dave
/
relstorage
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
relstorage/CHANGES.rst

888 lines
32 KiB
ReStructuredText
Raw Permalink Blame History

=========
Changes
=========
2.1.2 (unreleased)
==================
- Nothing changed yet.
2.1.1 (2019-01-07)
==================
- Avoid deleting attributes of DB driver modules we import. Fixes
:issue:`206` reported by Josh Zuech.
2.1.0 (2018-02-07)
==================
- Document that installing RelStorage from source requires a working
CFFI compilation environment. Fixes :issue:`187`, reported by
Johannes Raggam.
- Test with MySQL Connector/Python 8.0.6, up from 2.1.5. Note that
PyPy 5.8.0 is known to *not* work with MySQL Connector/Python
(although PyPy 5.6.0 did).
2.1a2 (2017-04-15)
==================
- Implemented the storage ``afterCompletion`` method, which allows
RelStorage storages to be notified of transaction endings for
transactions that don't call the two-phase commit API. This allows
resources to be used more efficiently because it prevents RDBMS
transactions from being held open.
Fixes: :issue:`147` (At least for ZODB 5.2.)
- Oracle: Fix two queries that got broken due to the performance work
in 2.1a1.
- MySQL: Workaround a rare issue that could lead to a ``TypeError``
when getting new OIDs. See :issue:`173`.
- The ``len`` of a RelStorage instance now correctly reflects the
approximate number of objects in the database. Previously it
returned a hardcoded 0. See :issue:`178`.
- MySQL: Writing blobs to the database is much faster and scales much
better as more blobs are stored. The query has been rewritten to use
existing primary key indexes, whereas before it used a table scan
due to deficiencies in the MySQL query optimizer. Thanks to Josh
Zuech and enfold-josh. See :issue:`175`.
2.1a1 (2017-02-01)
==================
- 3.6.0 final release is tested on CI servers.
- Substantial performance improvements for PostgreSQL, both on reading
and writing. Reading objects can be 20-40% faster. Writing objects
can be 15-25% faster (the most benefit will be seen by history-free
databases on PostgreSQL 9.5 and above). MySQL may have a (much)
smaller improvement too, especially for small transactions. This was
done through the use of prepared statements for the most important
queries and the new `'ON CONFLICT UPDATE'
<https://wiki.postgresql.org/wiki/What's_new_in_PostgreSQL_9.5#INSERT_..._ON_CONFLICT_DO_NOTHING.2FUPDATE_.28.22UPSERT.22.29>`_
syntax. See :pr:`157` and :issue:`156`.
- The umysqldb driver no longer attempts to automatically reconnect on
a closed cursor exception. That fails now that prepared statements
are in use. Instead, it translates the internal exception to one
that the higher layers of RelStorage recognize as requiring
reconnection at consistent times (transaction boundaries).
- Add initial support for the `MySQL Connector/Python
<https://dev.mysql.com/doc/connector-python/en/>`_ driver. See
:issue:`155`.
- Backport `ZODB #140
<https://github.com/zopefoundation/ZODB/pull/140>`_ to older
versions of ZODB. This improves write performance, especially in
multi-threaded scenarios, by up to 10%. See :pr:`160`.
- MySQL temporary tables now use the InnoDB engine instead of MyISAM.
See :pr:`162`.
2.0.0 (2016-12-23)
==================
- MySQL and Postgres now use the same optimized methods to get the
latest TID at transaction commit time as they do at poll time. This
is similar to :issue:`89`.
- MySQL now releases the commit lock (if acquired) during pre-pack
with GC of a history-free storage at the same time as PostgreSQL and
Oracle did (much earlier). Reported and initial fix provided in
:pr:`9` by jplouis.
2.0.0rc1 (2016-12-12)
=====================
- Writing persistent cache files has been changed to reduce the risk
of stale temporary files remaining. Also, files are kept open for a
shorter period of time and removed in a way that should work better
on Windows.
- RelStorage is now tested on Windows for MySQL and PostgreSQL thanks
to AppVeyor.
- Add support for Python 3.6.
2.0.0b9 (2016-11-29)
====================
- The MySQL adapter will now produce a more informative error if it
gets an unexpected result taking the commit lock. Reported by Josh
Zuech.
- Compatibility with transaction 2.0 on older versions of ZODB (prior
to the unreleased version that handles encoding meta data for us),
newer versions of ZODB (that do the encoding), while maintaining
compatibility with transaction 1.x. In particular, the ``history``
method consistently returns bytes for username and description.
- In very rare cases, persistent cache files could result in a corrupt
cache state in memory after loading them, resulting in
AttributeErrors until the cache files were removed and the instance
restarted. Reported in :issue:`140` by Carlos Sanchez.
2.0.0b8 (2016-10-02)
====================
- List CFFI in `setup_requires` for buildout users.
2.0.0b7 (2016-10-01)
====================
- Add the ability to limit the persistent cache files size. Thanks to
Josh Zuech for the suggestion, which led to the next change.
- Move the RelStorage shared cache to a `windowed-LFU with segmented
LRU
<http://highscalability.com/blog/2016/1/25/design-of-a-modern-cache.html>`_
instead of a pure LRU model. This can be a nearly optimal caching
strategy for many workloads. The caching code itself is also faster
in all tested cases.
It's especially helpful when using persistent cache files together
with a file size limit, as we can now ensure we write out the most
frequently useful data to the file instead of just the newest.
For more information see :issue:`127` and :pr:`128`. Thanks to Ben
Manes for assistance talking through issues related to the cache
strategy.
For write-heavy workloads, you may want to increase
``cache_delta_size_limit``.
The internal implementation details of the cache have been
completely changed. Only the ``StorageCache`` class remains
unchanged (though that's also an implementation class). CFFI is now
required, and support for PyPy versions older than 2.6.1 has been dropped.
- On CPython, use LLBTrees for the cache delta maps. This allows using
a larger, more effective size while reducing memory usage. Fixes :issue:`130`.
- Persistent cache files use the latest TID in the cache as the file's
modification time. This allows a more accurate choice of which file
to read at startup. Fixes :issue:`126`.
- Fix packing of history-preserving Oracle databases. Reported in
:issue:`135` by Peter Jacobs.
2.0.0b6 (2016-09-08)
====================
- Use ``setuptools.find_packages`` and ``include_package_data`` to
ensure wheels have all the files necessary. This corrects an issue
with the 2.0.0b5 release on PyPI. See :issue:`121` by Carlos Sanchez.
2.0.0b5 (2016-08-24)
====================
- Supporting new databases should be simpler due to a code
restructuring. Note that many internal implementation classes have
moved or been renamed.
- The umysqldb support handles query transformations more efficiently.
- umysqldb now raises a more informative error when the server sends
too large a packet.
.. note:: If you receive "Socket receive buffer full" errors, you
are likely experiencing `this issue <https://github.com/esnme/ultramysql/issues/34>`_ in ultramysql and
will need a patched version, such as the one provided in
`this pull request
<https://github.com/esnme/ultramysql/pull/61>`_.
- The local persistent cache file format has been changed to improve
reading and writing speed. Old files will be cleaned up
automatically. Users of the default settings could see improvements
of up to 3x or more on reading and writing.
- Compression of local persistent cache files has been disabled by
default (but there is still an option to turn it back on).
Operational experience showed that it didn't actually save that much
disk space, while substantially slowing down the reading and writing
process (2-4x).
- Add an option, ``cache-local-dir-read-count`` to limit the maximum
number of persistent local cache files will be used to populate a
storages's cache. This can be useful to reduce startup time if cache
files are large and workers have mostly similar caches.
2.0.0b4 (2016-07-17)
====================
- Add experimental support for umysqldb as a MySQL driver for Python
2.7. This is a gevent-compatible driver implemented in C for speed.
Note that it may not be able to store large objects (it has been
observed to fail for a 16M object---it hardcodes a
``max_allowed_packet`` of exactly 16MB for read and write buffers),
and has been observed to have some other stability issues.
2.0.0b3 (2016-07-16)
====================
- Add support for ZODB 5. RelStorage continues to run on ZODB 4 >=
4.4.2.
- Add support for tooling to help understand RelStorage cache
behaviour. This can help tune cache sizes and the choice to use
Memcached or not. See :issue:`106` and :pr:`108`.
- Fix a threading issue with certain database drivers.
2.0.0b2 (2016-07-08)
====================
Breaking Changes
----------------
- Support for cx_Oracle versions older than 5.0 has been dropped. 5.0
was released in 2008.
- Support for PostgreSQL 8.1 and earlier has been dropped. 8.2 is
likely to still work, but 9.0 or above is recommended. 8.2 was
released in 2006 and is no longer supported by upstream. The oldest
version still supported by upstream is 9.1, released in 2011.
Platform Support
----------------
- Using ZODB >= 4.4.2 (*but not 5.0*) is recommended to avoid
deprecation warnings due to the introduction of a new storage
protocol. The next major release of RelStorage will require ZODB
4.4.2 or above and should work with ZODB 5.0.
- Change the recommended and tested MySQL client for Python 2.7 away
from the unmaintained MySQL-python to the maintained mysqlclient
(the same one used by Python 3).
- PyMySQL now works and is tested on Python 3.
- A pure-Python PostgreSQL driver, pg8000, now works and is tested on
all platforms. This is a gevent-compatible driver. Note that it
requires a PostgreSQL 9.4 server or above for BLOB support.
- Support explicitly specifying the database driver to use. This can
be important when there is a large performance difference between
drivers, and more than one might be installed. (Also, RelStorage no
longer has the side-effect of registering ``PyMySQL`` as ``MySQLdb`` and
``psycopg2cffi`` as ``psycopg2``.) See :issue:`86`.
Bug Fixes
---------
- Memcache connections are explicitly released instead of waiting for
GC to do it for us. This is especially important with PyPy and/or
``python-memcached``. See :issue:`80`.
- The ``poll-interval`` option is now ignored and polling is performed
when the ZODB Connection requests it (at transaction boundaries).
Experience with delayed polling has shown it typically to do more
harm than good, including introducing additional possibilities for
error and leading to database performance issues. It is expected
that most sites won't notice any performance difference. A larger
discussion can be found in :issue:`87`.
Performance
-----------
- Support a persistent on-disk cache. This can greatly speed up
application warmup after a restart (such as when deploying new code).
Some synthetic benchmarks show an 8-10x improvement. See :issue:`92`
for a discussion, and see the options ``cache-local-dir`` and
``cache-local-dir-count``.
- Instances of :class:`.RelStorage` no longer use threading locks by
default and hence are not thread safe. A ZODB :class:`Connection
<ZODB.interfaces.IConnection>` is documented as not being
thread-safe and must be used only by a single thread at a time.
Because RelStorage natively implements MVCC, each Connection has a
unique storage object. It follows that the storage object is used
only by a single thread. Using locks just adds unneeded overhead to
the common case. If this is a breaking change for you, please open
an issue. See :pr:`91`.
- MySQL uses (what should be) a slightly more efficient poll query.
See :issue:`89`.
- The in-memory cache allows for higher levels of concurrent
operation via finer-grained locks. For example, compression and
decompression are no longer done while holding a lock.
- The in-memory cache now uses a better approximation of a LRU
algorithm with less overhead, so more data should fit in the same
size cache. (For best performance, CFFI should be installed; a
warning is generated if that is not the case.)
- The in-memory cache is now smart enough not to store compressed
objects that grow during compression, and it uses the same
compression markers as zc.zlibstorage to avoid double-compression.
It can also gracefully handle changes to the compression format in
persistent files.
2.0.0b1 (2016-06-28)
====================
Breaking Changes
----------------
- Update the ZODB dependency from ZODB3 3.7.0 to ZODB 4.3.1. Support
for ZODB older than 3.10 has been removed; ZODB 3.10 may work, but
only ZODB 4.3 is tested.
- Remove support for Python 2.6 and below. Python 2.7 is now required.
Platform Support
----------------
- Add support for PyPy on MySQL and PostgreSQL using PyMySQL and
psycopg2cffi respectively. PyPy can be substantially faster than
CPython in some scenarios; see :pr:`23`.
- Add initial support for Python 3.4+ for MySQL (using mysqlclient), PostgreSQL,
and Oracle.
Bug Fixes
---------
- Fixed ``loadBefore`` of a deleted/undone object to correctly raise a
POSKeyError instead of returning an empty state. (Revealed by
updated tests for FileStorage in ZODB 4.3.1.)
- Oracle: Packing should no longer produce LOB errors. This partially
reverts the speedups in 1.6.0b2. Reported in :issue:`30` by Peter
Jacobs.
- :meth:`.RelStorage.registerDB` and :meth:`.RelStorage.new_instance`
now work with storage wrappers like zc.zlibstorage. See :issue:`70`
and :issue:`71`.
Included Utilities
------------------
- zodbconvert: The ``--incremental`` option is supported with a
FileStorage (or any storage that implements
``IStorage.lastTransaction()``) as a destination, not just
RelStorages.
- zodbconvert: The ``--incremental`` option works correctly with a
RelStorage as a destination. See :pr:`22`. With contributions by
Sylvain Viollon, Mauro Amico, and Peter Jacobs. Originally reported
by Jan-Wijbrand Kolman.
- PostgreSQL: ``zodbconvert --clear`` should be much faster when the
destination is a PostgreSQL schema containing lots of data. *NOTE*:
There can be no other open RelStorage connections to the destination,
or any PostgreSQL connection in general that might be holding locks
on the RelStorage tables, or ``zodbconvert`` will block indefinitely
waiting for the locks to be released. Partial fix for :issue:`16`
reported by Chris McDonough.
- ``zodbconvert`` and ``zodbpack`` use :mod:`argparse` instead of
:mod:`optparse` for command line handling.
Performance
-----------
- MySQL: Use the "binary" character set to avoid producing "Invalid
utf8 character string" warnings. See :issue:`57`.
- Conflict resolution uses the locally cached state instead of
re-reading it from the database (they are guaranteed to be the
same). See :issue:`38`.
- Conflict resolution reads all conflicts from the database in one
query, instead of querying for each individual conflict. See
:issue:`39`.
- PostgreSQL no longer encodes and decodes object state in Base64
during database communication thanks to database driver
improvements. This should reduce network overhead and CPU usage for
both the RelStorage client and the database server. psycopg2 2.4.1
or above is required; 2.6.1 or above is recommended. (Or
psycopg2cffi 2.7.4.)
- PostgreSQL 9.3: Support ``commit-lock-timeout``. Contributed in :pr:`20`
by Sean Upton.
Other Enhancements
------------------
- Raise a specific exception when acquiring the commit lock
(:exc:`~relstorage.adapters.interfaces.UnableToAcquireCommitLockError`) or pack
lock (:exc:`~relstorage.adapters.interfaces.UnableToAcquirePackUndoLockError`)
fails. See :pr:`18`.
- ``RelStorage.lastTransaction()`` is more consistent with FileStorage
and ClientStorage, returning a useful value in more cases.
- Oracle: Add support for getting the database size. Contributed in
:pr:`21` by Mauro Amico.
- Support :class:`ZODB.interfaces.IExternalGC` for history-free
databases, allowing multi-database garbage collection with
``zc.zodbdgc``. See :issue:`47`.
Project Details
---------------
- Travis CI is now used to run RelStorage tests against MySQL and
PostgreSQL on every push and pull request. CPython 2 and 3 and PyPy
are all tested with the recommended database drivers.
- Documentation has been reorganized and moved to `readthedocs
<http://relstorage.readthedocs.io>`_.
- Updated the buildout configuration to just run relstorage tests and
to select which databases to use at build time.
1.6.1 (2016-08-30)
==================
- Tests: Basic integration testing is done on Travis CI. Thanks to
Mauro Amico.
- ``RelStorage.lastTransaction()`` is more consistent with FileStorage
and ClientStorage, returning a useful value in more cases.
- zodbconvert: The ``--incremental`` option is supported with a
FileStorage (or any storage that implements
``IStorage.lastTransaction()``) as a destination, not just
RelStorages.
- zodbconvert: The ``--incremental`` option is supported with a
RelStorage as a destination. See :pr:`22`. With contributions by
Sylvain Viollon, Mauro Amico, and Peter Jacobs. Originally reported
by Jan-Wijbrand Kolman.
- Oracle: Packing should no longer produce LOB errors. This partially
reverts the speedups in 1.6.0b2. Reported in :issue:`30` by Peter
Jacobs.
1.6.0 (2016-06-09)
==================
- Tests: Use the standard library doctest module for compatibility
with newer zope.testing releases.
1.6.0b3 (2014-12-08)
====================
- Packing: Significantly reduced the RAM consumed by graph traversal during
the pre_pack phase. (Tried several methods; encoded 64 bit IISets turned
out to be the most optimal.)
1.6.0b2 (2014-10-03)
====================
- Packing: Used cursor.fetchmany() to make packing more efficient.
1.6.0b1 (2014-09-04)
====================
- The local cache is now more configurable and uses ``zlib`` compression
by default.
- Added support for ``zodburi``, which means you can open a storage
using "postgres:", "mysql:", or "oracle:" URIs.
- Packing: Reduced RAM consumption while packing by using IIBTree.Set
instead of built-in set objects.
- MySQL 5.5: The test suite was freezing in checkBackwardTimeTravel. Fixed.
- Added performance metrics using the perfmetrics package.
- zodbconvert: Add an --incremental option to the zodbconvert script,
letting you convert additional transactions at a later date, or
update a non-live copy of your database, copying over missing
transactions.
- Replication: Added the ro-replica-conf option, which tells RelStorage
to use a read-only database replica for load connections. This makes
it easy for RelStorage clients to take advantage of read-only
database replicas.
- Replication: When the database connection is stale (such as when
RelStorage switches to an asynchronous replica that is not yet up to
date), RelStorage will now raise ReadConflictError by default.
Ideally, the application will react to the error by transparently
retrying the transaction, while the database gets up to date. A
subsequent transaction will no longer be stale.
- Replication: Added the revert-when-stale option. When this option is
true and the database connection is stale, RelStorage reverts the
ZODB connection to the stale state rather than raise
ReadConflictError. This option is intended for highly available,
read-only ZODB clients. This option would probably confuse users of
read-write ZODB clients, whose changes would sometimes seem to be
temporarily reverted.
- Caching: Use the database name as the cache-prefix by default. This
will hopefully help people who accidentally use a single memcached for
multiple databases.
- Fixed compatibility with persistent 4.0.5 and above.
1.5.1 (2011-11-12)
==================
- Packing: Lowered garbage collection object reference finding log level to
debug; this stage takes mere seconds, even in large sites, but could produce
10s of thousands of lines of log output.
- RelStorage was opening a test database connection (and was leaving it
idle in a transaction with recent ZODB versions that support
IMVCCStorage.) RelStorage no longer opens that test connection.
- zodbconvert: Avoid holding a list of all transactions in memory.
- Just after installing the database schema, verify the schema was
created correctly. This affects MySQL in particular.
1.5.0 (2011-06-30)
==================
- PostgreSQL: Fixed another minor compatibility issue with PostgreSQL 9.0.
Packing raised an error when the client used old an version of libpq.
- Delete empty transactions in batches of 1000 rows instead of all in one
go, to prevent holding the transaction lock for longer than absolutely
necessary.
- Oracle: Fix object reference downloading performance for large RelStorage
databases during the garbage collection phase of a pack.
- Oracle, PostgreSQL: Switch to storing ZODB blob in chunks up to 4GB
(the maximum supported by cx_Oracle) or 2GB (PostgreSQL maximum blob size)
to maximize blob reading and writing performance.
The PostgreSQL blob_chunk schema changed to support this, see
notes/migrate-to-1.5.txt to update existing databases.
- zodbconvert: When copying a database containing blobs, ensure the source
blob file exists long enough to copy it.
1.5.0b2 (2011-03-02)
====================
- Better packing based on experience with large databases. Thanks
to Martijn Pieters!
- Added more feedback to the packing process. It'll now report
during batch commit how much of the total work has been
completed, but at most every .1% of the total number of
transactions or objects to process.
- Renamed the --dry-run option to --prepack and added a
--use-prepack-state to zodbpack. With these 2 options the
pre-pack and pack phases can be run separately, allowing re-use
of the pre-pack analysis data or even delegating the pre-pack
phase off to a separate server.
- Replaced the packing duty cycle with a nowait locking strategy.
The pack operation will now request the commit lock but pauses if
it is already taken. It releases the lock after every batch
(defaulting to 1 second processing). This makes the packing
process faster while at the same time yielding to regular ZODB
commits when busy.
- Do not hold the commit lock during pack cleanup while deleting
rows from the object reference tables; these tables are
pack-specific and regular ZODB commits never touch these.
- Added an option to control schema creation / updating on startup.
Setting the ``create-schema`` option to false will let you use
RelStorage without a schema update.
- Fixed compatibility with PostgreSQL 9.0, which is capable of
returning a new 'hex' type to the client. Some builds of psycopg2
return garbage or raise an error when they see the new type. The fix
was to encode more SQL query responses using base 64.
- With the new shared-blob-dir option set to false, it was possible
for a thread to read a partially downloaded blob. Fixed. Thanks for
the report from Maurits van Rees.
- Support for "shared-blob-dir false" now requires ZODB 3.9 or better.
The code in the ZODB 3.8 version of ZODB.blob is not compatible with
BlobCacheLayout, leading to blob filename collisions.
1.5.0b1 (2011-02-05)
====================
- Added a state_size column to object_state, making it possible
to query the size of objects without loading the state. The new
column is intended for gathering statistics. A schema migration
is required.
- Added more logging during zodbconvert to show that something is
happening and give an indication of how far along the process is.
- Fixed a missing import in the blob cache cleanup code.
- Added a --dry-run option to zodbpack.
- Replaced the graph traversal portion of the pack code with
a more efficient implementation using Python sets (instead of SQL).
The new code is much faster for packing databases with deeply
nested objects.
1.5.0a1 (2010-10-21)
====================
- Added an option to store ZODB blobs in the database. The option is
called "shared-blob-dir" and it behaves very much like the ZEO
option of the same name. Blobs stored in the database are broken
into chunks to reduce the impact on RAM.
- Require setuptools or distribute. Plain distutils is not sufficient.
1.4.2 (2011-02-04)
==================
- Fixed compatibility with ZODB 3.10. As reported by JĂźrgen Herrmann,
there was a problem with conflict errors. The RelStorage patch of the
sync() method now works with ZODB 3.10.
- Fixed a bug in packing history-free databases. If changes were
made to the database during the pack, the pack code could delete
too many objects. Thanks to Chris Withers for writing test code
that revealed the bug. A schema migration is required for history-free
databases; see notes/migration-to-1.4.txt.
- Enabled logging to stderr in zodbpack.
1.4.1 (2010-10-21)
==================
- Oracle: always connect in threaded mode. Without threaded mode,
clients of Oracle 11g sometimes segfault.
1.4.0 (2010-09-30)
==================
- Made compatible with ZODB 3.10.0b7.
- Enabled ketama and compression in pylibmc_wrapper. Both options
are better for clusters. [Helge Tesdal]
- Oracle: Use a more optimal query for POSKeyError logging. [Helge Tesdal]
- Fixed a NameError that occurred when getting the history of an
object where transaction extended info was set. [Helge Tesdal]
1.4.0c4 (2010-09-17)
====================
- Worked around an Oracle RAC bug: apparently, in a RAC environment,
the read-only transaction mode does not isolate transactions in the
manner specified by the documentation, so Oracle users now have to
use serializable isolation like everyone else. It's slower but more
reliable.
- Use the client time instead of the database server time as a factor
in the transaction ID. RelStorage was using the database server time
to reduce the need for synchronized clocks, but in practice, that
policy broke tests and did not really avoid the need to synchronize
clocks. Also, the effect of unsynchronized clocks is predictable
and manageable: you'll get bunches of transactions with sequential
timestamps.
- If the database returns an object from the future (which should never
happen), generate a ReadConflictError, hopefully giving the application
a chance to recover. The most likely causes of this are a broken
database and threading bugs.
1.4.0c3 (2010-07-31)
====================
- Always update the RelStorage cache when opening a database connection for
loading, even when no ZODB Connection is using the storage. Otherwise,
code that used the storage interface directly could cause the cache
to fall out of sync; the effects would be seen in the next
ZODB.Connection.
- Added a ZODB monkey patch that passes the "force" parameter to the
sync method. This should help the poll-interval option do its job
better.
1.4.0c2 (2010-07-28)
====================
- Fixed a subtle bug in the cache code that could lead to an
AssertionError indicating a cache inconsistency. The inconsistency
was caused by after_poll(), which was ignoring the randomness of
the order of the list of recent changes, leading it to sometimes
put the wrong transfer ID in the "delta_after" dicts. Also expanded
the AssertionError with debugging info, since cache inconsistency
can still be caused by database misconfiguration and mismatched
client versions.
- Oracle: updated the migration notes. The relstorage_util package
is not needed after all.
1.4.0c1 (2010-06-19)
====================
- History-preserving storages now replace objects on restore instead of
just inserting them. This should solve problems people were
having with the zodbconvert utility.
- Oracle: call the DBMS_LOCK.REQUEST function directly instead of using
a small package named ``relstorage_util``. The ``relstorage_util``
package was designed as a secure way to access the DBMS_LOCK package,
but the package turned out to be confusing to DBAs and provided no
real security advantage. People who have already deployed
RelStorage 1.4.x on Oracle need to do the following:
GRANT EXECUTE ON DBMS_LOCK TO <zodb_user>;
You can also drop the ``relstorage_util`` package. Keep the
``relstorage_op`` package.
- Made compatible with ZODB 3.10.
- MySQL: specify the transaction isolation mode for every connection,
since the default is apparently not necessarily "read committed"
anymore.
1.4.0b3 (2010-02-02)
====================
- Auto-reconnect in new_oid().
1.4.0b2 (2010-01-30)
====================
- Include all test subpackages in setup.py.
- Raise an error if MySQL reverts to MyISAM rather than using the InnoDB
storage engine.
1.4.0b1 (2009-11-17)
====================
- Added the keep-history option. Set it to false to keep no history.
(Packing is still required for garbage collection and blob deletion.)
- Added the replica-conf and replica-timeout options. Set replica-conf
to a filename containing the location of database replicas. Changes
to the file take effect at transaction boundaries.
- Expanded the option documentation in README.txt.
- Revised the way RelStorage uses memcached. Minimized the number of
trips to both the cache server and the database.
- Added an in-process pickle cache that serves a function similar to the
ZEO cache.
- Added a wrapper module for pylibmc.
- Store operations now use multi-insert and multi-delete SQL
statements to reduce the effect of network latency.
- Renamed relstorage.py to storage.py to overcome import issues.
Also moved the Options class to options.py.
- Updated the patch for ZODB 3.7 and 3.8 to fix an issue with
blobs and subtransactions.
- Divided the implementation of database adapters into many small
objects, making the adapter code more modular. Added interfaces
that describe the duties of each part.
- Oracle: Sped up restore operations by sending short blobs inline.
- Oracle: Use a timeout on commit locks. This requires installation
of a small PL/SQL package that can access DBMS_LOCK. See README.txt.
- Oracle: Used PL/SQL bulk insert operations to improve write
performance.
- PostgreSQL: use the documented ALTER SEQUENCE RESTART WITH
statement instead of ALTER SEQUENCE START WITH.
- Moved MD5 sum computation to the adapters so they can choose not
to use MD5.
- Changed loadSerial to load from the store connection only if the
load connection can not provide the object requested.
- Stopped wrapping database disconnect exceptions. Now the code
catches and handles them directly.
- Use the store connection rather than the load connection for OID
allocation.
- Detect and handle backward time travel, which can happen after
failover to an out-of-date asynchronous slave database. For
simplicity, invalidate the whole ZODB cache when this happens.
- Replaced the speed test script with a separately distributed package,
``zodbshootout``.
- Added the ``zodbpack`` script.
1.3.0b1 (2009-09-04)
====================
- Added support for a blob directory. No BlobStorage wrapper is needed.
Cluster nodes will need to use a shared filesystem such as NFS or
SMB/CIFS.
- Added the blob-dir parameter to the ZConfig schema and README.txt.
1.2.0 (2009-09-04)
==================
- In Oracle, trim transaction descriptions longer than 2000 bytes.
- When opening the database for the first time, don't issue a warning
about the inevitable POSKeyError on the root OID.
- If RelStorage tries to unpickle a corrupt object state during packing,
it will now report the oid and tid in the log.
1.2.0b2 (2009-05-05)
====================
- RelStorage now implements IMVCCStorage, making it compatible with
ZODB 3.9.0b1 and above.
- Removed two-phase commit support from the PostgreSQL adapter. The
feature turned out to be unnecessary.
- Added MySQL 5.1.34 and above to the list of supportable databases.
- Fixed minor test failures under Windows. Windows is now a supportable
platform.
Information about older releases can be found :doc:`here <HISTORY>`.
Reference in New Issue View Git Blame Copy Permalink