Makefile: upload-tarballs: switch from xfer-client to flappclient, closes #350

[tahoe-lafs/tahoe-lafs.git] / NEWS

User visible changes in Tahoe-LAFS.  -*- outline -*-

* Release ?.?.? (?)

** New Features

*** Immutable Directories

Tahoe can now create and handle immutable directories. These are read just
like normal directories, but are "deep-immutable", meaning that all their
children (and everything reachable from those children) must be immutable
objects (i.e. immutable/literal files, and other immutable directories).

These directories must be created in a single webapi call, which provides all
of the children at once (instead of the usual create/add/add sequence, since
they cannot be changed after creation). They have URIs that start with
"URI:DIR2-CHK:" or "URI:DIR2-LIT:", and are described on the human-facing web
interface (aka the "WUI") with a "DIR-IMM" abbreviation (as opposed to "DIR"
for the usual read-write directories and "DIR-RO" for read-only directories).

Tahoe releases before 1.6.0 cannot read the contents of an immutable
directory. 1.5.0 will tolerate their presence in a directory listing (and
display it as an "unknown node"). 1.4.1 and earlier cannot tolerate them: a
DIR-IMM child in any directory will prevent the listing of that directory.

Immutable directories are repairable, just like normal immutable files.

The webapi "POST t=mkdir-immutable" call is used to create immutable
directories. See docs/frontends/webapi.txt for details.

*** "tahoe backup" now creates immutable directories, backupdb has dircache

The "tahoe backup" command has been enhanced to create immutable directories
(in previous releases, it created read-only mutable directories). This is
significantly faster, since it does not need to create an RSA keypair for
each new directory. In addition, "DIR-IMM" immutable directories are
repairable, unlike "DIR-RO" read-only mutable directories (at least in this
release; a future Tahoe release should be able to repair DIR-RO).

In addition, the backupdb (used by "tahoe backup" to remember what it has
already copied) has been enhanced to store information about existing
immutable directories. This allows it to re-use directories that have moved
but still contain identical contents, or which have been deleted and later
replaced. (the 1.5.0 "tahoe backup" command could only re-use directories
that were in the same place as they were in the previous backup). With this
change, the backup process no longer needs to read the previous snapshot out
of the Tahoe grid, reducing the network load considerably.

As noted above, Tahoe versions earlier than 1.5.0 cannot read immutable
directories.

** Minor Changes

The webapi acquired a new "t=mkdir-with-children" command, to create and
populate a directory in a single call. This is significantly faster than
using separate "t=mkdir" and "t=set_children" operations (it uses one
gateway-to-grid roundtrip, instead of three or four).

The t=set_children operation is now documented in docs/frontends/webapi.txt .

Several small packaging improvements were made to facilitate the "tahoe-lafs"
package being added to Ubuntu's "Karmic Koala" 9.10 release. Several
mac/win32 binary libraries were removed, some figleaf code-coverage files
were removed, a bundled copy of darcsver-1.2.1 was removed, and additional
licensing text was added.


* Release 1.5.0 (2009-08-01)

** Improvements

Uploads of immutable files now use pipelined writes, improving upload speed
slightly (10%) over high-latency connections. (#392)

Processing large directories has been sped up, by removing a O(N^2) algorithm
from the dirnode decoding path and retaining unmodified encrypted entries.
(#750, #752)

The human-facing web interface (aka the "WUI") received a significant CSS
makeover by Kevin Reid, making it much prettier and easier to read. The WUI
"check" and "deep-check" forms now include a "Renew Lease" checkbox,
mirroring the CLI --add-lease option, so leases can be added or renewed from
the web interface.

The CLI "tahoe mv" command now refuses to overwrite directories. (#705)

The CLI "tahoe webopen" command, when run without arguments, will now bring
up the "Welcome Page" (node status and mkdir/upload forms).

The 3.5MB limit on mutable files was removed, so it should be possible to
upload arbitrarily-sized mutable files. Note, however, that the data format
and algorithm remains the same, so using mutable files still requires
bandwidth, computation, and RAM in proportion to the size of the mutable file.
(#694)

This version of Tahoe-LAFS will tolerate directory entries that contain filecap
formats which it does not recognize: files and directories from the future.
This should improve the user experience (for 1.5.0 users) when we add new cap
formats in the future. Previous versions would fail badly, preventing the user
from seeing or editing anything else in those directories. These unrecognized
objects can be renamed and deleted, but obviously not read or written. Also
they cannot generally be copied. (#683)

** Bugfixes

deep-check-and-repair now tolerates read-only directories, such as the ones
produced by the "tahoe backup" CLI command. Read-only directories and mutable
files are checked, but not repaired. Previous versions threw an exception
when attempting the repair and failed to process the remaining contents. We
cannot yet repair these read-only objects, but at least this version allows
the rest of the check+repair to proceed. (#625)

A bug in 1.4.1 which caused a server to be listed multiple times (and
frequently broke all connections to that server) was fixed. (#653)

The plaintext-hashing code was removed from the Helper interface, removing
the Helper's ability to mount a partial-information-guessing attack. (#722)

** Platform/packaging changes

Tahoe-LAFS now runs on NetBSD, OpenBSD, ArchLinux, and NixOS, and on an
embedded system based on an ARM CPU running at 266 MHz.

Unit test timeouts have been raised to allow the tests to complete on
extremely slow platforms like embedded ARM-based NAS boxes, which may take
several hours to run the test suite. An ARM-specific data-corrupting bug in
an older version of Crypto++ (5.5.2) was identified: ARM-users are encouraged
to use recent Crypto++/pycryptopp which avoids this problem.

Tahoe-LAFS now requires a SQLite library, either the sqlite3 that comes
built-in with python2.5/2.6, or the add-on pysqlite2 if you're using
python2.4. In the previous release, this was only needed for the "tahoe backup"
command: now it is mandatory.

Several minor documentation updates were made.

To help get Tahoe-LAFS into Linux distributions like Fedora and Debian,
packaging improvements are being made in both Tahoe-LAFS and related libraries
like pycryptopp and zfec.

The Crypto++ library included in the pycryptopp package has been upgraded to
version 5.6.0 of Crypto++, which includes a more efficient implementation of
SHA-256 in assembly for x86 or amd64 architectures.

** dependency updates

 foolscap-0.4.1
 no python-2.4.0 or 2.4.1 (2.4.2 is good)
  (they contained a bug in base64.b32decode)
 avoid python-2.6 on windows with mingw: compiler issues
 python2.4 requires pysqlite2 (2.5,2.6 does not)
 no python-3.x
 pycryptopp-0.5.15


* Release 1.4.1 (2009-04-13)

** Garbage Collection

The big feature for this release is the implementation of garbage collection,
allowing Tahoe storage servers to delete shares for old deleted files. When
enabled, this uses a "mark and sweep" process: clients are responsible for
updating the leases on their shares (generally by running "tahoe deep-check
--add-lease"), and servers are allowed to delete any share which does not
have an up-to-date lease. The process is described in detail in
docs/garbage-collection.txt .

The server must be configured to enable garbage-collection, by adding
directives to the [storage] section that define an age limit for shares. The
default configuration will not delete any shares.

Both servers and clients should be upgraded to this release to make the
garbage-collection as pleasant as possible. 1.2.0 servers do not have the
code to perform the update-lease operation, while 1.3.0 servers have
update-lease but will return an exception for unknown storage indices,
causing clients to emit an Incident for each exception, slowing the add-lease
process down to a crawl. 1.3.0 clients did not have the add-lease operation
at all.

** Security/Usability Problems Fixed

A super-linear algorithm in the Merkle Tree code was fixed, which previously
caused e.g. download of a 10GB file to take several hours before the first
byte of plaintext could be produced. The new "alacrity" is about 2 minutes. A
future release should reduce this to a few seconds by fixing ticket #442.

The previous version permitted a small timing attack (due to our use of
strcmp) against the write-enabler and lease-renewal/cancel secrets. An
attacker who could measure response-time variations of approximatly 3ns
against a very noisy background time of about 15ms might be able to guess
these secrets. We do not believe this attack was actually feasible. This
release closes the attack by first hashing the two strings to be compared
with a random secret.

** webapi changes

In most cases, HTML tracebacks will only be sent if an "Accept: text/html"
header was provided with the HTTP request. This will generally cause browsers
to get an HTMLized traceback but send regular text/plain tracebacks to
non-browsers (like the CLI clients). More errors have been mapped to useful
HTTP error codes.

The streaming webapi operations (deep-check and manifest) now have a way to
indicate errors (an output line that starts with "ERROR" instead of being
legal JSON). See docs/frontends/webapi.txt for details.

The storage server now has its own status page (at /storage), linked from the
Welcome page. This page shows progress and results of the two new
share-crawlers: one which merely counts shares (to give an estimate of how
many files/directories are being stored in the grid), the other examines
leases and reports how much space would be freed if GC were enabled. The page
also shows how much disk space is present, used, reserved, and available for
the Tahoe server, and whether the server is currently running in "read-write"
mode or "read-only" mode.

When a directory node cannot be read (perhaps because of insufficent shares),
a minimal webapi page is created so that the "more-info" links (including a
Check/Repair operation) will still be accessible.

A new "reliability" page was added, with the beginnings of work on a
statistical loss model. You can tell this page how many servers you are using
and their independent failure probabilities, and it will tell you the
likelihood that an arbitrary file will survive each repair period. The
"numpy" package must be installed to access this page. A partial paper,
written by Shawn Willden, has been added to docs/proposed/lossmodel.lyx .

** CLI changes

"tahoe check" and "tahoe deep-check" now accept an "--add-lease" argument, to
update a lease on all shares. This is the "mark" side of garbage collection.

In many cases, CLI error messages have been improved: the ugly HTMLized
traceback has been replaced by a normal python traceback.

"tahoe deep-check" and "tahoe manifest" now have better error reporting.
"tahoe cp" is now non-verbose by default.

"tahoe backup" now accepts several "--exclude" arguments, to ignore certain
files (like editor temporary files and version-control metadata) during
backup.

On windows, the CLI now accepts local paths like "c:\dir\file.txt", which
previously was interpreted as a Tahoe path using a "c:" alias.

The "tahoe restart" command now uses "--force" by default (meaning it will
start a node even if it didn't look like there was one already running).

The "tahoe debug consolidate" command was added. This takes a series of
independent timestamped snapshot directories (such as those created by the
allmydata.com windows backup program, or a series of "tahoe cp -r" commands)
and creates new snapshots that used shared read-only directories whenever
possible (like the output of "tahoe backup"). In the most common case (when
the snapshots are fairly similar), the result will use significantly fewer
directories than the original, allowing "deep-check" and similar tools to run
much faster. In some cases, the speedup can be an order of magnitude or more.
This tool is still somewhat experimental, and only needs to be run on large
backups produced by something other than "tahoe backup", so it was placed
under the "debug" category.

"tahoe cp -r --caps-only tahoe:dir localdir" is a diagnostic tool which,
instead of copying the full contents of files into the local directory,
merely copies their filecaps. This can be used to verify the results of a
"consolidation" operation.

** other fixes

The codebase no longer rauses RuntimeError as a kind of assert(). Specific
exception classes were created for each previous instance of RuntimeError.

Many unit tests were changed to use a non-network test harness, speeding them
up considerably.

Deep-traversal operations (manifest and deep-check) now walk individual
directories in alphabetical order. Occasional turn breaks are inserted to
prevent a stack overflow when traversing directories with hundreds of
entries.

The experimental SFTP server had its path-handling logic changed slightly, to
accomodate more SFTP clients, although there are still issues (#645).


* Release 1.3.0 (2009-02-13)

** Checker/Verifier/Repairer

The primary focus of this release has been writing a checker / verifier /
repairer for files and directories.  "Checking" is the act of asking storage
servers whether they have a share for the given file or directory: if there
are not enough shares available, the file or directory will be
unrecoverable. "Verifying" is the act of downloading and cryptographically
asserting that the server's share is undamaged: it requires more work
(bandwidth and CPU) than checking, but can catch problems that simple
checking cannot. "Repair" is the act of replacing missing or damaged shares
with new ones.

This release includes a full checker, a partial verifier, and a partial
repairer. The repairer is able to handle missing shares: new shares are
generated and uploaded to make up for the missing ones. This is currently the
best application of the repairer: to replace shares that were lost because of
server departure or permanent drive failure.

The repairer in this release is somewhat able to handle corrupted shares. The
limitations are:

 * Immutable verifier is incomplete: not all shares are used, and not all
   fields of those shares are verified. Therefore the immutable verifier has
   only a moderate chance of detecting corrupted shares.
 * The mutable verifier is mostly complete: all shares are examined, and most
   fields of the shares are validated.
 * The storage server protocol offers no way for the repairer to replace or
   delete immutable shares. If corruption is detected, the repairer will
   upload replacement shares to other servers, but the corrupted shares will
   be left in place.
 * read-only directories and read-only mutable files must be repaired by
   someone who holds the write-cap: the read-cap is insufficient. Moreover,
   the deep-check-and-repair operation will halt with an error if it attempts
   to repair one of these read-only objects.
 * Some forms of corruption can cause both download and repair operations to
   fail. A future release will fix this, since download should be tolerant of
   any corruption as long as there are at least 'k' valid shares, and repair
   should be able to fix any file that is downloadable.

If the downloader, verifier, or repairer detects share corruption, the
servers which provided the bad shares will be notified (via a file placed in
the BASEDIR/storage/corruption-advisories directory) so their operators can
manually delete the corrupted shares and investigate the problem. In
addition, the "incident gatherer" mechanism will automatically report share
corruption to an incident gatherer service, if one is configured. Note that
corrupted shares indicate hardware failures, serious software bugs, or malice
on the part of the storage server operator, so a corrupted share should be
considered highly unusual.

By periodically checking/repairing all files and directories, objects in the
Tahoe filesystem remain resistant to recoverability failures due to missing
and/or broken servers.

This release includes a wapi mechanism to initiate checks on individual
files and directories (with or without verification, and with or without
automatic repair). A related mechanism is used to initiate a "deep-check" on
a directory: recursively traversing the directory and its children, checking
(and/or verifying/repairing) everything underneath. Both mechanisms can be
run with an "output=JSON" argument, to obtain machine-readable check/repair
status results. These results include a copy of the filesystem statistics
from the "deep-stats" operation (including total number of files, size
histogram, etc). If repair is possible, a "Repair" button will appear on the
results page.

The client web interface now features some extra buttons to initiate check
and deep-check operations. When these operations finish, they display a
results page that summarizes any problems that were encountered. All
long-running deep-traversal operations, including deep-check, use a
start-and-poll mechanism, to avoid depending upon a single long-lived HTTP
connection. docs/frontends/webapi.txt has details.

** Efficient Backup

The "tahoe backup" command is new in this release, which creates efficient
versioned backups of a local directory. Given a local pathname and a target
Tahoe directory, this will create a read-only snapshot of the local directory
in $target/Archives/$timestamp. It will also create $target/Latest, which is
a reference to the latest such snapshot. Each time you run "tahoe backup"
with the same source and target, a new $timestamp snapshot will be added.
These snapshots will share directories that have not changed since the last
backup, to speed up the process and minimize storage requirements. In
addition, a small database is used to keep track of which local files have
been uploaded already, to avoid uploading them a second time. This
drastically reduces the work needed to do a "null backup" (when nothing has
changed locally), making "tahoe backup' suitable to run from a daily cronjob.

Note that the "tahoe backup" CLI command must be used in conjunction with a
1.3.0-or-newer Tahoe client node; there was a bug in the 1.2.0 webapi
implementation that would prevent the last step (create $target/Latest) from
working.

** Large Files

The 12GiB (approximate) immutable-file-size limitation is lifted. This
release knows how to handle so-called "v2 immutable shares", which permit
immutable files of up to about 18 EiB (about 3*10^14). These v2 shares are
created if the file to be uploaded is too large to fit into v1 shares. v1
shares are created if the file is small enough to fit into them, so that
files created with tahoe-1.3.0 can still be read by earlier versions if they
are not too large. Note that storage servers also had to be changed to
support larger files, and this release is the first release in which they are
able to do that. Clients will detect which servers are capable of supporting
large files on upload and will not attempt to upload shares of a large file
to a server which doesn't support it.

** FTP/SFTP Server

Tahoe now includes experimental FTP and SFTP servers. When configured with a
suitable method to translate username+password into a root directory cap, it
provides simple access to the virtual filesystem. Remember that FTP is
completely unencrypted: passwords, filenames, and file contents are all sent
over the wire in cleartext, so FTP should only be used on a local (127.0.0.1)
connection. This feature is still in development: there are no unit tests
yet, and behavior with respect to Unicode filenames is uncertain. Please see
docs/frontends/FTP-and-SFTP.txt for configuration details. (#512, #531)

** CLI Changes

This release adds the 'tahoe create-alias' command, which is a combination of
'tahoe mkdir' and 'tahoe add-alias'. This also allows you to start using a
new tahoe directory without exposing its URI in the argv list, which is
publicly visible (through the process table) on most unix systems.  Thanks to
Kevin Reid for bringing this issue to our attention.

The single-argument form of "tahoe put" was changed to create an unlinked
file. I.e. "tahoe put bar.txt" will take the contents of a local "bar.txt"
file, upload them to the grid, and print the resulting read-cap; the file
will not be attached to any directories. This seemed a bit more useful than
the previous behavior (copy stdin, upload to the grid, attach the resulting
file into your default tahoe: alias in a child named 'bar.txt').

"tahoe put" was also fixed to handle mutable files correctly: "tahoe put
bar.txt URI:SSK:..." will read the contents of the local bar.txt and use them
to replace the contents of the given mutable file.

The "tahoe webopen" command was modified to accept aliases. This means "tahoe
webopen tahoe:" will cause your web browser to open to a "wui" page that
gives access to the directory associated with the default "tahoe:" alias. It
should also accept leading slashes, like "tahoe webopen tahoe:/stuff".

Many esoteric debugging commands were moved down into a "debug" subcommand:

 tahoe debug dump-cap
 tahoe debug dump-share
 tahoe debug find-shares
 tahoe debug catalog-shares
 tahoe debug corrupt-share

The last command ("tahoe debug corrupt-share") flips a random bit of the
given local sharefile. This is used to test the file verifying/repairing
code, and obviously should not be used on user data.

The cli might not correctly handle arguments which contain non-ascii
characters in Tahoe v1.3 (although depending on your platform it
might, especially if your platform can be configured to pass such
characters on the command-line in utf-8 encoding).  See
http://allmydata.org/trac/tahoe/ticket/565 for details.

** Web changes

The "default webapi port", used when creating a new client node (and in the
getting-started documentation), was changed from 8123 to 3456, to reduce
confusion when Tahoe accessed through a Firefox browser on which the
"Torbutton" extension has been installed. Port 8123 is occasionally used as a
Tor control port, so Torbutton adds 8123 to Firefox's list of "banned ports"
to avoid CSRF attacks against Tor. Once 8123 is banned, it is difficult to
diagnose why you can no longer reach a Tahoe node, so the Tahoe default was
changed. Note that 3456 is reserved by IANA for the "vat" protocol, but there
are argueably more Torbutton+Tahoe users than vat users these days. Note that
this will only affect newly-created client nodes. Pre-existing client nodes,
created by earlier versions of tahoe, may still be listening on 8123.

All deep-traversal operations (start-manifest, start-deep-size,
start-deep-stats, start-deep-check) now use a start-and-poll approach,
instead of using a single (fragile) long-running synchronous HTTP connection.
All these "start-" operations use POST instead of GET. The old "GET
manifest", "GET deep-size", and "POST deep-check" operations have been
removed.

The new "POST start-manifest" operation, when it finally completes, results
in a table of (path,cap), instead of the list of verifycaps produced by the
old "GET manifest". The table is available in several formats: use
output=html, output=text, or output=json to choose one. The JSON output also
includes stats, and a list of verifycaps and storage-index strings.

The "return_to=" and "when_done=" arguments have been removed from the
t=check and deep-check operations.

The top-level status page (/status) now has a machine-readable form, via
"/status/?t=json". This includes information about the currently-active
uploads and downloads, which may be useful for frontends that wish to display
progress information. There is no easy way to correlate the activities
displayed here with recent wapi requests, however.

Any files in BASEDIR/public_html/ (configurable) will be served in response
to requests in the /static/ portion of the URL space. This will simplify the
deployment of javascript-based frontends that can still access wapi calls
by conforming to the (regrettable) "same-origin policy".

The welcome page now has a "Report Incident" button, which is tied into the
"Incident Gatherer" machinery. If the node is attached to an incident
gatherer (via log_gatherer.furl), then pushing this button will cause an
Incident to be signalled: this means recent log events are aggregated and
sent in a bundle to the gatherer. The user can push this button after
something strange takes place (and they can provide a short message to go
along with it), and the relevant data will be delivered to a centralized
incident-gatherer for later processing by operations staff.

The "HEAD" method should now work correctly, in addition to the usual "GET",
"PUT", and "POST" methods. "HEAD" is supposed to return exactly the same
headers as "GET" would, but without any of the actual response body data. For
mutable files, this now does a brief mapupdate (to figure out the size of the
file that would be returned), without actually retrieving the file's
contents.

The "GET" operation on files can now support the HTTP "Range:" header,
allowing requests for partial content. This allows certain media players to
correctly stream audio and movies out of a Tahoe grid. The current
implementation uses a disk-based cache in BASEDIR/private/cache/download ,
which holds the plaintext of the files being downloaded. Future
implementations might not use this cache. GET for immutable files now returns
an ETag header.

Each file and directory now has a "Show More Info" web page, which contains
much of the information that was crammed into the directory page before. This
includes readonly URIs, storage index strings, object type, buttons to
control checking/verifying/repairing, and deep-check/deep-stats buttons (for
directories). For mutable files, the "replace contents" upload form has been
moved here too. As a result, the directory page is now much simpler and
cleaner, and several potentially-misleading links (like t=uri) are now gone.

Slashes are discouraged in Tahoe file/directory names, since they cause
problems when accessing the filesystem through the wapi. However, there are
a couple of accidental ways to generate such names. This release tries to
make it easier to correct such mistakes by escaping slashes in several
places, allowing slashes in the t=info and t=delete commands, and in the
source (but not the target) of a t=rename command.

** Packaging

Tahoe's dependencies have been extended to require the "[secure_connections]"
feature from Foolscap, which will cause pyOpenSSL to be required and/or
installed. If OpenSSL and its development headers are already installed on
your system, this can occur automatically. Tahoe now uses pollreactor
(instead of the default selectreactor) to work around a bug between pyOpenSSL
and the most recent release of Twisted (8.1.0). This bug only affects unit
tests (hang during shutdown), and should not impact regular use.

The Tahoe source code tarballs now come in two different forms: regular and
"sumo". The regular tarball contains just Tahoe, nothing else. When building
from the regular tarball, the build process will download any unmet
dependencies from the internet (starting with the index at PyPI) so it can
build and install them. The "sumo" tarball contains copies of all the
libraries that Tahoe requires (foolscap, twisted, zfec, etc), so using the
"sumo" tarball should not require any internet access during the build
process. This can be useful if you want to build Tahoe while on an airplane,
a desert island, or other bandwidth-limited environments.

Similarly, allmydata.org now hosts a "tahoe-deps" tarball which contains the
latest versions of all these dependencies. This tarball, located at
http://allmydata.org/source/tahoe/deps/tahoe-deps.tar.gz, can be unpacked in
the tahoe source tree (or in its parent directory), and the build process
should satisfy its downloading needs from it instead of reaching out to PyPI.
This can be useful if you want to build Tahoe from a darcs checkout while on
that airplane or desert island.

Because of the previous two changes ("sumo" tarballs and the "tahoe-deps"
bundle), most of the files have been removed from misc/dependencies/ . This
brings the regular Tahoe tarball down to 2MB (compressed), and the darcs
checkout (without history) to about 7.6MB. A full darcs checkout will still
be fairly large (because of the historical patches which included the
dependent libraries), but a 'lazy' one should now be small.

The default "make" target is now an alias for "setup.py build", which itself
is an alias for "setup.py develop --prefix support", with some extra work
before and after (see setup.cfg). Most of the complicated platform-dependent
code in the Makefile was rewritten in Python and moved into setup.py,
simplifying things considerably.

Likewise, the "make test" target now delegates most of its work to "setup.py
test", which takes care of getting PYTHONPATH configured to access the tahoe
code (and dependencies) that gets put in support/lib/ by the build_tahoe
step. This should allow unit tests to be run even when trial (which is part
of Twisted) wasn't already installed (in this case, trial gets installed to
support/bin because Twisted is a dependency of Tahoe).

Tahoe is now compatible with the recently-released Python 2.6 , although it
is recommended to use Tahoe on Python 2.5, on which it has received more
thorough testing and deployment.

Tahoe is now compatible with simplejson-2.0.x . The previous release assumed
that simplejson.loads always returned unicode strings, which is no longer the
case in 2.0.x .

** Grid Management Tools

Several tools have been added or updated in the misc/ directory, mostly munin
plugins that can be used to monitor a storage grid.

The misc/spacetime/ directory contains a "disk watcher" daemon (startable
with 'tahoe start'), which can be configured with a set of HTTP URLs
(pointing at the wapi '/statistics' page of a bunch of storage servers),
and will periodically fetch disk-used/disk-available information from all the
servers. It keeps this information in an Axiom database (a sqlite-based
library available from divmod.org). The daemon computes time-averaged rates
of disk usage, as well as a prediction of how much time is left before the
grid is completely full.

The misc/munin/ directory contains a new set of munin plugins
(tahoe_diskleft, tahoe_diskusage, tahoe_doomsday) which talk to the
disk-watcher and provide graphs of its calculations.

To support the disk-watcher, the Tahoe statistics component (visible through
the wapi at the /statistics/ URL) now includes disk-used and disk-available
information. Both are derived through an equivalent of the unix 'df' command
(i.e. they ask the kernel for the number of free blocks on the partition that
encloses the BASEDIR/storage directory). In the future, the disk-available
number will be further influenced by the local storage policy: if that policy
says that the server should refuse new shares when less than 5GB is left on
the partition, then "disk-available" will report zero even though the kernel
sees 5GB remaining.

The 'tahoe_overhead' munin plugin interacts with an allmydata.com-specific
server which reports the total of the 'deep-size' reports for all active user
accounts, compares this with the disk-watcher data, to report on overhead
percentages. This provides information on how much space could be recovered
once Tahoe implements some form of garbage collection.

** Configuration Changes: single INI-format tahoe.cfg file

The Tahoe node is now configured with a single INI-format file, named
"tahoe.cfg", in the node's base directory. Most of the previous
multiple-separate-files are still read for backwards compatibility (the
embedded SSH debug server and the advertised_ip_addresses files are the
exceptions), but new directives will only be added to tahoe.cfg . The "tahoe
create-client" command will create a tahoe.cfg for you, with sample values
commented out. (ticket #518)

tahoe.cfg now has controls for the foolscap "keepalive" and "disconnect"
timeouts (#521).

tahoe.cfg now has controls for the encoding parameters: "shares.needed" and
"shares.total" in the "[client]" section. The default parameters are still
3-of-10.

The inefficient storage 'sizelimit' control (which established an upper bound
on the amount of space that a storage server is allowed to consume) has been
replaced by a lightweight 'reserved_space' control (which establishes a lower
bound on the amount of remaining space). The storage server will reject all
writes that would cause the remaining disk space (as measured by a '/bin/df'
equivalent) to drop below this value. The "[storage]reserved_space="
tahoe.cfg parameter controls this setting. (note that this only affects
immutable shares: it is an outstanding bug that reserved_space does not
prevent the allocation of new mutable shares, nor does it prevent the growth
of existing mutable shares).

** Other Changes

Clients now declare which versions of the protocols they support. This is
part of a new backwards-compatibility system:
http://allmydata.org/trac/tahoe/wiki/Versioning .

The version strings for human inspection (as displayed on the Welcome web
page, and included in logs) now includes a platform identifer (frequently
including a linux distribution name, processor architecture, etc).

Several bugs have been fixed, including one that would cause an exception (in
the logs) if a wapi download operation was cancelled (by closing the TCP
connection, or pushing the "stop" button in a web browser).

Tahoe now uses Foolscap "Incidents", writing an "incident report" file to
logs/incidents/ each time something weird occurs. These reports are available
to an "incident gatherer" through the flogtool command. For more details,
please see the Foolscap logging documentation. An incident-classifying plugin
function is provided in misc/incident-gatherer/classify_tahoe.py .

If clients detect corruption in shares, they now automatically report it to
the server holding that share, if it is new enough to accept the report.
These reports are written to files in BASEDIR/storage/corruption-advisories .

The 'nickname' setting is now defined to be a UTF-8 -encoded string, allowing
non-ascii nicknames.

The 'tahoe start' command will now accept a --syslog argument and pass it
through to twistd, making it easier to launch non-Tahoe nodes (like the
cpu-watcher) and have them log to syslogd instead of a local file. This is
useful when running a Tahoe node out of a USB flash drive.

The Mac GUI in src/allmydata/gui/ has been improved.


* Release 1.2.0 (2008-07-21)

** Security

This release makes the immutable-file "ciphertext hash tree" mandatory.
Previous releases allowed the uploader to decide whether their file would
have an integrity check on the ciphertext or not. A malicious uploader could
use this to create a readcap that would download as one file or a different
one, depending upon which shares the client fetched first, with no errors
raised. There are other integrity checks on the shares themselves, preventing
a storage server or other party from violating the integrity properties of
the read-cap: this failure was only exploitable by the uploader who gives you
a carefully constructed read-cap. If you download the file with Tahoe 1.2.0
or later, you will not be vulnerable to this problem. #491

This change does not introduce a compatibility issue, because all existing
versions of Tahoe will emit the ciphertext hash tree in their shares.

** Dependencies

Tahoe now requires Foolscap-0.2.9 . It also requires pycryptopp 0.5 or newer,
since earlier versions had a bug that interacted with specific compiler
versions that could sometimes result in incorrect encryption behavior. Both
packages are included in the Tahoe source tarball in misc/dependencies/ , and
should be built automatically when necessary.

** Web API

Web API directory pages should now contain properly-slash-terminated links to
other directories. They have also stopped using absolute links in forms and
pages (which interfered with the use of a front-end load-balancing proxy).

The behavior of the "Check This File" button changed, in conjunction with
larger internal changes to file checking/verification. The button triggers an
immediate check as before, but the outcome is shown on its own page, and does
not get stored anywhere. As a result, the web directory page no longer shows
historical checker results.

A new "Deep-Check" button has been added, which allows a user to initiate a
recursive check of the given directory and all files and directories
reachable from it. This can cause quite a bit of work, and has no
intermediate progress information or feedback about the process. In addition,
the results of the deep-check are extremely limited. A later release will
improve this behavior.

The web server's behavior with respect to non-ASCII (unicode) filenames in
the "GET save=true" operation has been improved. To achieve maximum
compatibility with variously buggy web browsers, the server does not try to
figure out the character set of the inbound filename. It just echoes the same
bytes back to the browser in the Content-Disposition header. This seems to
make both IE7 and Firefox work correctly.

** Checker/Verifier/Repairer

Tahoe is slowly acquiring convenient tools to check up on file health,
examine existing shares for errors, and repair files that are not fully
healthy. This release adds a mutable checker/verifier/repairer, although
testing is very limited, and there are no web interfaces to trigger repair
yet. The "Check" button next to each file or directory on the wapi page
will perform a file check, and the "deep check" button on each directory will
recursively check all files and directories reachable from there (which may
take a very long time).

Future releases will improve access to this functionality.

** Operations/Packaging

A "check-grid" script has been added, along with a Makefile target. This is
intended (with the help of a pre-configured node directory) to check upon the
health of a Tahoe grid, uploading and downloading a few files. This can be
used as a monitoring tool for a deployed grid, to be run periodically and to
signal an error if it ever fails. It also helps with compatibility testing,
to verify that the latest Tahoe code is still able to handle files created by
an older version.

The munin plugins from misc/munin/ are now copied into any generated debian
packages, and are made executable (and uncompressed) so they can be symlinked
directly from /etc/munin/plugins/ .

Ubuntu "Hardy" was added as a supported debian platform, with a Makefile
target to produce hardy .deb packages. Some notes have been added to
docs/debian.txt about building Tahoe on a debian/ubuntu system.

Storage servers now measure operation rates and latency-per-operation, and
provides results through the /statistics web page as well as the stats
gatherer. Munin plugins have been added to match.

** Other

Tahoe nodes now use Foolscap "incident logging" to record unusual events to
their NODEDIR/logs/incidents/ directory. These incident files can be examined
by Foolscap logging tools, or delivered to an external log-gatherer for
further analysis. Note that Tahoe now requires Foolscap-0.2.9, since 0.2.8
had a bug that complained about "OSError: File exists" when trying to create
the incidents/ directory for a second time.

If no servers are available when retrieving a mutable file (like a
directory), the node now reports an error instead of hanging forever. Earlier
releases would not only hang (causing the wapi directory listing to get
stuck half-way through), but the internal dirnode serialization would cause
all subsequent attempts to retrieve or modify the same directory to hang as
well. #463

A minor internal exception (reported in logs/twistd.log, in the
"stopProducing" method) was fixed, which complained about "self._paused_at
not defined" whenever a file download was stopped from the web browser end.


* Release 1.1.0 (2008-06-11)

** CLI: new "alias" model

The new CLI code uses an scp/rsync -like interface, in which directories in
the Tahoe storage grid are referenced by a colon-suffixed alias. The new
commands look like:
 tahoe cp local.txt tahoe:virtual.txt
 tahoe ls work:subdir

More functionality is available through the CLI: creating unlinked files and
directories, recursive copy in or out of the storage grid, hardlinks, and
retrieving the raw read- or write- caps through the 'ls' command. Please read
docs/CLI.txt for complete details.

** wapi: new pages, new commands

Several new pages were added to the web API:

 /helper_status : to describe what a Helper is doing
 /statistics : reports node uptime, CPU usage, other stats
 /file : for easy file-download URLs, see #221
 /cap == /uri : future compatibility

The localdir=/localfile= and t=download operations were removed. These
required special configuration to enable anyways, but this feature was a
security problem, and was mostly obviated by the new "cp -r" command.

Several new options to the GET command were added:

 t=deep-size : add up the size of all immutable files reachable from the directory
 t=deep-stats : return a JSON-encoded description of number of files, size
                distribution, total size, etc

POST is now preferred over PUT for most operations which cause side-effects.

Most wapi calls now accept overwrite=, and default to overwrite=true .

"POST /uri/DIRCAP/parent/child?t=mkdir" is now the preferred API to create
multiple directories at once, rather than ...?t=mkdir-p .

PUT to a mutable file ("PUT /uri/MUTABLEFILECAP", "PUT /uri/DIRCAP/child")
will modify the file in-place.

** more munin graphs in misc/munin/

  tahoe-introstats
  tahoe-rootdir-space
  tahoe_estimate_files
  mutable files published/retrieved
  tahoe_cpu_watcher
  tahoe_spacetime

** New Dependencies

  zfec 1.1.0
  foolscap 0.2.8
  pycryptopp 0.5
  setuptools (now required at runtime)

** New Mutable-File Code

The mutable-file handling code (mostly used for directories) has been
completely rewritten. The new scheme has a better API (with a modify()
method) and is less likely to lose data when several uncoordinated writers
change a file at the same time.

In addition, a single Tahoe process will coordinate its own writes. If you
make two concurrent directory-modifying wapi calls to a single tahoe node,
it will internally make one of them wait for the other to complete. This
prevents auto-collision (#391).

The new mutable-file code also detects errors during publish better. Earlier
releases might believe that a mutable file was published when in fact it
failed.

** other features

The node now monitors its own CPU usage, as a percentage, measured every 60
seconds. 1/5/15 minute moving averages are available on the /statistics web
page and via the stats-gathering interface.

Clients now accelerate reconnection to all servers after being offline
(#374). When a client is offline for a long time, it scales back reconnection
attempts to approximately once per hour, so it may take a while to make the
first attempt, but once any attempt succeeds, the other server connections
will be retried immediately.

A new "offloaded KeyGenerator" facility can be configured, to move RSA key
generation out from, say, a wapi node, into a separate process. RSA keys
can take several seconds to create, and so a wapi node which is being used
for directory creation will be unavailable for anything else during this
time. The Key Generator process will pre-compute a small pool of keys, to
speed things up further. This also takes better advantage of multi-core CPUs,
or SMP hosts.

The node will only use a potentially-slow "du -s" command at startup (to
measure how much space has been used) if the "sizelimit" parameter has been
configured (to limit how much space is used). Large storage servers should
turn off sizelimit until a later release improves the space-management code,
since "du -s" on a terabyte filesystem can take hours.

The Introducer now allows new announcements to replace old ones, to avoid
buildups of obsolete announcements.

Immutable files are limited to about 12GiB (when using the default 3-of-10
encoding), because larger files would be corrupted by the four-byte
share-size field on the storage servers (#439). A later release will remove
this limit. Earlier releases would allow >12GiB uploads, but the resulting
file would be unretrievable.

The docs/ directory has been rearranged, with old docs put in
docs/historical/ and not-yet-implemented ones in docs/proposed/ .

The Mac OS-X FUSE plugin has a significant bug fix: earlier versions would
corrupt writes that used seek() instead of writing the file in linear order.
The rsync tool is known to perform writes in this order. This has been fixed.