path: root/src/third_party/wiredtiger/NEWS.MONGODB

3.0-RC9, Feb 11 2015
--------------------

commit f445f3bf63e3fa096479c5963f75d91e02f9b616
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Wed Feb 11 17:49:04 2015 +1100

    If logs crossed the threshold size while we were taking a checkpoint, don't take another one immediately.

    Should help with SERVER-17206, where we saw two checkpoints every iteration.

commit 0d85a9716b786de5fc90c00fb31765ade8aefd1f
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Wed Feb 11 17:48:03 2015 +1100

    Check if a page was recently split before doing forced eviction.  We used to do this, but it got lost in a recent reorg of __wt_page_release.  This change should mean that after an in-memory split, application threads that are appending have time to move to the new page at the end of the tree, rather than getting stuck trying to force out a page.

    SERVER-16938, SERVER-17121

commit 545d064fd4cbb0b35dc536e772c60b26a193d3f2
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Wed Feb 11 17:45:26 2015 +1100

    When doing truncates, if we see a clean page in memory, try to evict it before truncating.  It should be cheap (just freeing memory), and if the eviction succeeds, the fast truncate code can kick in and mark the whole page deleted immediately.  Otherwise, truncate will mark each record on the page deleted, and the next time through will try to force that page out, which has to go through reconciliation to figure out that all of the records are deleted.

    SERVER-17157

commit 9bbb8595abd6ac962a0debf20a6cdcef73d83855
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Feb 10 16:09:36 2015 +1100

    Allow size-limited LSM trees to have Bloom filters, based on the normal configuration.

commit c040f84a765c7c39f03e173a555eb50f85e2e698
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Feb 10 15:58:02 2015 +1100

    Re-enable the global setting to disable LSM merges.
    
    refs #1657

commit 8f14899ba0ce5b1a8df689e3c68db9a68bfeee66
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Tue Feb 10 04:36:57 2015 +0000

    Fix a bug when re-opening an LSM tree.
    
    We could have attempted to update the last chunk that is already on disk.

commit d8263d46c1aa136d24ef194a8f62f0b02b92b9b0
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Feb 10 15:11:59 2015 +1100

    Improve LRU eviction of large pages: don't give up because a large page has recent updates: push on and try to do eviction and restore updates.

commit eb02caa2564a18e857d18ef4b3f25683b438111c
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Feb 10 15:01:37 2015 +1100

    Fix a local variable read when looking for pages evict racing with a page becoming dirty.
    
    refs #1660

commit da4d99e7ad57057a1b8397629d59a3c83c28de21
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Tue Feb 10 02:32:25 2015 +0000

    Fix a bug in LSM cursor open.
    
    The bug caused us to re-open more cursors than necessary in open.
    
    Related to fix: 439a655e

commit eec16c3052af107bbe57aaf547eb8e70d2de4966
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Tue Feb 10 00:26:37 2015 +0000

    Don't do LSM merge throttling if merges are disabled.

commit fcee4c8ce0b5db9d3340169deb321601b81f4a1b
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Mon Feb 9 14:04:02 2015 -0500

    Track splits during eviction by data-source as well as by connection.
    
    Don't double-count in-memory splits (we're incrementing cache_eviction_split in the underlying split-parent routine, not in the caller, so it's counting both normal and in-memory splits). Instead, cache_eviction_split is normal eviction splits, cache_inmem_split is in-memory
splits,

commit 3d1f9eace79b1aff84369d0caee245f9d6d96a60
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Mon Feb 9 06:25:36 2015 +0000

    Add a mode to LSM where we can limit the size of data in the tree.
    
    A feature request to allow for a high insert throughput into a table with a size limitation.
    
    Adds a new configuration option to WT_SESSION::create which is lsm=(chunk_count_limit=0), default to 0 which is disabled.
    
    Refs #1652

commit c63ba34c915d95c156aaf6c47a04fe6d361b91ad
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Feb 9 14:07:57 2015 +1100

    Don't double-count the on-disk header size when setting split boundaries.
    
    refs #1655

commit 152a0efdbd3ea66b142f52eed3c9224437143eec
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Mon Feb 9 12:25:10 2015 +1100

    Fix a bug in table create. A crash could cause recovery to break.
    
    Refs SERVER-17204
    
    The bug is that we weren't doing an fsync of the file after it was
    created. Recovery assumes that if there are records for a particular
    file, then it will exist on disk.

commit 4d50f5878073e582567848ae03ee506bb5058227
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Mon Feb 9 00:43:27 2015 +0000

    Remove obsolete updates every time we add a new update.
    
    We used to only do the check when the cache was full. That could lead to update chains growing immensely long, which is bad.
    
    Refs: #1647, SERVER-17195

commit e891a1f312850bcaaf5183f3fd2e091567044a96
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Sun Feb 8 17:59:49 2015 -0500

    If we find a "removed" page, clean or dirty, leaf or internal, fast-path eviction, it helps with append-only workloads.

commit ab2a7e9b397adf657081458e11f3dc472b10fd2b
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Thu Feb 5 15:54:01 2015 -0500

    There's a problem that went in in #1282, the key difference is that we are setting a split boundary at the end of the first page when there is more than a page worth of data. See also #1630 and #1631. This is an alternate approach to #1631: the real change is to fallthrough into the split case if the next item won't fit, callers of the split code can't handle failure from split, it has to create enough room for the next item to be entered into the buffer.
    
commit 90a352717a45a40d047b33c9fb00e7174e1ae04f
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Thu Feb 5 12:59:08 2015 -0500

    Initialize first_lsn if we have no logs. #1638

commit 7cc7efb75c90e778f9757b954ad3ec85912b58fd
Author: Don Anderson <dda@ddanderson.com>
Date:   Thu Feb 5 12:20:55 2015 -0500

    For wt printlog, make operations into a JSON array.  Without that, any tool that parses JSON is almost certain to merge successive values of repeated fields.
    Refs #1438.

commit 5bf11d893548804b890836a3d9ef4335c4319bb7
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Wed Feb 4 15:46:30 2015 -0500

    Add name_hash and hash bucket queues for fh and block. #1643

    SERVER-17078

commit 3b0c18f612c9cf4d61bc13785ff7125fa67b265a
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Wed Feb 4 06:27:07 2015 +0000

    Keep filling pages in reconciliation until we hit a boundary.
    
    This reverts some of a change for #1282 (without reverting the functionality in that change, AFAI can tell).
    
    Refs #1630


3.0-RC8 Feb 4 2015
------------------

commit d8b7f0b8db92a2ad6d64b95cafeaf20f0a90c8ce
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Wed Feb 4 16:00:11 2015 +1100

    Updates should always mark pages dirty (before checking for obsolete updates to free).

commit 0947f82e01587836277d911b147bc98eefb58507
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Tue Feb 3 10:28:00 2015 -0500

    Fixes for split cache accounting: multi-page splits weren't correctly accounting for the allocated WT_REFs, insert splits weren't correctly handling the new right-page's instantiated key (the parent needs to be incremened by both the left- and right-hand page's keys, and cannot assume it's the same size as the original WT_REF's key), insert splits need to increment the parent page's WT_REF size by two, not one.

commit df96addef5f3ffcb495b4bf54390cf3fd41ac924
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Feb 2 16:45:46 2015 +1100

    In recovery, track the maximum file ID in the metadata, regardless of whether there are any updates to roll forward.
    
    Previously, we tracked the largest file ID that was updated in the logs being rolled forward.  It was usually the case that the most recently created file was also the most recently updated, so that calculation usually worked and wasn't detected until the repro in SERVER-17142 that created tables, did a clean shutdown and restart, then created more tables and did a dirty shutdown and restart, which was rolling forward updates into the wrong tables.
    
    refs SERVER-17142, SERVER-17131(?)

commit 71f1559c91ed119082ebe42772da15e28915e1c8
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Feb 3 10:40:27 2015 +1100

    Start with clean trees so we can detect updates racing with sweep.
    
    Use a deleted ref to a leaf page that is created on first update, which is the same state the tree should be in if an empty leaf page is evicted.  The only wrinkle is that bulk operations expect to find a leaf page in the tree: create it explicitly before the bulk insert starts.  This was probably a bug before: if we had created a tree and kept it around for long enough with cache pressure before a bulk load started, the initial leaf page could have been evicted.

commit 8545c4b3b7f5ed306215c82f1ad1cbe3664f0c50
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Mon Feb 2 17:13:09 2015 -0500

    Make the "split to deepen the tree" configuration values real, stored
    in the metadata file, but they remain undocumented for now.

commit fb769dafee4aca91a60a28cd89317c268ac79d4f
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Mon Feb 2 16:36:22 2015 -0500

    WT_CELL_ADDR_DEL is 0, so we can't test vtype against 0 to know if it's
    been set or not. Reference SERVER-16866.

commit feca80738c1b9103b4faa04ddb0718344347f640
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Mon Feb 2 13:53:21 2015 -0500

    Wrap calls to functions using pindex with WT_WITH_PAGE_INDEX.

commit 23f2e1ba0680a2e8fa7a081f1b46e1ae2ab220d4
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Feb 2 17:18:03 2015 +1100

    Once we decide to force-evict a page, do it directly rather than setting read_gen and hoping page release agrees.

commit 5f00de07b5bad20a6ffb5ec7d412c4ca0b10c64f
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Feb 2 17:11:27 2015 +1100

    split_gen paranoia: always increment split_gen once per split, use the allocated value to check for existing readers.  Make sure that publishing a split_gen doesn't miss an update.

commit 10a74d6af4f945e34368bc5754797ef1d684d8ab
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Feb 2 16:52:34 2015 +1100

    If discarding a tree for sweep races with an update, give up the discard gracefully.
    
    refs #1618, SERVER-17048

commit a2d20dc49cac870977d91213a7fe6dabf362ce70
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Feb 2 16:45:46 2015 +1100

    In recovery, track the maximum file ID in the metadata, regardless of whether there are any updates to roll forward.
    
    Previously, we tracked the largest file ID that was updated in the logs being rolled forward.  It was usually the case that the most recently created file was also the most recently updated, so that calculation usually worked and wasn't detected until the repro in SERVER-17142 that created tables, did a clean shutdown and restart, then created more tables and did a dirty shutdown and restart, which was rolling forward updates into the wrong tables.
    
    refs SERVER-17142, SERVER-17131(?)

commit b0a828b262a2d0d3cf1361eed98aa25a1168a7a6
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Sat Jan 31 12:59:34 2015 -0500

    We no longer calculate allocation overhead per allocation chunk, revert the workaround the problem with page memory size calculations during splits where we forced the new parent page memory size to 5% of its current value; reference #1564, #1565. This fixes a problem where 5% of a page's memory footprint isn't large enough to accommodate the cache decrements that will be done in the page's future, leading to page underflow.
    
    Minor cleanups: we no longer calculate allocation overhead per allocation chunk, the macro WT_MEMSIZE_ADD is no longer needed at all, and WT_MEMSIZE_TRANSFER is renamed to WT_MEM_TRANSFER.

commit b640366c28fc66744e482c20c16973cb052aef8e
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Fri Jan 30 10:19:31 2015 -0500

    I believe we can race with pages being marked clean or dirty, which means we need to entirely divorce the page's dirty-byte count from page state: the page's dirty-byte count is just a value that tells us how many dirty bytes this page has contributed to the cache's total dirty-bytes count.  Sync the cache's information to that value when possible, but don't worry if we can't.

commit d02ea7246ec33e05b5fd60c499fea3ffe25c57d2
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Fri Jan 30 17:38:09 2015 +1100

    Use reads to measure cache pressure with shared caches.  We previously tracked writes, which were skewed by checkpoints.
    
    refs #1569

commit a326c3ba10e0d299944a650b890f8c2d851db34a
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Thu Jan 29 17:19:06 2015 -0500

    Simplify the cache calculation when a page is marked clean, use the page's dirty-byte count (which allows a race between the page being marked clean and being re-dirtied).
    
    This branch is still not correct, but appears to be able to run the CONFIG from #1582 without underflow for a much longer time).
    
    Reference #1605.

commit 1c60c4966dd68ea2bf05ebe62e3f1d8de1a7d033
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Thu Jan 29 14:33:36 2015 +1100

    Use a copy of the oldest transaction ID when sweeping cached overflow items.  Otherwise, we could free structures that are still hooked into the skiplist.
    
    refs #1615

commit 42724267278c64f5af68b281c9ee5742d1a56d72
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Wed Jan 28 10:31:51 2015 -0500

    Adjust logging yield and timeout values. #1610

commit ae102f4fe604f7fd547dece8ee138e8292d4f07c
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Wed Jan 28 17:40:06 2015 +1100

    Cleanup accounting for update lists when restoring updates to evicted pages.  Previously, we only accounted for the first update in a list.
    
    refs SERVER-16997

commit 4adf9c929b1b46f273239214b4e2757fcfdb8f96
Author: Mark Benvenuto <mark.benvenuto@mongodb.com>
Date:   Fri Jan 23 18:29:15 2015 -0500

    Windows Install Documentation

commit 8faa218d27e7f21091f0b51a973f27047db1d950
Author: Mark Benvenuto <mark.benvenuto@mongodb.com>
Date:   Tue Jan 27 13:47:24 2015 -0500

    MCI configuration update

commit 422cbb6cea5fa5be6829044215ae46dc10be5f70
Author: Mark Benvenuto <mark.benvenuto@mongodb.com>
Date:   Mon Jan 26 16:11:29 2015 -0500

    Add Install Target to SCons

commit 41e7ab083d79a650e93a34d09e01e973ca4100d9
Author: Mark Benvenuto <mark.benvenuto@mongodb.com>
Date:   Fri Jan 23 15:54:40 2015 -0500

    WiredTiger DLL support
    - Examples that only depend on public API use DLL now

commit 23b2493e75cd166075eaccdaef75c8beee4576db
Author: Mark Benvenuto <mark.benvenuto@mongodb.com>
Date:   Fri Jan 23 15:50:16 2015 -0500

    Scons Improvements
    - Added --enable-attach, --enable-diagnostic, --enable-verbose
    - Renamed --enable-swig to --enable-python for consistency + swig cleanup
    - Renamed wiredtiger static library to libwiredtiger.lib

commit 96ab0ef67eee20fa75fa6d52c97d98bc119b74ae
Author: Mark Benvenuto <mark.benvenuto@mongodb.com>
Date:   Thu Jan 15 15:07:01 2015 -0500

    Struct alignment and packing for MSVC

commit f3b65997ece52382eed91730416d5f919bea79cd
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Fri Jan 9 10:49:59 2015 -0500

    Fix huffman config and add huffman tests. #1536

2.8-RC7 Jan 27 2015
-------------------

commit 2b4172f17008ff36dbeb50cadaf4fb97fc859e4e
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Jan 27 15:50:09 2015 +1100

    Revert a workaround for splits during truncate.
    
    refs #1583, #1563, SERVER-16868

commit c2e108e2774ae79504579bcdca33f26fcff8cb07
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Jan 27 09:58:32 2015 +1100

    Change recovery to start from the checkpoint LSN in the metadata.  Don't assert that we see a checkpoint complete in the available log: if the application crashes in between syncing the metadata and writing the final checkpoint record, there is no need to roll anything forward but we don't have the final checkpoint.
    
    refs #1529

commit 2555e80d2020ba9833c436a22d1031f6c5778a64
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Mon Jan 26 14:31:25 2015 -0500

    Coverity CID 50796 (#1 of 1): Unintentional integer overflow (OVERFLOW_BEFORE_WIDEN)
    SERVER-17001

commit 1ce3b94d6e40d37a77e62eda500f286bd3816eb9
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Jan 26 15:56:25 2015 +1100

    Grab the table list lock while building the list of handles to checkpoint.
    
    This avoids a potential deadlock during compact operations and/or checkpoints with a target list (and an assertion about lock ordering in diagnostic builds).
    
    Note that nested locking is not ideal: the medium-term fix here is #1598.
    
    refs #1589, SERVER-16967

commit db3943563a87c3e4c42445ae9f3a07efacfdf4ac
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Jan 26 14:54:47 2015 +1100

    Free WT_REFs deleted by truncate.  We were doing this when a page spontaneously became empty, but not if the "fast truncate" code kicked in.
    
    refs SERVER-16921

commit 2063efb22c3c29b980f86f7fee77b6d03ba63ec1
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Fri Jan 23 16:21:06 2015 -0500

    Don't count pages evicted by a worker thread as an "application thread" eviction; add a new statistic to distinguish between the server itself evicting pages and the eviction worker threads evicting.
    
    Don't increment the eviction counters unless we find a page to evict, __evict_lru_pages() gets called a huge number of times in any workload where eviction is happening.
    
    Reference SERVER-16997, SERVER-17020.

commit 3abb99d58aaa46b0b3fcd338293a668422e3fcaf
Author: Mark Benvenuto <mark.benvenuto@mongodb.com>
Date:   Fri Jan 23 15:05:16 2015 -0500

    Close Thread Handle after thread join on Windows

commit 7d677aedfdcaa5458e900e556b662def460d0281
Author: Don Anderson <dda@ddanderson.com>
Date:   Fri Jan 23 08:52:13 2015 -0500

    Fix drop index on a newly opened session.
    Fix __wt_schema_open_index to return WT_NOTFOUND when opening a single index.  This fixes opening a cursor on a non-existant index.
    Refs #1567.

commit 3626081dff24e1448281d10658752b996897ca82
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Thu Jan 22 18:08:15 2015 -0500

    Add the cache_overhead configuration string to allow applications to configure their cache overhead.

commit 4843cd78e7f90937ebdb23f84fbd7c133a7e5256
Author: Don Anderson <dda@ddanderson.com>
Date:   Thu Jan 22 10:40:11 2015 -0500

    Prepend underscores to SWIG methods that could have name conflicts
    with WT internal names.  refs #1574.

commit ebb1d9402c0ce2911069b0437d71766b92c3dc12
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Wed Jan 21 12:57:20 2015 -0500

    Add log code to ensure write-no-sync. #1585

commit 44fa4fbff95d0689b20c3fe3f4a55202554f0d9f
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Mon Jan 19 14:25:39 2015 -0500

    Make compact more aggressive about finding blocks to move.

2.8-RC6, Jan 20 2015
--------------------

commit ab1d63d3aa2371ce53287c6c6c77833eb281a38a
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Tue Jan 20 15:37:46 2015 -0500

    Check for valid log_fh handle in wt_log_write. #1580

commit e2de971061abea9451e92d60f0870136c9c0af42
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Tue Jan 20 13:24:06 2015 -0500

    Quit page eviction immediately if we're trying to evict a tree, that is, an internal page that has other internal pages as children.

commit 6f3c5a933ef8ce79efc03a22a8c03526ffb2197b
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Mon Jan 19 12:38:24 2015 -0500

    The size of the file is decreasing each time, so compacting 10 times, at 10%, is not sufficient to drive a file to its smallest size. The right fix is probably to get better information from the block manager as to exactly how much the size of the file has decreased, but that's messy, especially when you consider the checkpoints requires to get to that smallest size. For now, do 100 compaction attempts instead of 10, and depend on the no-progress state and/or the compaction timeout to limit the amount of time we spend here.

commit 72172b088fba6769866aecabba8176303140f5c4
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Mon Jan 19 10:25:13 2015 -0500

    Coverity 1264611, memory leak (WT_RET that should have been a WT_ERR).

commit f61f984cf5241ac54bc2ea368c8c15e0cdfa91aa
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Sat Jan 17 22:25:02 2015 +0000

    Fix a deadlock opening statistics cursors.
    
    Refs #1575 and JIRA SERVER-16738

commit c5fa51a0f18e4117d9f7b841de86eb35af751264
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Sat Jan 17 09:07:04 2015 -0500

    Log close thread needs to wait for any outstanding writes. #1571

commit 9cd8120f491595ad6ac1c25c4b154ad6556b5fe7
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Sat Jan 17 09:49:09 2015 +1100

    Close the session for the log close server thread.  Fixes a leak detected by address sanitizer.

commit bd7364ea9a0542bee61db0a89e771faf814f6f53
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Fri Jan 16 21:03:55 2015 +0000

    Fix a bug in raw compression, where we were overflowing memory.
    
    We weren't growing the buffer enough when adding new items in.
    
    Refs SERVER-16664

commit 76addf73581c53f24462ab5fd724048aec36eaf3
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Sat Jan 17 05:48:03 2015 +1100

    Have WT_CURSOR::equals return 1 when cursors are equal, 0 when not.

commit b2841dfc015d9502e1def870605968144b935570
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Thu Jan 15 21:58:39 2015 -0500

    Add log thread to fsync and close log files. #1560

commit ebb93969ebfb6b9bb9dc60621933f2fbeac4b472
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Thu Jan 15 22:47:52 2015 +0000

    Don't do memory adjustments for the WT_REF's WT_ADDR structures, we don't do those adjustments in other places we set addresses.
    
    Workaround the problem with page memory size calculations during splits by forcing the new parent page memory size to 5% of its current value; reference #1564.
    
    Minor cleanups/renaming of the code instantiating the WT_REF structures during a tree-deepening split to clarify what's going on there.

commit e0031209183c880fb1a1b99399013e7675a75e88
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Fri Jan 16 09:26:34 2015 +1100

    Don't look at a page after it may be freed during split.
    
    During the process of doing a split we switch the ref to WT_REF_MEM - after which it's no longer safe to refer to the page. Shuffle the code so that we don't.
    
    SERVER-16868

commit b6d7532cbf823d537b8f1733169fe4de08173c09
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Thu Jan 15 16:55:00 2015 -0500

    Only advance sync_lsn to the end of our write.  We waited until the log->sync_lsn is advanced into our file.  It was a bug to set the sync_lsn to the current write_lsn as that can be too far ahead in a new log file when earlier log files aren't done yet.

commit 85851933a938c53dfa57d1621cab1a959db672eb
Author: Thomas Rueckstiess <thomas@rueckstiess.net>
Date:   Mon Dec 15 11:04:43 2014 +1100

    wtstats.py: removed python-nvd3 dependencies, rewriting with HTML template

commit 4c26d2324bae1d7030b0142d50dbd2ccf11ddeb6
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Thu Dec 11 19:32:50 2014 -0500

    Add support for a WT_CURSOR.reconfigure method, reference #1381.

2.8-RC5 Jan 15 2015
-------------------

commit 2e54a27683c5e2fd88918575383c76d3f60c3c78
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Thu Jan 15 07:17:21 2015 +1100

    Workaround a read-after-free involving eviction during truncates.  We were implicitly relying on first_dirty_txn to prevent pages being immediately force-evicted by truncate.  The bug is not fully understood, but this change restores the previous window where reads can complete before the page is evicted for real.
    
    refs BF-759

commit 8a1bfe3c35f0c1d90ea3e8e70c2aae8dff1fdbb3
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Wed Jan 14 15:07:39 2015 -0500

    Force log file closes to go in sequence. #1555

    Update the sync_lsn after sync'ing and closing an earlier log file and make sure archive doesn't try to remove a file that is still in use.

2.8-RC4 Dec 22 2014
-------------------

commit fbb96d94cdba9a28f5c5d737ce6c96543f3289f4
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Dec 22 15:59:46 2014 +1100

    Use the original page's first_dirty_txn when restoring updates to match what we do for in-memory splits.
    
    refs #1475

commit 4df72e8e20139ddf667e1f0d3b6b7dcf91deb006
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Dec 22 13:12:07 2014 +1100

    Avoid EBUSY returns to verify and salvage caused by checkpoints.  The "fix" involves blocking checkpoints while salvage or verify are in progress.
    
    refs #1404, SERVER-16457

commit 864f3495721b1311b49df19ee241bfca9adf0863
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Sun Dec 21 20:47:52 2014 -0500

    Make the cache bytes-written and bytes-read match, both should
    ignore compression. Reference #1505.

commit 995d6f8c26ae19013a1eb921fd871481ca643f47
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Mon Dec 22 12:42:46 2014 +1100

    Eviction should do update-restore if upper layers are trying to force out a page, regardless of its size.  Also, only look at ref->page after checking for exclusive access.  It is possible (but very unlikely) that a child page pointer could be replaced in the window where we are checking hazard pointers.

commit d4abc51ea61211f90f4b70a0486442264ededc27
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Mon Dec 22 10:43:04 2014 +1100

    Fix a bug where a custom extractor terminate was being called twice.
    
    Resolves issue #1503.
    
    Clarify the custom extractor and collator terminate documentation while I'm here.

commit 16972ef63de1283d85146530c35f522b053e2c1e
Author: Don Anderson <dda@ddanderson.com>
Date:   Fri Dec 19 09:56:47 2014 -0500

    Remove version numbering from the pkg-config file.  We don't create include files that are named by version.  Programs linked using -lwiredtiger will follow the symlink to wiredtiger-a.b.c.so, so their referred library name is forever stamped as wiredtiger-a.b.c.so, which won't conflict even when we ship wiredtiger-a.b.d.so.
    Refs #1458.

commit e913b0811114d65b543cd78824e809eb487fd330
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Fri Dec 19 17:19:58 2014 +1100

    Check that handles are not being walked by eviction before discarding.
    
    refs #1497

commit 0d21e437917bc7cf08393852a3074957431ea30e
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Fri Dec 19 15:44:15 2014 +1100

    Use the eviction server to write pages with READGEN_OLDEST set.
    
    Even before the eviction trigger has been reached. This should mean that we clear those pages out of cache earlier, and hopefully will save application threads from doing the evictions (at least sometimes).

commit e0adfba3c4011c49b73ff3e4a165a4a938f69cb3
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Fri Dec 19 12:23:21 2014 +1100

    Don't try to write leaves from the sweep server.  Previously, this was done without locking the handle, and so could race with LSM discarding a handle.
    
    We know the handle has been idle, so there is a good chance that a checkpoint has run since the last update and the write leaves was wasted effort.  If not, this change will keep the handle locked for longer preventing new opens, but it has been idle for a while anyway.
    
    refs #1495, #1497 (maybe)

commit 87328a8d5c1d4a201a1df604ba32a87863948bbb
Author: Mark Benvenuto <mark.benvenuto@mongodb.com>
Date:   Thu Dec 18 14:51:51 2014 -0500

    fix test_salvage on Windows

commit 1953776ada137f3deae50169bf889d2063b353d3
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Thu Dec 18 16:45:06 2014 +1100

    Don't try to set and clear session->split_gen in WT_INTL_FOREACH_BEGIN: there are too many of those loops to ensure that none of them skip clearing it.  Instead, make sure all calls are wrapped in WT_WITH_APAGE_INDEX.
    
    refs SERVER-16546

commit 235f747e2df80d9899497595a2b649e7d6df8601
Author: Mark Benvenuto <mark.benvenuto@mongodb.com>
Date:   Wed Dec 17 14:11:41 2014 -0500

    snprintf - Implement a custom version of snprintf match the truncation behavior of C99 standard snprintf until MSVC supports snprintf.

commit 857a6fd0c4b6b001c78cbbc507674e2129029dff
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Wed Dec 17 15:51:41 2014 -0500

    Check the block header checksum before we clear it, it should be the same as the passed-in checksum, and if just those 4 bytes are corrupted, we wouldn't noticed. SERVER-16457.

commit dfa706056c4a359f7f894047bc9e5399efcec776
Author: Don Anderson <dda@ddanderson.com>
Date:   Tue Dec 16 15:58:34 2014 -0500

    Some refactoring of python packing.  More checks for standalone unit tests.
    Refs #1429.

2.8-RC3 Dec 17 2014
-------------------

commit bb064847e1c45f2b396d3f65f4e08cd10f33ed6e
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Wed Dec 17 15:45:23 2014 +1100

    Detect write-write conflicts before no-overwrite cursors decide to skip an update.
    
    refs SERVER-16351

commit 91abf8e35d5246a653bd9615ffd9723d87999c38
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Wed Dec 17 13:36:02 2014 +1100

    Add support for none configuration string to log compressor.
    
    To be consistent with block compression configuration.

commit 5438fee4942b4dbf484799dad6e12e042d253e99
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Wed Dec 17 12:09:33 2014 +1100

    Return an error if a shared cache configuration is set, but not enabled.
    
    Check for a configuration via shared_cache=(size=).
    
    Refs #1487

commit 390a5b71b25492dc3030e908a65a11a04401852b
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Tue Dec 16 15:32:09 2014 -0500

    We documented that huffman_key and huffman_value took "none" as an argument, but they didn't.
    
    Worse, if key was set but not value, we'd set value anyway, free of charge, and if value was set but not key, we'd fail. I doubt this is a problem (it's pretty surprising if anyone would set key but not also set value).
    
    Use __wt_config_gets_none() to support the "none" setting, re-work the logic to ignore keys or values that are zero-length.
    
    Reference #1417.

commit 662e26eeb31f76f2c4aeebf6690c9056612de32e
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Tue Dec 16 14:20:01 2014 -0500

    Atomically create all log files and move them into place. #1482

commit 5c30d62dbf7ec0976d6ec4d2aed4ba272aadd499
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Mon Dec 15 19:41:05 2014 -0500

    If we race with the logging thread and get to __wt_logmgr_destroy() while __log_archive_once() is still using conn->log_path, we can free it out from under the running server. If there's a logging thread running, don't free conn->log_path until we've joined that thread.
    Reference #1480.

commit d77d35db407fd74c266bdb728b12c74fdab26ba2
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Mon Dec 15 18:33:54 2014 -0500

    Even if we don't track any overflow pages during our read of the file, we still have to process the list of pages looking for leaf pages that reference unavailable overflow pages, no overflow pages doesn't imply there are no references to overflow pages.

commit cedf8cfe69bf964629aab498feb20a0b1ab77bc0
Author: Don Anderson <dda@ddanderson.com>
Date:   Mon Dec 15 11:54:15 2014 -0500

    Fix use of 'compressed' flag for printlog.
    Added printlog call to test case for log compression.
    Refs #1472.

commit 3210b11cf7bfb79f3ed52cd1c17a13c644a82e7a
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Mon Dec 15 11:02:10 2014 -0500

    Fix memory leak.  Always free log_path. #1473

commit 097c61e5f3326bc71f2d645b3f539c5c6d1ae3fb
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Mon Dec 15 05:58:06 2014 +0000

    Allow printlog to work without recovery.
    
    It now works even if compression or a custom path are setup.

commit b8921272755ce66d09ab2a001745573420bd41ac
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Wed Dec 10 15:40:06 2014 -0500

    We can't use the corrected page size to calculate the buffer's space available, but we don't have a page size either. We do know how much space we added, so use that to increment the space available.
    
    The raw compression handler can no longer pass a page size buffer to the underlying compression function, because a single key/value object could be larger than the page size, instead, pass a buffer of the same length as the source to compress, whatever that is.

commit 402041727de02931be1dd385f3c970f31a53341c
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Wed Dec 10 10:53:27 2014 +1100

    Remove the min / max bounds on overflow sizes.

commit 5088ee53fce569915e8de8c168da50cff7991ec1
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Tue Dec 9 16:22:20 2014 -0500

    Separate the btree maximum key/value sizes from the underlying page size, reference #1282.
    
    Deprecate the internal_item_max and leaf_item_max configuration strings, replace with internal_key_max, leaf_key_max and leaf_value_max.
    
    Remove examples/c/ex_file.c (there's no real need for a "file" URI example, and it's easy to replace the one place the documentation used it).

commit e1e187e8fdfb48526f2a62e3f0f48072c30db53e
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Dec 16 10:29:08 2014 +1100

    Update the swept handles statistic any time we close the underlying handle, regardless of whether a session still references the data handle.
    
    reds #1460, #1466

commit 080b34fde5de97459c383c67ba93d9fdc88090a8
Author: Don Anderson <dda@ddanderson.com>
Date:   Fri Dec 12 11:13:22 2014 -0500

    Allow pruning scenarios with different limits for default vs. long runs.  The pruned scenario list now matches the original ordering.  Tests now print with both the scenario number and the scenario args.  This should make it easier to diagnose and debug problems that effect only certain scenarios.  Refs #1461.

commit d5b88e08e2f6e39d098cfff3c013f4aa035c88bc
Author: Don Anderson <dda@ddanderson.com>
Date:   Thu Dec 11 15:13:11 2014 -0500

    Changed python test suite to allow for shorter runs by default, with a --long (or -l) option for the complete runs.  Txn02 in particular now runs only a small number of scenarios by default.  Also added a @longtest("description") decorator for individual tests that can be marked to be run only under --long.  Refs #1461.

commit 2f37332e5bbd14823f0c78ad38672dbce074e87f
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Fri Dec 12 17:21:50 2014 +1100

    Ensure metadata table is open at start of checkpoints.

commit 2cb10882f4f7189a3c2de4d7e187117873fded32
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Fri Dec 12 17:07:11 2014 +1100

    Switch to stashing the metadata dhandle, not the btree.
    
    Also update checkpoint to use the saved handle instead of
    searching for it again.

commit 42c05161cf8cc74606b259ceeeb41dd38ea7fb4e
Author: Susan LoVerso <sue@wiredtiger.com>
Date:   Thu Dec 11 15:24:14 2014 -0500

    Use dhandle hash lists in more places.  Adjust sweep timings. #1460

commit 7fb6315e45e74f0bef0a04505018e1ab0b68d144
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Thu Dec 11 23:06:19 2014 +1100

    If LSM search_near finds a matching tombstone, step the whole LSM cursor next to find the closest key.  We can't step individual chunk cursors, or we could return a record that is deleted in a more recent chunk.
    
    MongoDB BF-694, BF-700

commit 5f6bbc898564aefb312255555abd34202cb98815
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Wed Dec 10 14:39:25 2014 +1100

    Track whether eviction is making progress regardless of whether the cache is 100% full.  Otherwise we can get into a tight loop.  Use the count of pages evicted rather than a flag, now that there are multiple eviction threads.

commit cbe9e9bdbc508f95076b8097d41bb4cc799eab1c
Author: Don Anderson <dda@ddanderson.com>
Date:   Tue Dec 9 19:11:52 2014 -0500

    Change timing to allow archive thread to complete on tests that
    do archive.  Since this can make each test run substantially longer,
    reduce the number of tests that are doing archive from ~4000 to
    something under 100. Refs #1452.

commit 62af85890179abb9fda17a619fcd5ae69fb369e0
Merge: b83bf08 b24c7af
Author: Michael Cahill <mjc@wiredtiger.com>
Date:   Wed Dec 10 11:03:48 2014 +1100

    Merge pull request #1449 from wiredtiger/lsm-switch-simplify
    
    Improve and simplify the LSM switch logic

commit 7e0f7d7b803f9af04ad10b2bec6ef5073aa79248
Author: Don Anderson <dda@ddanderson.com>
Date:   Tue Dec 9 16:04:48 2014 -0500

    SESSION->drop with "force" of nonexistant index/colgroup should be silent.
    Refs #1436.

commit be364821d75c0c42169d79c486fa582c777f7082
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Dec 9 15:58:28 2014 +1100

    Sweep old handles more aggressively:
    
    1. don't have checkpoint or other periodic operations like statistics logging keep old handles alive;
    2. don't wait for all sessions to empty the file from their cache before closing;
    3. only update the time of death from the sweep thread.

commit c96a4c954ccc73744f8a1fbcf2fea6debdfca018
Merge: cc8eb0b 2e332b9
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Tue Dec 9 14:41:36 2014 +1100

    Merge pull request #1443 from wiredtiger/cursor-open-optimize
    
    Cursor open optimize

commit 993c8ede8ff64eac9e87d1adcd39f8575039222b
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Tue Dec 9 13:21:13 2014 +1100

    Improve and simplify the LSM switch logic: it is fine to keep writing into a chunk while it is being switched, until either it hits the hard chunk size limit or a switch transaction ID is chosen that is larger than the writer's.  Fixes an assertion failure introduced by #1432 that could write to an old chunk after the switch completed.
    
    refs #1432, #1418

commit 5551461cd5f26249e4330c9f87b4945d7ec2bb34
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Mon Dec 8 18:01:31 2014 +1100

    If there are only two LSM worker threads, don't let the cache get full.
    
    Allow the first thread to do flushes as well as switches and drops if there are only two threads.
    
    Refs #1441, but this is a hang seen from that test/format configuration, not a segfault.

commit 8f06d6b79dabed54ad1e05515bbdb31e23c4e991
Author: Don Anderson <dda@ddanderson.com>
Date:   Fri Dec 5 14:22:41 2014 -0500

    Modify printlog output so that arbitrary strings are shown as using the JSON Unicode standard. refs #1438.

commit 68090796dea07e7b2d3d5bee8d69aafcd8febe16
Author: Alex Gorrod <alexg@wiredtiger.com>
Date:   Fri Dec 5 04:36:25 2014 +0000

    Avoid string comparisons when looking up tables.
    
    Save a hash value in the table so we can do integer comparisons rather than string comparisons.

commit 4de5e3a71bfad1c2a9ef1eccccdd45ec02fecba9
Author: Michael Cahill <michael.cahill@wiredtiger.com>
Date:   Fri Dec 5 13:01:51 2014 +1100

    Force eviction if we see many consecutive deletes when scanning through a page.  This fixes quadratic behavior in find-first+delete workloads.

commit bbced52c939e16ad5662b3a177cef3e52abddd6e
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Thu Dec 4 07:57:13 2014 -0500

    In the final close, continue and remove the handle no matter what errors we see, otherwise the handle-close code in __wt_conn_dhandle_discard() can become infinite loops, where we repeatedly attempt to close the same file handles. Reference #1434.

commit 46fa7f0b6397fe765c5e8c2853f9cd0b067bc808
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Wed Dec 3 13:46:01 2014 -0500

    Changes in #1204, #1288 mistakenly changed the values for some error defines, which breaks backward compatibility. (The WT_DEADLOCK error name sorted differently from WT_ROLLBACK, and we were assigning error values based on the sort order in a script.) Revert the change, and make sure it doesn't happen again.

commit 249e88485c75951a0584a7c7a8dd4b8f8b6a3382
Author: Keith Bostic <keith@wiredtiger.com>
Date:   Wed Dec 3 06:25:16 2014 -0500

    Support "none" in all configuration strings as an alternative to an empty string. Reference #1417.

commit 63d7c7869f8c2ab5a3e6ee935d1e37f21d40755f
Author: Don Anderson <dda@ddanderson.com>
Date:   Tue Dec 2 14:00:11 2014 -0500

    Added log compression.  When configured, we attempt to compress each log record.  Added printlog output to show before/after compression sizes.  Refs #1359.