2.26

  * Added support for py.test 3.3.0
  * Improved metadata rotation performance.

2018-01-05, S3QL 2.25

  * s3qlctrl now accepts a new *dropcache* command.
  * Fixed a race condition that resulted in mount.s3ql crashing with
    "I/O operation on closed file". Thanks to www.imCode.com for
    sponsoring this work!

2017-11-04, S3QL 2.24

  * fsck.s3ql now accepts a new --force-remote parameter. It does
    dangerous things, and you should probably not use it.

  * When terminated by a signal (rather than by an unmount call),
    mount.s3ql now terminates as quickly as possible (i.e., without
    cleaning the cache), resulting in an unclean unmount.

  * The Swift backend feature detection is no longer affected by the
    no-ssl backend option (which is intended only for connections to
    the authentication server).

  * Accessing the cn-north-1 S3 storage region is now working.

2017-08-18, S3QL 2.23

  * Accessing the us-east-1 S3 storage region is now working again.

  * Retry for some additional SSL Errors. Unfortunately we're playing
    whack-a-mole here.

  * The swift backend now dynamically detects the features supported
    by the backend server.

  * The swift backend now supports optimized object copy and bulk
    deletion (if supported by the server).

2017-06-23, S3QL 2.22

  * The storage URL for Amazon S3 buckets has changed. It now includes
    the bucket's region.

  * Added support for AWS v4 authentication. Thanks to ibbchina.com
    for sponsoring this work!

2016-10-28, S3QL 2.21

  * Added workaround for an issue where S3QL crashes with an
    "ssl.SSLError: [SSL: BAD_WRITE_RETRY]" exception. This problem has
    to be fixed in Python, but we now work around it by closing and
    re-opening the connection when it occurs.

2016-08-24, S3QL 2.20

  * Fixed a problem with exporting S3QL file systems over NFS4,
    where directory contents would appear in duplicate.

2016-06-27, S3QL 2.19

  * Fixed a crash when running s3qlstat on an almost empty
    file system.

2016-04-29, S3QL 2.18

  * The `runtests.py` file has been dropped in favor of requiring
    installation of py.test.

  * S3QL now uses semantic versioning. This means that
    backwards-incompatible versions (i.e., versions that require an
    upgrade of the file system revision) will be reflected in an
    increase of the major version number, i.e. the next
    backwards-incompatible version will have version 3.0.

  * Enabled several FUSE optimizations (most importantly, the
    ``big_writes`` feature) that got accidentally disabled at
    some point in the past. This should result in significant
    performance improvements.

    Thanks to Simon Brown for the report!

  * Fixed a crash with "BUG ALERT: Dirty inode was destroyed".

2016-03-08, S3QL 2.17.1

  * Fixed a bug in the upgrade procedure that prevented file system
    modifications when successive commands used the local metadata
    cache.

  * Fixed a bug in the upgrade procedure that resulted in corrupted
    file [acm]times when there was no cached metadata available for
    both s3qladm and the first fsck.s3ql/mount.s3ql invocation after
    the upgrade.

2016-03-08, S3QL 2.17

  * The internal file system revision has changed. File systems
    created with S3QL 2.17 or newer are not compatible with prior S3QL
    versions. To update an existing file system to the newest
    revision, use the `s3qladm upgrade` command.

  * Due to a bug in the FUSE kernel module
    (cf. https://github.com/libfuse/libfuse/issues/23) S3QL does not
    (and did not) honor the write restriction bit. This limitation is
    now documented. (It is unknown if the same problem exists when
    running under BSD or OS-X).

  * S3QL now honors the setgid bit on directories (as opposed to just
    storing them).

  * The `s3qlstat` command now accepts a --raw option to inhibit
    pretty-printing of sizes.

  * The `s3qlctrl log` command is now working again.

  * The `s3qlctrl cachesize` command is now working again.

  * The s3 backend now supports the "infrequent access" storage class
    (cf. https://aws.amazon.com/s3/storage-classes/)

  * The 's3qlstat' command now prints more information about cache
    usage, including the number of cache entries and pending object
    removals.

2016-02-23, S3QL 2.16

  * S3QL now supports (and requires) Python-LLFUSE 0.43 or newer
    (but not never than 1.0).

  * S3QL now retries when receiving a "ServiceUnavailable" error from
    an S3 (or compatible) backend.

  * When retrying network requests, the back-off time now includes a
    random factor to avoid accidentally flooding the server with too
    many simultanuous requests.

  * Fixed a problem with fsck.s3ql aborting with a "ConstraintError"
    in some situations.

  * Fixed a problem with fsck.s3ql repeatedly renaming a file in
    /lost+found because it may lack data.

  * In the authinfo2 file, storage urls for the local backend may now
    be specified with a trailing slash.

2015-09-18, S3QL 2.15

  * Fixed a compatibility problem with recent versions of
    Python-LLFUSE (0.41 and newer).

  * S3QL now supports using special characters in the prefix
    for S3 buckets. Previously, characters that would need to
    be percent-encoded when part of an URL (like + or %) would
    result in a crash.

2015-07-31, S3QL 2.14

  * S3QL now longer crashes when the server sends a zero-length
    response that claims to be XML.

  * The default tcp-timeout has been increased from 10 seconds to 20
    seconds.

  * Fixed test failures when FUSE is not available, tests should now
    again be skipped.

  * S3QL now tracks the frequency of server errors and logs warnings
    if they occur more often than 5 times per second over a period of
    at least 60 seconds. Previously, retries would only be logged
    if the same request has been retried more than 3 times.

  * The "s3qlctrl cachesize" command is now working again (previously
    it would appear to succeed, but not actually change the cache
    size).

  * 404 errors when retrying a delete request are now ignored - most
    likely they mean that the server actually received and executed
    the previous try.

  * Fixed an issue with fsck.s3ql crashing when trying to create a
    file with a very long name in /lost+found.

2015-01-31, S3QL 2.13

  * The internal file system revision has changed. File systems
    created with S3QL 2.13 or newer are not compatible with prior S3QL
    versions.

    To update an existing file system to the newest revision, use the
    's3qladm upgrade' command. During the upgrade, all storage objects
    that were created by S3QL 1.1 or earlier will be downloaded and
    re-uploaded. For objects created by S3QL 1.1.2 or later, the
    upgrade will only affect the storage object's metadata (so no
    contents need to be transferred).

  * SECURITY ADVISORY: Previous mkfs.s3ql versions used /dev/urandom
    to generate the master key when creating a new encrypted file
    system. However, /dev/urandom is not guaranteed to provide the
    256-bits of entropy requested by S3QL. This may have allowed an
    attacker to predict parts of the master key in situations where
    little entropy is available (e.g. right after system boot, or if a
    different program has previously drained the entropy pool).

    Note that the master key is not affected when the file system
    passphrase is changed. The only way to recover from a potentially
    compromised master key is to create a new file system and copy
    over all data.

  * When creating new file systems, the master key is now generated by
    reading /dev/random rather than /dev/urandom to ensure sufficient
    entropy.

  * The 'no-ssl' option for the swift backend is now used only when
    connecting to the authentication server. Whether SSL is used when
    connecting to the storage server is determined by the
    authentication server.

  * Fixed a crash when using the "disable-expect100" swift backend
    option.

  * Fixed a race condition that could lead to a "KeyError" crash
    when refreshing a Google Storage OAuth2 access token.

  * Fixed a race condition that could lead to a hanging mount.s3ql
    process and hanging test_thread_hang() unit test.

  * Updated internal metadata storage and checksum format. The old
    format was difficult to secure against malicious backend data and
    could have resulted in false-positive checksum mismatches with
    future or different Python interpreters.

  * Sizes (e.g. in the s3qlstat output) are now always reported
    with three significant digits.

  * Fixed a bug that caused fsck.s3ql to either abort with a
    "apsw.ConstraintError" or to incorrectly consider storage
    objects as missing when the connection to remote server is
    interrupted while retrieving the object list.

  * Storage urls without any prefix (e.g. s3://foo, but also
    s3://foo/) are now using the same local cache
    directory. Previously they would use different directories,
    despite pointing to the same file system.

    (Note that this does not affect storage urls with a prefix,
    s3://foo/bar and s3://foo/bar/ refer to different locations in the
    bucket, and thus correctly use different cache directories).

  * Fixed a problem where mount.s3ql would crash when unmouting the
    file system because it could not delete the cache directory. This
    could happen when the file system was not unmounted cleanly, but
    fsck.s3ql was then run on a different system (or using a different
    cache directory).

  * S3QL now requires at least version 3.4 of the dugong Python
    module.

2014-11-09, S3QL 2.12

  * The s3c backend has a new 'dumb-copy' option. If this option
    is enabled, copy operations are assumed to have succeeded
    as soon as the server returns a '200 OK' status, without
    checking the contents of the response body.

  * The swift and s3c backends have a new 'disable-expect100' to
    disable support for the 'Expect: continue' header. Using this
    option allows S3QL to work with storage servers without proper
    HTTP 1.1 support, but may decrease performance as object data will
    be transmitted to the server more than once in some circumstances.

  * contrib/benchmark.py is now working again.

  * The `tcp-timeout` backend option is now actually working
    instead of resulting in a "TypeError".

  * Fixed a problem where saving metadata would fail with
    "ValueError: Can't dump NULL values".

  * s3qlstat now also gives information about cache usage.

2014-09-04, S3QL 2.11.1

  * By popular demand, mount.s3ql is now able to daemonize again
    and also does so by default.

  * Un-encrypted file systems (created with mkfs.s3ql --plain) can now
    be upgraded to the newest revision.

  * S3QL now automatically retries a request if a backend server
    reports a 408 (Request Timeout) status code.

  * The Google Storage backend now copes better with server errors
    when refreshing an OAuth2 access token.

2014-08-27, S3QL 2.11

  * SECURITY UPDATE (CVE-2014-0485).

    A remote code execution vulnerability was fixed.

    An attacker with control over the communication with the storage
    backend or the ability to manipulate the data stored in the
    backend was able to trigger execution of arbitrary code by
    mount.s3ql, fsck.s3ql, mkfs.s3ql, s3qladm and s3ql_verify. Both
    encrypted and unencrypted file systems were vulnerable.

  * s3ql_verify no longer crashes when checking an empty file system.

  * Fixed a crash when using Google OAuth2 and the first request after
    the access token has expired is a write request.

  * mount.s3ql now supports systemd-style readyness notification. To
    use this feature, make sure that you have the 'systemd' module
    installed for Python 3.

    This feature has not been tested, if you use it or encounter
    problems, please report back.

  * Fixed a race condition that could cause tests/t5_failsafe.py to
    fail.

  * mount.s3ql no longer daemonizes on its own. With a modern init
    system this should no longer be necessary, and when running
    mount.s3ql from the command line the shell can be used to put the
    process into background.

  * There is a new --backend-options parameter that can be used to
    pass backend-specific options to any S3QL command that accepts a
    storage url.

  * The --no-ssl and --ssl-ca-path parameters have been removed. For
    those backends were these parameters make sense, you can use the
    backend options of the same name instead (e.g. instead of
    `--no-ssl` use `--backend-options no-ssl`).

  * Several backends now accept a `tcp-timeout` option. If S3QL is
    unable to communicate with the remote server for longer than this
    period, the TCP/IP connection is re-established.

  * The Amazon S3 backend now accepts a 'sse' option to enable server
    side encryption. Both costs & benefits of S3 server side
    encryption are probably very small, and this option does *not*
    affect any client side encryption performed by S3QL itself.

  * The Amazon S3 backend now accepts a 'rrs' option to enable reduced
    redundancy storage for any newly created objects.

2014-07-28, S3QL 2.10.1

  * 'setup.py install' now also installs the s3ql_verify command.

2014-07-27, S3QL 2.10

  * The internal file system revision has changed. File systems
    created with S3QL 2.10 or newer are not compatible with prior S3QL
    versions. To update an existing file system to the newest
    revision, use the 's3qladm upgrade' command.

    It is strongly recommended to run the (new) s3ql_verify command
    with the --data option at shortly after the upgrade. This is
    necessary to ensure that the upgrade to the next (2.11) S3QL
    release will run smoothly.

  * The User's Guide now contains a description of the possible
    failure modes of mount.s3ql.

  * The --debug command line parameter now generates a bit less
    output by default, and there is an additional --debug-module
    parameter to activate additional messages.

  * When using encryption, S3QL now checks that a storage object's
    key corresponds to the data stored under it.

    The lack of this check in previous versions allowed an attacker
    with control over the storage server to interchange blocks within
    the same file system (which would have resulted in apparent data
    corruption in the file system). Targeted modification of specific
    files, however, would have been very unlikely, because the
    interchange of blocks had to be done blindly (with the attacker
    not knowing which file any block belongs to, nor what its contents
    are).

    Fixes https://bitbucket.org/nikratio/s3ql/issue/52/.

  * S3QL now aborts immediately instead of retrying if the server
    storage server reports that the local clock is skewed.

  * There is a new 's3ql_verify' command. This program retrieves and
    checks every storage object to ensure that all objects are
    available and have not been tampered with. In contrast to
    fsck.s3ql, s3ql_verify does not trust the object list provided by
    the storage server and actually attempts to download the objects
    one by one.

  * S3QL now requires version 3.2 or newer of the dugong module.

2014-06-28, S3QL 2.9

  * Fix crash when using swift backend and server uses an
    authentication URL other than /v1.0.

  * Fixed two test failures when running unit tests as root.

  * Fixed problems when receiving an HTTP error without a well-formed
    XML body from the remote server (this may happen e.g. due to
    failure in a transparent proxy or load balancer).

  * S3QL now depends on the defusedxml Python module
    (https://pypi.python.org/pypi/defusedxml/).

  * S3QL is no longer vulnerable to DOS attacks from malicious backend
    servers. Previously, a malicious server could induce S3QL to
    consume arbitrary amounts of memory by recursive XML entity
    expansion.

  * S3QL now supports Google OAuth2 authentication. To use it, specify
    'oauth2' as backend login, and a valid OAuth2 refresh token as
    backend password. To obtain the refresh token, the (new)
    s3ql_oauth_client command may be used.

  * S3QL now requires version 3.1 or newer of the dugong module.

  * In some cases, transmission errors when reading storage objects
    from a backend may have been misreported as corrupted backend
    objects. This has been fixed.

  * S3QL no longer crashes when data corruption occurs in the first
    few bytes of an LZMA compressed storage object.

  * S3QL now honors the "Retry-After" HTTP header when receiving an
    XML error from a storage server.

  * Fixed a crash that could occur when the remote server (or some
    intermediate proxy) sends a non-XML error response.

  * mount.s3ql and fsck.s3ql now use different exit codes to indicate
    different failure conditions.

2014-03-29, S3QL 2.8.1

  * No changes in S3QL itself.

  * The S3QL 2.8 tarball accidentally included a copy of the Python
    dugong module. This has been fixed in the 2.8.1 release.

2014-03-13, S3QL 2.8

  * Fixed various problems with using a proxy server.

  * Sending SIGUSR1 to mount.s3ql now generates a stacktrace
    (debugging feature).

  * When passing --installed to the test runner, S3QL commands
    are now loaded from $PATH instead of the packages bin/
    directory.

  * The runtest.py script now comes with the correct shebang
    (i.e., it can now be called as "./runtests.py" instead of
    "python3 runtests.py").

  * S3QL now requires the python dugong module
    (https://bitbucket.org/nikratio/python-dugong)

  * Fixed a file system hang when all upload threads encounter
    unexpected backend problems.

2013-12-16, S3QL 2.7

  * Fixed another race condition that could lead to mount.s3ql
    crashing with `ValueError: I/O operation on closed file`.

  * S3QL no longer generates warning messages for the first two times
    that it has to resend a request to the storage backend. If there
    is no success after the second try, messages are emitted as before.

  * S3QL now stores multiple copies of the master encryption key to
    allow recovery if the backend looses the object holding the
    primary copy. To take advantage of this functionality for existing
    file systems, change the file system passphrase with s3qladm.

  * Fixed problem with automatic cache size detection (mount.s3ql was
    treating bytes as kilobytes). Thanks to GVormayr@gmail.com for the
    patch!

  * Fixed "AttributeError in LegacyDecryptDecompressFilter" crash when
    reading objects written by old S3QL versions.

  * Fixed a problem with umount.s3ql giving a strange error when
    the mountpoint is still in use.

2013-12-03, S3QL 2.6

  * Fixed a problem with mount.s3ql crashing with `KeyError` in
    `collections/abc.py`.

  * Fixed a problem with mount.s3ql crashing with `ValueError: I/O
    operation on closed file`.

  * Fixed a race condition that could lead to data corruption when
    multiple processes trigger download of the same data block
    simultaneously.

  * Fixed a problem with S3QL crashing with "SSLEOFError" when an SSL
    connection is terminated unexpectedly.

  * Increased deduplication performance. Previously, block checksums
    were accidentally computed without parallelism.

  * Fixed a problem with mount.s3ql (incorrectly) reporting corrupted
    data for compressed blocks of some specific sizes. Many thanks to
    Balázs for extensive debugging of this problem.

  * Specifying --debug when using syslog or file logging no longer
    suppresses informational messages to stdout.

  * Fixed build under MacOS X. Thanks to john@nachtimwald.com for the
    patch!

  * mount.s3ql now autodetects a sensible defaults for the
    --max-cache-entries and --cachesize options.

2013-10-19, S3QL 2.5

  * Removed a race condition that resulted in a crash when a file was
    deleted while the upload was in progress, and the upload was then
    restarted because of a connection problem.

  * Specifying bzip2 as the compression algorithm for mount.s3ql
    no longer gives an error.

  * Fixed another race condition that could lead to the (apparent)
    vanishing of the file system metadata when the mount.s3ql is
    killed at an unfortunate time.

  * S3QL programs no longer create an empty file in the default log
    file location if a custom log file is specified with --log.

  * contrib/benchmark.py is now working again - thanks to
    davidthomson@gmail.com.

  * If a backend returns corrupted data, mount.s3ql no longer crashes
    but switches to read-only operation.

  * Fixed a problem with error messages from mount.s3ql being lost,
    resulting in a file system shutdown for apparently no reason.

2013-08-27, S3QL 2.4

  * Attempts to upgrade a file system that is already at the newest
    revision no longer render the local metadata copy unusable.

  * Google storage bucket names containing dots no longer cause
    problems with SSL certificate validation.

  * Fixed crash when statvfs() is called on an almost empty file
    system (which may happen automatically when using some desktop
    environments, resulting in an immediate crash after mounting an
    empty filesystem).

  * There is a new "swiftks" backend for OpenStack/Swift servers that
    supports keystone authentication and allows specification of the
    desired storage region.

  * Eliminated a race condition that could lead to mount.s3ql crashing
    with a "KeyError in collections/abc.py" under some circumstances.

  * Fixed a crash that could occur when trying to clean up after
    receiving a connection or server error when uploading an object.

2013-07-18, S3QL 2.3

  * Reading the file system passphrase from the authinfo2 file
    is now working again.

  * The precedence of the authinfo2 sections has been fixed to agree
    with the documentation: later entries override earlier entries
    rather than the other way around.

  * Fixed a problem with mount.s3ql not recognizing the upgraded
    revision when `s3qladm upgrade` was run with a different
    --cachedir argument.

  * Fixed a crash of the s3 backend when multiple objects are deleted
    and the s3 connection needs to be re-established.

2013-07-12, S3QL 2.2

  * Eliminated a race condition that resulted in an apparent loss of
    remote metadata when the mount.s3ql process was interrupted after
    renaming the remote metadata but before having completely uploaded
    the local metadata.

  * Attempting to run mount.s3ql with --allow-user but without having
    the necessary permissions no longer results in the file system
    being marked as needing fsck.

  * The S3 backend now deletes multiple storage objects using one
    request if possible, reducing latency, traffic and number of
    requests to S3.

  * S3QL is now able to detect server errors and redirections before
    all data has been sent to the server. This avoids pointless data
    transfers and should fix a problem with S3QL operating very slow
    when operating on recently created storage buckets/containers.

  * There is a new 'rackspace' backend that allows specification of
    the object storage region in the storage URL. The rackspace
    backend also uses a different authentication API than the swift
    backend (which can still be used for Rackspace as well). When
    using the 'rackspace' backend, use your regular Rackspace password
    instead of the API key as the backend password.

  * The mount.s3ql --compress option now allows specification of the
    compression level in addition to the compression algorithm.

2013-06-14, S3QL 2.1 (beta)

  * S3QL now uses SSL connections by default.

  * S3QL now verifies server certificates when using SSL connections.

  * There is a new --no-ssl option that now works for every backend.

  * The Amazon S3 and Swift backends are now working again (regression
    was introduced in 2.0-beta).

2013-05-12, S3QL 2.0 (beta)

  * S3QL now requires Python 3.3 or newer.

  * S3QL now requires the pycrypto third-party Python module.

  * S3QL now requires at least version 0.39 of the llfuse Python
    module.

  * Cycling of metadata backups when using the local backend now
    takes much less time.

  * S3QL no longer requires the unittest2, pycryptopp and lzma
    third-party Python modules.

  * It is no longer required that S3QL and Python APSW be linked to
    the same SQLite shared library.

  * mount.s3ql: fixed crash when using --metadata-upload-interval 0.

  * Instead of completely unmounting the file system (and leaving the
    mount point inaccessible) when backend data is missing or
    corrupted, S3QL now returns an error and switches to read-only
    operation.

2013-04-09, S3QL 1.14

  * Reduced CPU consumption when S3QL is idle.

  * S3QL now automatically retries requests when S3 reports an
    "internal error" (this functionaly got accidentally broken
    in S3QL 1.13).

  * Fixed a hang when using s3qlrm to delete a directory with
    a very large number of sub-directories.

2013-03-03, S3QL 1.13.2

  * Allow S3 bucket names starting with a number.

  * Return EOPNOTSUP error on ACL operations. The previous value of
    EINVAL works with FUSE, but confuses several applications.

2013-02-24, S3QL 1.13.1

  * Change error code returned to applications trying to set ACL
    extended attributes. The previous error code of ENOSYS had the
    unintended side effect that some FUSE versions gave up trying to
    get/set any extended attributes at all.

2013-02-23, S3QL 1.13

  * Fixed issues when trying to access a directory at the root
    of the file system that has recently been removed with s3qlrm.
    (In that case the directory was no longer included in the
    readdir() output, but it was still posssible to stat() it because
    was still present in the kernel's dentry cache).

  * When trying to connect to the backend, S3QL now also retries at
    increasing intervals if DNS appears to be unavailable (i.e., if
    there's currently no network at all). This means that mount.s3ql
    will no longer bail out when called without a network connection
    but simply retry (printing warning messages in the process).

  * Work around undocumented "Timeout" reply from Google Storage and
    retry instead of crashing.

  * Instead of failing with a bogus "Invalid credentials or skewed system
    clock" error when trying to access a bucket with an invalid DNS
    name, S3QL now prints a proper error message.

    (Such bucket names are still allowed when using the US Standard
    regions, but generally not a good idea).

  * Removed ostensible ACL support. It was found that S3QL never kept
    ACLs and file permission bits synchronized as mandated by POSIX.
    This is due to a bug in the FUSE library and cannot easily be
    fixed in S3QL. Consequently, S3QL no longer reports ACL support
    unless a FUSE version with the necessary bugfix version is
    installed.

  * Fixed a bug that caused malformed metadata to be written out when
    using recent eglibc versions (the relevant change in eglibc
    was introduced between eglibc versions 2.13 and 2.17).


2012-09-03, S3QL 1.12

  * Be more verbose about how to remedy the situation when attempting
    to upgrade file system needing fsck.

  * Fsck now detects and fixes directory entries which do not have an
    entry type stored in their mode field.

2012-05-04, S3QL 1.11.1

  * Fixed crash when using S3 backend.
    (Regression introduced in release 1.11).

  * Increase minimum reported file system size to 1 TB, and work
    around df bug so that size is reported properly.

  * umount.s3ql now produces a more helpful error message if the
    file system has crashed.

2012-04-29, S3QL 1.11

  * S3QL no longer uses the confusing "bucket" term (which is S3
    specific). The "bucket-passphrase" option in the authentication
    file has been renamed to "fs-passphrase".

  * Metadata about stored objects is now split into multiple fields,
    so that no field is longer than 256 bytes. This makes S3QL
    compatible with a wider range of storage providers.

    Thanks to Stuart Wallace <stuart@wapbot.co.uk> for reporting this
    issue.

  * The SWIFT backend now retries if it receives 5xx error codes from
    the remote server, and no longer fails when attempting to upload
    data with an expired auth token. Thanks to Ken
    <forken.liu@delta.com.tw> for the report and initial patch!

2012-02-24, S3QL 1.10

  * The s3s, s3cs and gss backends have been removed. Use the new
    --ssl option together with the s3, s3c and gs backends instead.

  * S3QL no longer keeps track of consistency guarantees for the
    different backends. The increasing number of different storage
    providers offering different regions, redundancy and availability
    levels makes this no longer feasible.

  * The User's Guide contains a new section "Important Rules to Avoid
    Losing Data". Reading it is strongly recommended.

2012-01-21, S3QL 1.9

  * Fixed a problem with file and directory time stamps
    changing with the time zone of the mount.s3ql process.

  * Fixed a crash in contrib/benchmark.py and made output
    (hopefully) easier to understand.

  * Fixed "Too many chained redirections" bug when trying to access
    a new bucket immediately after creation.

  * When receiving 5xx errors from the server, S3QL now waits and
    retries instead of aborting immediately.

  * contrib/s3ql_upstart.conf now supports running S3QL as an
    ordinary user.

  * S3QL now has experimental, native support for OpenStack/SWIFT
    servers. Thanks to the guys from www.bitbackup.de for providing
    support and a test server!

  * S3QL now has experimental proxy support and will honor the
    http_proxy and https_proxy environment variables.

2011-12-06, S3QL 1.8.1

  * Fixed direct updates from 1.6 to 1.8.

2011-12-05, S3QL 1.8

  * Fixed "ValueError: Can't dump NULL values" crash when file system
    metadata was uploaded during heavy file system activity.

  * Deleting files no longer takes O(number of distinct filenames)
    time (bug introduced in 1.7).

  * Improved s3qlcp and directory listing performance for very large
    directories.

  * The --blocksize option of mkfs.s3ql has often been misunderstood,
    leading people to create file systems with block sizes
    significantly smaller than 1 MB.

    As a matter of fact, a small S3QL block size does *not* have any
    advantage over a large block size when storing lots of small
    files. A small block size, however, seriously degrades performance
    when storing larger files. This is because S3QL is effectively
    using a dynamic block size, and the --blocksize value merely
    specifies an upper limit.

    To make this more clear, the --blocksize option has been renamed
    to --max-obj-size. If you created your file system with a block
    size of less than 1 MB, it is strongly recommended to recreate it
    with a larger block size. Unfortunately, an in-place conversion of
    the block size is not possible.

  * mount.s3ql has a new --min-obj-size option. In the future, files
    smaller than the minimum object size may be combined into groups
    that are stored as single objects in the storage backend.

  * Depend on python-llfuse 0.37. This fixes a problem with the file system
    effectively blocking any other request while an s3qlrm, s3qllock
    or s3qlcp instance is running.

  * Fixed some crashes happening under heavily concurrent file system
    usage.

2011-11-27, S3QL 1.7

  * Extended attribute names are now de-duplicated.

  * Metadata is now stored in a custom, delta-encoded binary format and
    then BZIP2 compressed, resulting in a 5-fold speedup when dumping
    and compressing.

  * Inodes are now assigned sequentially rather than randomly, and old
    inodes are not reused. This makes S3QL fully NFS compatible and
    allows metadata to be stored much more efficiently, resulting in
    a 4 to 8 fold decrease in compressed metadata size.

  * s3qlcp now also copies extended attributes.

  * s3qlcp no longer generates incorrect block reference counts when
    copying a file that has identical blocks (i.e., that can be
    de-duplicated within itself).

  * Eliminated a race condition in s3qlcp. When copying a file with
    s3qlcp immediately after it was modified or created, it was
    possible that s3qlcp would copy the new file attributes (size,
    modification time, etc.) but the old data blocks.

    A copy of a newly created file would then contain only zero bytes,
    while a copy of a modified file would look like the original but
    contain data from before the modification.

  * "mkfs.s3ql --force" and "s3qladm clear" are working again.

2011-11-20, S3QL 1.6

  * fsck: gracefully recover if temporary indices already exist
    (e.g. if previous fsck was interrupted).

  * Due not fail with "BadDigest" error when objects are only
    partially retrieved from backend.

2011-10-20, S3QL 1.5

  * Fixed parsing of storage urls, s3s:// no longer generates bogus
    error message.

  * Fixed support for prefix in storage urls.

  * Retry on timeout when transmitting data to remote server.

  * Do not free mount point when terminating due to unhandled
    exception in background thread.

2011-10-06, S3QL 1.4

  * Metadata is now always LZMA compressed, mount.s3ql's --compress
    option only affects compression of file contents.

  * Network errors that occur in the middle of a read operation
    are now handled correctly as well.

  * s3qladm now uses cached metadata if available.

2011-10-04, S3QL 1.3

  * Fixed an s3qlcp crash happening when source or destination inode
    number was a 64 bit value.

  * "Connection reset by peer" network errors are now handled
    correctly.

  * fsck.s3ql can now renumber inodes. This is useful if the file
    system was mounted using S3QL 1.2 or 1.1.x without the --nfs
    option and can now no longer be accessed on 32bit systems.

  * Use only 32 bits for inodes, even when --nfs is not specified.
    This ensures that file systems can be safely used by both 32 and
    64 bit systems.

2011-09-28, S3QL 1.2

  * Fixed a database problem that was responsible for file
    system access becomer slower and slower the more data
    was stored in the file system.

  * Fixed a race condition that could cause applications to get just
    zero bytes when trying to read from a file that has just been
    copied with s3qlcp.

2011-09-20, S3QL 1.1.4

  * Fixed a typo that caused errors when trying to remove any blocks
    that have been committed to the backend.

  * Improved accuracy of s3qlstat during active file transfers
    (but still not 100% accurate).

  * Fixed some theoretical deadlocks.

  * contrib/benchmark.py is now working again and also takes into
    account the throughput from userspace to S3QL.

2011-09-18, S3QL 1.1.3

  * Fixed a race condition in the local backend that resulted in
    errors of the form "[Errno 17] File exists: [bucket path]".

  * Added Google Storage backend.

  * Added backend for general, S3 compatible storage services.

  * Fixed a bug that caused S3QL to complain about the backend having
    lost objects when trying to download an object before its upload
    was completed. This could happen because locking was done based on
    inode and block number rather than object id, and the
    de-duplication feature can reuse an object for another inode
    before the upload is completed.

  * Fixed a data corruption bug. If a data block was changed while
    being uploaded to the backend, and a second, identical data block
    was flushed while the upload was in progress, but before the first
    block was changed, the second data block was linked to the
    *modified* data. This happened because changes to an object in
    transit were only checked for after the upload completed, leaving
    a window in which the contents of the upload object did not agree
    with its stored hash.

    This problem can be detected by verifying the hash values of all
    stored data blocks. This procedure will automatically be done when
    the file system is updated to the newest revision, but may take a
    longer time since every object needs to be downloaded and checked.

2011-09-08, S3QL 1.1.2

  * The modules for communicating with the different storage providers
    have been completely rewritten, resulting in improved performance,
    more features and better maintainability.

  * S3 buckets can now be used with arbitrary prefixes, allowing to
    store more than one S3QL file system in a bucket.

  * The format of the --authinfo file has changed. See the
    documentation for details of the new format. To avoid breaking
    backwards compatibility, the default file is now called authinfo2.

  * Network errors are now handled much more consistently.

  * There is a new --nfs option for mount.s3ql that needs to be used
    when the S3QL file system will be exported over NFS.

  * The local backend now stores data and metadata in the same file,
    so it needs only half as many files as before.

  * The --homedir option has been replaced by the more finely grained
    --authfile, --cachedir and --log options.

  * S3QL can now log directly to syslog.

  * The sftp storage backend has been dropped. The recommended way to
    use S3QL over ssh is to use sshfs
    (http://fuse.sourceforge.net/sshfs.html) with S3QL's local
    backend.

  * fsck now checks if all indices have been created. This avoids
    a huge performance problem when mount.s3ql was interrupted
    after downloading metadata, but before creating the indices.

2011-07-23, S3QL 1.1 (development version)

  * Restructured metadata. This should also significantly reduce the
    size of the SQLite database file.

  * Fixed license typo in file header comments, license
    is GNU GPL Version 3, not LGPL.

  * Fixed problem with fsck.s3ql generating extraordinary long
    filenames in /lost+found and then crashing.

  * When called as root, use umount rather than fusermount for
    compatibility with FUSE4BSD.

2011-05-20, S3QL 1.0.1

  * Disabled WAL mode again for now because of unexpected problems
    with s3qlcp, s3qllock and s3qlrm (performance going down orders
    of magnitude, and *very* large *.db-wal file in ~/.s3ql).

2011-05-13, S3QL 1.0

  * S3QL has been declared stable after 2 years of beta-testing did
    not reveal any data-critical bugs.

  * Fixed occasional assertion error when calling s3qlctrl flushcache
    or unmounting the file system.

  * Fixed a race condition when a block is expired while it is in
    transit but has already been modified again.

  * expire_backups.py no longer has an --init option, the state file
    is created automatically if this operation is safe. Instead, there
    is a --reconstruct-state option that can be used to try to
    reconstruct a lost state file.

  * The size of symbolic links is now reported as the length of
    the target instead of zero. This used to confuse revision
    control systems like git and hg.

  * Added man pages for all S3QL commands.

2011-02-04, S3QL 0.30

  * S3QL now defaults to use unencrypted HTTP connections, which
    significantly improves performance when using the S3 backend. For
    an encrypted file system, all data is already encrypted anyway,
    and authentication data is never transmitted in plain text even
    for unencrypted file systems. Therefore, the use of SSL brings
    little gain for most users. To force SSL usage, the new --ssl
    option can be used.

  * mkfs.s3ql now has a --force option to overwrite an existing file
    system.

2010-12-30, S3QL 0.29

  * The FUSE interface has been rewritten using Cython and factored
    out into a separate package, http://code.google.com/p/python-llfuse/.

    This should result in easier installation, better performance and
    better maintainability.


2010-12-19, S3QL 0.28

  * "s3qlctrl upload-meta" now works even if the file system has been
    mounted with --metadata-upload-interval=0.

  * File system metadata is now permanently cached locally. This
    significantly reduces the time required to mount the file system.

  * The documentation is now also included in PDF format.

2010-12-11, S3QL 0.27

  * The authinfo file now supports passwords that include white space.

  * The s3qladm command can now be used to download metadata backups.

  * The --strip-meta option for mount.s3ql has been removed, redundant
    data is now always stripped before upload.

  * mount.s3ql now has a --upstart option so that it can easily run as
    an upstart job. An example job definition is included in
    contrib/s3ql.conf.

  * s3qlctrl now has an 'upload-meta' command to trigger a metadata
    upload while the file system is mounted.

  * Fixed a bug that preserved old data when truncating a file to
    zero. If you ever had fsck errors of the form

    Size of inode [ddd] ([filename]) does not agree with number of
    blocks

    then the affected files may contain additional, old data at the
    end that is not supposed to be there.


2010-10-30, S3QL 0.26

  * Fixed a problem with the sftp backend failing because it tries to
    access a file locally.

  * Various minor bugfixes

2010-09-28, S3QL 0.25

  * A critical bug that could cause backups to be deleted too early
    and potentially break the whole backup strategy has been fixed in
    contrib/expire_backups.py.

    The new version has changed semantics that and also copes
    significantly better when backups are made in irregular time
    intervals.

  * S3QL should now respond with more consistent performance when
    accessing the file system while s3qlcp, s3qllock or s3qlrm is
    running at the same time.

  * When enabling debug output for the `UploadManager` module, S3QL
    now logs detailed messages about de-duplication, upload and
    compression performance.

2010-09-18, S3QL 0.24

  * Fixed a deadlock that could cause the mount.s3ql process to hang
    around after umount.

  * Fixed a bug that caused S3QL to consider all downloaded blocks as
    dirty and resulted in unneccessary checksum calculations on
    expiration.

  * s3qlctrl can now change the log level at runtime.

  * s3qladm delete now also deletes any local stale cache files.

  * Periodic metadata upload can now be disabled completely by
    specifying an interval of zero.


2010-09-03, S3QL 0.23

  * Fixed problem with global_lock.py not being installed by setup.py

2010-08-31, S3QL 0.22

  * Fixed a bug that could cause file system errors when calling
    s3qlrm on a lot of really large files.

  * The sftp backend is now significantly faster, thanks to skyang2009
    for the patch.

  * s3qlctrl can now change the cache size of a mounted file system.

2010-08-09, S3QL 0.21

  * The metadata is now only uploaded if there have been any changes.

  * mount.s3ql now supports several parallel compression and
    encryption threads with the --compression-threads option.

  * S3QL now supports "immutable directories". This important new
    feature can be used to protect backups from modification after
    they have completed. See the User's Guide for details.

  * Using S3 RRS is now deprecated, see User's Guide for details.

  * fsck.s3ql now moves damaged files due to data lost by a backend
    into /lost+found/

  * expire_backups is no longer installed automatically and can now be
    found in the contrib/ directory.

  * S3QL now comes with sample backup script in contrib/s3_backup.sh

  * Added contrib/pcp.py, an rsync wrapper to recursively copy
    directory trees with several parallel rsync processes.

2010-08-01, S3QL 0.20.1

  * Hotfix for s3qladm upgrade.

2010-08-01, S3QL 0.20

  * Added contrib/make_dummy.py. This script creates a dummy copy of a
    bucket that contains only the file system metadata. The resulting
    dummy can be used for debugging.

  * Mounting with the local and sftp backends is now significantly
    faster, because the object tree is no longer traversed completely.

  * Fixed a race condition that occasionally produced errors when
    deleting large files (spanning multiple blocks).

  * The file system now stays responsive to other requests while
    compressing blocks.

  * s3qlstat is now much faster since the size after de-duplication
    does not need to be queried from the backend anymore but is stored
    in the metadata.


2010-07-25, S3QL 0.19

  * Fixed various smaller bugs, see Mercurial changelog for details.

2010-07-11, S3QL 0.18

  * Added --strip-meta option to mount.s3ql

  * Added --metadata-upload-interval option to mount.s3ql. This allows
    to periodically upload updated metadata even while the file system
    is mounted.

  * stat.s3ql, tune.s3ql and cp.s3ql have been renamed to s3qlstat,
    s3qladm and s3qlcp respectively.

  * sftp backend is working again

  * Added the s3qlctrl command.

2010-06-29, S3QL 0.17

  * The local and sftp backends now spread their files into different
    sub-directories.

  * Amazon S3 Reduced Redundancy Storage (RRS) is now supported. To
    use it, use a storage url of the form s3rr://<bucket> instead of
    s3://<bucket>.

2010-06-15, S3QL 0.16

  * Fixed problem with readdir() returning entries twice or skipping
    them if files are added or removed while readdir() is in progress.

  * Fixed build problem on Gentoo.

  * fsck.s3ql now does a rudimentary check if the file system is still
    mounted to prevent checking a mounted file system.


2010-05-28, S3QL 0.15

  * Fixed test cases to handle systems with low system clock
    resolution.

  * Corrected installation instructions for Debian

  * mount.s3ql: instead of --bzip2, --zlib and --no-compress, there is
    now just one option --compress=<alg>.

  * File system metadata is now uploaded with all indices. This makes
    mounting the file system much faster. Only if LZMA compression has
    been chosen, indices are stripped for storage and regenerated on
    the next mount.

2010-05-14, S3QL 0.14

  * fsck now detects if a cached block is dirty and commits only dirty
    blocks to the backend.

  * Installation in Debian and Ubuntu is now much simpler, it can be
    done completely with aptitude.

2010-05-04, S3QL 0.13

  * S3QL now works with Ubuntu Karmic / 10.04 LTS

  * The test command no longer produces bogus error messages after all
    tests have completed.

  * The umount.s3ql command now properly handles the 'fuser' output
    with Kernel 2.6.32 (before it always refused to unmount, claiming
    that the mount point was busy).

  * The compression can now be specified independently from the
    encryption, so it is possible to have an unencrypted, but
    compressed file system.

  * Apart from zlib, bzip2 and lzma, data can now also be stored
    without any compression.

  * S3QL no longer emulates the . and .. directory entries since this
    is not required by POSIX and makes the code much simpler. This
    also means that the st_nlink value of a directory is not longer 2
    + number of subdirectories.

  * Fixed a bug that caused files to be listed with wrong sizes under
    certain conditions.

  * Added `expire_backups` command and instructions for a simple
    backup solution using S3QL and rsync.

2010-04-27, S3QL 0.12

  * fsck.s3ql now accepts a --batch option to not prompt for any user
    input and requires a --force option to check the file system even
    if it is marked as clean.

  * Fixed a bug in cp.s3ql that caused incorrect st_nlink values in
    the copy.

  * fsck.s3ql has been even more optimized.

  * Fixed a problem with running out of file descriptors when lots of
    objects are deleted.

  * Bucket encryption passwords can now also be stored in the
    ~/.s3ql/authinfo file.

  * mount.s3ql doesn't complain any more if it receives any of the
    standard mount(8) mount options.


2010-04-24, S3QL 0.11

  * S3QL file system can now be mounted from /etc/fstab

  * Metadata now takes significantly less space.

  * Files with extended attributes can now be deleted.

  * Extended attributes can now be listed.

  * It is now possible to choose between zlib, BZip2 and LZMA
    compression every time the file system is mounted.

  * Added contrib/benchmark.py to find out optimal compression method
    for given network bandwidth.

  * fsck.s3ql no longer uses mknod(3) , since that lead to errors on
    e.g. NFS mounted volumes under Fedora.

  * File access, modification and inode change times before 1972 are
    now supported.

  * Fixed a deadlock when removing or overwriting files.

2010-04-21, S3QL 0.10

  * S3QL now depends on FUSE version 2.8

  * Cached blocks are now flushed to the backend as soon as they have
    not been accessed for more than 10 seconds.

  * The setup script now automatically checks and, if necessary,
    installs the Python module dependencies.

  * mkfs.s3ql now creates compressed and encrypted file systems by
    default.

  * Data is now compressed with LZMA instead of Bzip2.

  * Data compression and data upload is now done in parallel, so the
    full network can bandwidth is continuously without breaks for data
    compression.

  * fsck.s3ql is now several orders of magnitude faster. The code has
    been optimized and some special database indices are now
    precalculated.

  * When running cp.s3ql, the attributes of the target directory are
    now immediately refreshed (so that e.g. ls immediately shows the
    correct number of hard links).

  * File removal is now significantly faster, since the network
    transactions are carried out asynchronously as part of the cache
    expiration.

  * mount.s3ql no longer tries to create files with mknod(), since
    that lead to errors on NFS mounted volumes under Fedora.

  * This releases includes a lot of new code, so it may be slightly
    more beta-ish than usual.


2010-04-04, S3QL 0.9

  * The --cachedir, --logfile, --awskey and --credfile options are
    gone and have been replaced by --homedir.

  * ~/.awssecret is no longer read, instead there is a common file
    with authentication data for all backends in ~/.s3ql/authinfo

  * The syntax for mounting S3 buckets is now s3://bucketname instead
    of just the bucket name

  * There is now an SFTP backend. Thanks to Ron Knapp
    <ron.siesta@gmail.com> for most of the code.


2010-03-07, S3QL 0.8

  * S3QL now uses Python's default unittest.py instead of shipping its
    own.

  * Most documentation has been moved from the Wiki into the tarball,
    so that it always corresponds to the correct version.

  * setuptools is now used for installation. This allows .egg
    creation, dependency resolution and generation of the HTML
    documentation.

  * The S3 part of boto has been integrated into S3QL.

2010-02-22, beta7

  * mount.s3ql no longer chdirs into / when daemonizing but into the
    cache directory.

  * Added example backup script in contrib/backup.py

  * tune.s3ql --change-passphrase is working again

  * Added testcase for tune.s3ql --change-passphrase

  * Internal S3 errors are now handled properly by retrying the
    upload.

2010-02-19, beta6

 * tune.s3ql --copy is now *really* working properly (brrr)

2010-02-19, beta5

 * mkfs.s3ql now makes strict checks on the bucket name

 * Removed obsolete mount.s3ql_local from distribution

 * tune.s3ql --copy is now working properly

2010-02-19, beta4

    * tune.s3ql now has a --copy option to copy buckets

    * Storage location for new buckets can be specified in mkfs.s3ql
      and tune.s3ql with --s3-location

    * Fixed a deadlock in mount.s3ql when using local buckets without
      --fg

    * The local: bucket specifier is no longer artificially slow.

    * mount.s3ql: --allow_other is now working

2010-02-04, beta3

    * Fixed a deadlock that caused umount.s3ql to hang indefinitely
      when mount was called without --fg

    * The '.' and '..' directory entries are no longer explicitly
      stored in the database.

    * Rewrote the mechanism to handle delayed updates. Now it no
      longer depends on a small object being propagated relatively
      fast, but is able to handle arbitrary network outages.
