diff options
| author | Michael Cahill <michael.cahill@wiredtiger.com> | 2015-03-27 10:53:09 +1100 |
|---|---|---|
| committer | Michael Cahill <michael.cahill@wiredtiger.com> | 2015-03-27 10:53:09 +1100 |
| commit | d862d7a60b6196e49bca741e90b6b28041ec2358 (patch) | |
| tree | 076495ca0a088d3de7d5df00f96efbaee3bf774f | |
| parent | 3e37e1fca16f135a56068996bd37e28165cef0dc (diff) | |
| download | mongodb-3.0.2.tar.gz | |
Re-add the NEWS.MONGODB file, update for 3.0.2mongodb-3.0.2
| -rw-r--r-- | NEWS.MONGODB | 1702 |
1 files changed, 1702 insertions, 0 deletions
diff --git a/NEWS.MONGODB b/NEWS.MONGODB new file mode 100644 index 00000000000..a24eec18966 --- /dev/null +++ b/NEWS.MONGODB @@ -0,0 +1,1702 @@ +3.0.2, Mar 26 2015 +------------------ + +commit 17bd2356a5b17893e626749be399fb8fda23db1e +Author: Mark Benvenuto <mark.benvenuto@mongodb.com> +Date: Fri Mar 20 01:07:59 2015 -0400 + + Use beginthreadex, and ensure we use the correct C calling conventions in all callbacks + +commit 269e847ad64dd12dfcadb58f84f905069e5b8dce +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Wed Mar 25 15:47:46 2015 +1100 + + Split sweep into two passes: one that walks the handle list without locking and finds candidate handles to close, and a second pass holding the handle list lock that removes dead handles. + + refs #1814, #1811, #1808 + +commit 065a435f6b1d1b3fcb640d59c3109d0c2e24d308 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Tue Mar 24 16:53:28 2015 +1100 + + Add statistics to track internal/overflow/leaf pages in cache. + + It is often useful to know if a cache is full of internal pages. I could have added page count tracking, but byte count tracking seems more useful (enough to justify the extra performance overhead). + +commit f4616895f1ac35aaa50fc1a68e0c6dd9e0cf7717 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Mon Mar 23 17:00:15 2015 +1100 + + If a walk for pages to evict ends on a page requiring forced eviction, don't keep it pinned. + + refs SERVER-16662, SERVER-17382 + +commit 49ddcca712db3a50c390f48b518f0835f28bc9d4 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Sat Mar 21 16:17:51 2015 -0400 + + Separate out the code to flush a file handle and rename the file into place. + +commit a0e88c41781c5b7ea0155fb57a58c91a964af4a3 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Fri Mar 20 19:08:45 2015 -0400 + + If the underlying FILE was opened for writing, flush and fsync it during close. + +commit 8d918f0ef8588056ecf729e72ffdd8bc0a79fd6c +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Fri Mar 20 22:54:13 2015 +1100 + + Fix a race closing eviction helper threads: close the main eviction thread first, so the number of helper threads can be read safely. + + Partial fix for #1698 + +commit 294b0bce296bebf9790418e5575f59be5bec000c +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Fri Mar 20 17:19:28 2015 +1100 + + Only align buffers on Linux if direct I/O is configured. Clarify why we round up the allocation size: it usually won't matter because we allocate multiples of the alignment size anyway. + +commit ee7456799277e19ff4771de79ac4eafac43aff9b +Author: Mark Benvenuto <mark.benvenuto@mongodb.com> +Date: Thu Mar 19 16:16:15 2015 -0400 + + Visual Studio 2015 Preview support + +commit 9c60b462a25a36915207dde240579c5aa673f2c7 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Fri Mar 20 00:09:42 2015 +0000 + + Update the eviction server to not set stuck if it isn't full. + + When only looking for pages to force out, the cache isn't really stuck. The trouble is that if we set stuck the eviction server doesn't clear walks which can lead to it always holding a reference t othe page we (really) want to evict. + + refs #1777 + +commit f0138353697d8706bd1f26e83b9fff8f4e83af8d +Author: Keith Bostic <keith@wiredtiger.com> +Date: Thu Mar 19 13:04:17 2015 -0400 + + Create the base configuration file in a temporary file and rename it into place so a crash can't result in a corrupted base configuration file. + + refs #1775, #1776, SERVER-17571 + +commit 2e15cd6fc322c90c763394c52056bfebd4153aad +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Thu Mar 19 10:25:31 2015 +1100 + + Avoid EBUSY errors from truncate caused by checkpoints. + + refs #1643 + +commit 3188352d623de85803db3dd6e5b5188822e2db4f +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Wed Mar 18 13:00:53 2015 +1100 + + If the session-level attempt to lock a handle gets EBUSY, fall back to the slow path. The sweep server relies on there being a retry loop if an exclusive operation such as verify conflicts with a sweep. + + refs #1404 + +commit 5d8641568ac27e21d6671cb646a818708eb0aa28 +Author: Susan LoVerso <sue@wiredtiger.com> +Date: Tue Mar 17 15:38:57 2015 -0400 + + Set the end of the log if we find a zero hole. + + refs #1766, SERVER-17569, SERVER-17613 + +commit f38e325f87d94f178c932328632857361fc3c92c +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Tue Mar 17 04:30:11 2015 +0000 + + Implement review feedback. The new configuration strings are: + + file_manager=(close_idle_time=30,close_scan_interval=10) + + Update the sweep test to take advantage of new configurations and + shorten the runtime. + +commit 87dba78b65bc869c426363d86d02134aa8f8f2ac +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Tue Mar 17 14:39:49 2015 +1100 + + Round up the "in-memory size" for WT_UPDATEs to a multiple of 32. They are a very common case, and for tiny updates, this seems to to a better job of matching tcmalloc's behavior at least. + + refs SERVER-17424 + +commit d277a088385401ff6e6664dd688a2fcab2ac4087 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Mon Mar 16 11:42:56 2015 -0400 + + Sanity check all splits for at least 100 keys on the page, if there are huge keys and a too-small cache, there's nothing to be done. + +commit d4aa136a20c9baf151c3185058ce089679e0de0e +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Mon Mar 16 06:10:12 2015 +0000 + + Ensure we free memory when statistics cursor open fails. + + refs #1760 + +commit 66622ac33357d53212ddfa362fdf3c6b439bf34b +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Mon Mar 16 05:55:45 2015 +0000 + + Update the split deepen sanity check. Allow a page that is using more than 1/4 of the cache to be split. + + refs #1759 + +commit f99114f785985d152ba8ddfa735771574e6bff5d +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Mon Mar 16 16:43:27 2015 +1100 + + Fix the search for a free hazard pointer slot. Previously, it gave up early once all slots were allocated, even if they weren't all in use. + +commit 15fe04460c1f057811e0f3e02feb249c710d05ce +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Mon Mar 16 16:40:49 2015 +1100 + + Track the number of hazard pointers active in the eviction server. Don't start a new walk if we are close to the limit. + + refs SERVER-17551 + +commit e6d8c8a56c4dc83f206d43fcaa21902b35b1e4e6 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Mon Mar 16 05:20:22 2015 +0000 + + Add a new statistic tracking range of IDs pinned by a transaction. + + refs #1746 + +commit f954ab6b3fb6a51ecf3cf625cd3f5c5f818e0fde +Author: Keith Bostic <keith@wiredtiger.com> +Date: Sat Mar 14 12:31:00 2015 -0400 + + Pass a allocated memory size to __wt_page_inmem(). + +commit 308dc500adcde7b62c8dbb9aa0fb401795445546 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Sat Mar 14 10:51:21 2015 -0400 + + Don't drop core if duplicate symbols are detected (caller's table gets free'd twice). + +commit dc19643859063a03b985d97e2f24b1262ec4b15d +Author: Keith Bostic <keith@wiredtiger.com> +Date: Fri Mar 13 15:35:28 2015 -0400 + + Use SCNxxx #defines in sscanf() calls, not PRIxxx. + + We can't cast a pointer to a WT_LSN.offset to a pointer to a different type, we can't know WT_LSN.offset is the same size as that type. + +commit 99a992ed5207822ce2bc2fd69cb5b37408e6080f +Author: Keith Bostic <keith@wiredtiger.com> +Date: Fri Mar 13 14:17:34 2015 -0400 + + Switch sscanf from SCNu46 to SCNi64 so we handle both hexadecimal and integer symbol/frequency values + + Add testing support for out-of-range symbol/frequency values. + + refs #1536 + +commit 48a6f6e76aafbf7b33dbb22325ddb4e8e27603a1 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Fri Mar 13 10:04:06 2015 -0400 + + Fix problems with configuration value parsing, break out the Huffman configuration file parsing code into a single routine. + +commit e7ef6d0c3d107f374de9924d981d731fe36fb4ec +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Fri Mar 13 16:32:37 2015 +1100 + + Update memory allocation accounting for new pages and realloc. + + They used to not always account for alignment overhead. + +commit 1e97f9730a6fd70c03784e292cb0435c8fb82354 +Author: Don Anderson <dda@ddanderson.com> +Date: Thu Mar 12 14:02:24 2015 -0400 + + Fix to track deleted cells. Remove unused vars. + +commit 5778e4098862665d4cdaf2cda3ef40d47d951efb +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Thu Mar 12 16:35:27 2015 +1100 + + Add ability to configure sweep server timings via API. + + Still needs some test code. + +commit a1c9f8f8833482c9773fc30b7b9a7b5f1ab09014 +Author: Don Anderson <dda@ddanderson.com> +Date: Wed Mar 11 22:48:12 2015 -0400 + + Made some adjustments to stats so that the btree entries is an accurate total for row and column stores. Entries that appear in internal pages (pointing to leaf pages or other internal pages) are no longer counted. Added and modified tests to check the number of btree entries reported by stats against the number of KV pairs expected to be in btree. + + refs #1733 + +commit 6d7c061a8292021195b32260a8b41d3e92e1958f +Author: Susan LoVerso <sue@wiredtiger.com> +Date: Wed Mar 11 11:11:57 2015 -0400 + + Support no-logging mode in wt command for debug. + + refs #1732 + +commit 4a802bd592b6c8506b07900bf89a9d9fe53cfc25 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Wed Mar 11 09:30:39 2015 -0400 + + If the application only has 1 outstandin async operation at a time, we won't let them specify that. Change the minimum ops_max value to 1, but set the minimum maximum bound in the code to 10 as before. + +commit 3a412a2371e80ed9e26e546b6fe3c09d2d4e4091 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Wed Mar 11 09:29:55 2015 -0400 + + If the application re-uses a WT_SYNC_OP structure, we assert: return EINVAL instead, it's easier to debug than a core dump. + +commit 16119f98fee9525bf1990ccbcc979b9f08a8b000 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Wed Mar 11 20:24:20 2015 +1100 + + Don't cache an ikey before winning the race to split into a parent. + + refs #1582 + +commit 4d0b4093e3d3c9fa0be2bbb01467579f05deddc5 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Wed Mar 11 03:44:22 2015 +0000 + + Update visibility check to know about checkpoints. + + This should allow us to evict more pages while a checkpoint is in progress. We can evict dirty pages from a file once the checkpoint is finished in that file. Similarly for the row store specific obsolete update check. + + refs #1745 + +commit 272daddb3734cd196cd303df4271f7e9e8f00cd3 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Wed Mar 11 02:55:56 2015 +0000 + + Add a global checkpoint generation and track it per data handle. Not yet used, but will be used to allow more eviction during checkpoints. + + refs #1745 + +commit 89db28287d7b212efb31203eca253fbf144ea207 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Tue Mar 10 07:19:59 2015 +0000 + + Mark statistics with the right aggregation flags (specifically cache dirty tracking and log preallocation). This helps wtstats graph generator. + + refs #1742 + +commit e51cc35a88ed497f911a25fec9f4bfcc6617d6a2 +Author: Mark Benvenuto <mark.benvenuto@mongodb.com> +Date: Mon Mar 9 15:40:59 2015 -0400 + + SERVER-17471: Use uint64_t instead of long since long is compiler specific + +commit 105903c1beb0d16b40e243ed3624417c4fa0702c +Author: Keith Bostic <keith@wiredtiger.com> +Date: Mon Mar 9 12:19:50 2015 -0400 + + Remove pockets of case-sensitivity for configuration strings. + +commit 1e8c2b89de09462253231213329674b8769d11fe +Author: Keith Bostic <keith@wiredtiger.com> +Date: Sun Mar 8 13:05:51 2015 -0400 + + Change the statistics cursor WT_CURSOR.reset method to re-load statistics values. + + refs #1533 + +commit 769dc5976fe0d2448fb4fde511c5bc29eea39bfb +Author: Keith Bostic <keith@wiredtiger.com> +Date: Tue Mar 3 16:45:09 2015 -0500 + + Change checkpoints to do first-fit allocation when we start writing the actual checkpoint blocks, that way if we delete significant space, the checkpoint blocks won't prevent file truncation. + + +3.0.1, Mar 9 2015 +----------------- + +commit d654795bb763b95d14604b9b65d09ae79b8ee5b6 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Mon Mar 9 05:35:33 2015 +0000 + + Add a test case for checkpoint consistency. + + We recently fixed a bug where updating a page after the write leaves phase of a checkpoint had completed, then either checkpointed again or closed the database. We would end up with inconsistent data. That was the case because we weren't marking trees dirty all the time. + + This test case reproduces the scenario so we don't reintroduce the failure. + + refs #1735 SUPPORT-1248, SERVER-17506 + +commit 0315ee75f712ed0ccddca0616339de93b17835a3 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Mon Mar 9 11:21:06 2015 +1100 + + Bump release version to 2.5.2 + +commit 4235c69d37474fb4e14673e0ea99337659db948d +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Mon Mar 9 08:32:50 2015 +1100 + + When skipping a dirty page during a checkpoint, make sure the tree is marked dirty. + + refs SUPPORT-1248, SERVER-17506 + +commit 8382d14f32efc53b19aecd596cab3ba0d682b22d +Author: Keith Bostic <keith@wiredtiger.com> +Date: Fri Mar 6 16:10:37 2015 -0500 + + Only increment/decrement from the current position on the first position set by our caller, that is, the position passed-in from our caller is the only position the caller has already seen. + + refs SERVER-17345. + +commit ed9c48c7b8fa5dd3362e417fda8337f1690585ed +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Fri Mar 6 22:48:45 2015 +1100 + + Ignore empty child pages in column stores. + +commit 553a351ca6a81179c7f1db9c04d6f96aef0545ac +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Thu Mar 5 05:46:15 2015 +0000 + + Don't allow LSM bloom create to block waiting for space in the cache. + + We are only accessing one page at a time, and allowing the bloom create to block can lead to LSM not making progress. + + Refs: #1720 + +commit 5285b768de3237b4379f1526176efdeb55860971 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Thu Mar 5 16:15:08 2015 +1100 + + Remove the special "discard" path for trees marked clean: always go through eviction, which checks page modify flags. This may mean additional writes to internal pages when discarding a tree, but means that if a discard stops part-way through, the remaining in-memory tree has not lost any context. + + refs SUPPORT-1248, SERVER-17510 + +commit 00edf7a47115923bc9f3eaa7eee84c9f7d6b0d77 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Thu Mar 5 16:15:07 2015 +1100 + + If a page is split by eviction, mark the tree dirty to avoid having dirty pages in a tree marked clean. + + refs SUPPORT-1248 + +commit 57d3eba53fb91a0287374b9642b7cd4ef644854a +Author: Susan LoVerso <sue@wiredtiger.com> +Date: Wed Mar 4 10:00:59 2015 -0500 + + Add fsync before closing log files and after header. #1717 + + refs SERVER-17451 + +commit d970bfe6b1bed7d1919b800bf2d65a3789b74d6f +Author: Keith Bostic <keith@wiredtiger.com> +Date: Tue Mar 3 09:38:59 2015 -0500 + + Don't set eviction_workers_min/eviction_workers_max in the connection structure before checking the values are OK. + +commit 90f3f34a97440b6788a1a558e560a33fd116f166 +Author: Susan LoVerso <sue@wiredtiger.com> +Date: Tue Mar 3 09:26:38 2015 -0500 + + Move writing into log worker thread. #1683 + +commit 1266bbb1143b22fec6b1c255b3aade5d0506477e +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Tue Mar 3 03:37:18 2015 +0000 + + Fix a bug in the reconfigure API related to shared cache quotas. + + While fixing the bug simplify the code flow for reconfigure and caches/shared caches. + + refs #1712 + +commit 1ebd617e6dfcf542983d62d9666c5a328dd41bb8 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Mon Mar 2 15:31:05 2015 -0500 + + Add overflow key/value counts to the statistics code + + refs #1520, #1703 + +commit a2166ead528ef61da478db67e3c0209a6bef5ac6 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Mon Mar 2 14:18:37 2015 +1100 + + Allow memory_page_max to be at most 1/4 of the cache size not 1/2. + + If we let a single page grow to half the cache size, it's too easy for the cache to get pinned full if it is tiny to start with. + +commit b036921625e415bb66ac458922b81a9fae07740b +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Mon Mar 2 14:17:12 2015 +1100 + + Take all eviction candidates if we are aggressive. + +commit 0ddd3face0b99f5653001825bf3df0662ffcdc10 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Mon Mar 2 13:07:08 2015 +1100 + + Reverse the direction of the LRU walk each time the eviction server goes to sleep. Keep looking for candidates if eviction is stuck. Don't give up our walk position if eviction is stuck. + +commit cff10ba30f2eac379197e5d7dea49da2b8159890 +Author: Don Anderson <dda@ddanderson.com> +Date: Fri Feb 27 15:05:24 2015 -0500 + + Added test for wt dump on an index. + + refs #1709 + +commit a9f0e3ac769a060a8c3b06bac71fbed0e5f46cc6 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Fri Feb 27 16:59:35 2015 +1100 + + If we encounter a dirty page when closing a clean tree, switch to checkpointing. + + refs SERVER-17319, #1643?, #1404? + +commit bfcf5987b2b6f08d931d620330aed46837e3a8c2 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Thu Feb 26 16:32:09 2015 +1100 + + Add some paranoia to LSM around transaction checks: make sure we have allocated an ID before using it for anything. + + Recheck switch transactions before doing update checks in old chunks. + + refs #1641, #1701, #1702 + +commit da0bc67c821282e9fd0da725279811b59b25a675 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Thu Feb 26 16:32:09 2015 +1100 + + Add some paranoia around setting row-store internal keys. + + refs #1582 + +commit 5ea91f6ed0e0677530c5ab8215f81eb48ea307b8 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Thu Feb 26 04:51:25 2015 +0000 + + Several optimizations to large page eviction: + + * Don't update the read generation on page in if it's set to oldest. + * Clear the walk positions before the eviction server sleeps. + * If only looking for pages that would block add them all to the queue. + * If evicting dirty pages use the worker threads, not the server. + +commit 0eecd0a2d97771380ecbd7fd27bd44988db1148c +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Thu Feb 26 02:08:36 2015 +0000 + + Fix a bug in checkpoint, where it could get an EBUSY return. + + The case that could return EBUSY was when checkpointing with a specific target, while that target was open exclusively or for a bulk load. + + Refs #1404 #1589 + +commit 392a540deec817c5d6738b8e848a68882df3ac8a +Author: Susan LoVerso <sue@wiredtiger.com> +Date: Wed Feb 25 15:03:50 2015 -0500 + + If the LSN given doesn't exist, don't return an error, but do force recovery. Fix recover.sh to grep the CONFIG. #1700 + +commit 59b699b7085868d1b12a41ae4cd7a01f25f6e865 +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Tue Feb 24 23:55:20 2015 +0000 + + Handle the case where a large record pushes us over a page boundary and we spill across two pages correctly. The previous fix could let us span additional space. + + Fix another bug in the fixup code - ensure there is enough space in the temporary buffer when fixing up after the fist page is full. + + refs #1697 + +commit ad8b58188961943e74c57b85e3b976aa03b79617 +Author: Susan LoVerso <sue@wiredtiger.com> +Date: Tue Feb 24 18:54:18 2015 -0500 + + Set flags to SLOT_INIT_FLAGS on free. refs #1683 + +commit 45e4c049044efc96c682f390466a35d22dac555f +Author: Keith Bostic <keith@wiredtiger.com> +Date: Tue Feb 24 16:17:32 2015 -0500 + + Move WT_PAGE.u.row.d next to WT_PAGE.u.row.entries, it reduces cache misses inside of row-search. + + refs #1665 + +commit 3c6d7adf422f432ae117e2292dcce00cc3b531a3 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Tue Feb 24 16:07:16 2015 -0500 + + Turn off key prefix-compression and rebuild the key before doing the boundary split. The size of the key is likely to increase by a few bytes, and if the value is large enough to consume almost all of the buffer, we overflow the space available. + + refs #1697 + +commit ca9ab16c320f6f154ff1fd3d0b65316f87e8bddc +Author: Alex Gorrod <alexg@wiredtiger.com> +Date: Tue Feb 24 06:51:14 2015 +0000 + + Fix a bug in reconciliation. In cases where we decide to squeeze one more large entry onto a page, we could have attempted to decrement space available negative. + +commit 777c35b074fce3656c14ca5770b424a65c719134 +Author: Susan LoVerso <sue@wiredtiger.com> +Date: Thu Feb 19 12:32:16 2015 -0500 + + Add log worker thread to advance write_lsn. + + refs #1683 + +commit ae686a225a011bac07119e2d66e837e08d5a3a0e +Author: Keith Bostic <keith@wiredtiger.com> +Date: Tue Feb 10 12:46:04 2015 -0500 + + Sasha says (my paraphrasing): I simply rearranged the fields in the WT_PAGE struct, so that u.row.d and u.row.entries are close together at the very end of the definition of union u, and the "uint8_t type;" immediately follows that. + + +3.0-RC10, Feb 24 2015 +--------------------- + +commit 2fdfb2bbed56e42e1717e567828c68d0b2eb868d +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Tue Feb 24 14:44:48 2015 +1100 + + Review places that set/clear session->dhandle, replace with macros. Change callers to save/restore if they need to release a handle after a call. + +commit ce89b608835561b11ce4e525a5ebdad86558f115 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Mon Feb 23 19:55:56 2015 -0500 + + The statistics server has open handles and may be opening underlying handles, make sure we don't overwrite them, reference #1694. + +commit 8827d909d7f14e9ad767d909d65598508ff0025c +Author: Keith Bostic <keith@wiredtiger.com> +Date: Mon Feb 23 09:39:21 2015 -0500 + + Add an explicit barrier after setting the checkpointing value. (I'm pretty sure it's not actually necessary to have a barrier in the current code, but the barrier we rely on is in a different function and isn't always called, depending the eviction configuration of this file, I'd just rather be safe than try and debug this some time in the future.) + +commit 10abb2c47cffbcc215c3507256cee3e2cae2dd5f +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Mon Feb 23 15:23:54 2015 +1100 + + If eviction is walking a file with only one live page (e.g., the last page), be careful to step over it, or eviction's hazard reference can make forced eviction stall. + + SERVER-17344 + +commit f72367f97357cf3e77f0d57d39992686b400ebe7 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Sun Feb 22 15:40:37 2015 -0500 + + Make sure we have a local copy of the start/stop cursor keys before calling into the underlying Btree range truncate function + + SERVER-17141. + +commit 09d345951d43e7ed928980048e5c5c927406a66b +Author: Keith Bostic <keith@wiredtiger.com> +Date: Fri Feb 20 11:25:22 2015 -0500 + + If WT_SESSION.truncate isn't given a start cursor, instantiate one, it's always faster to traverse Btree objects in a forward direction. + +commit ce38bc5b40f1aa3ffc07cdd2cf993e32143135b5 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Fri Feb 20 08:57:57 2015 -0500 + + We don't need to do a search in the row-store cursor truncate setup code, the WT_SESSION.truncate API code already did one. + + This does not mean WT_SESSION.truncate can't return WT_NOTFOUND in some cases, the first thing cursor truncate does is a cursor remove call in order to acquire the page's write generation information -- that remove call does the usual "discard my current reference and get a new one" work, which allows another deleting thread to race and remove one of the truncate's start/stop keys. + + refs SERVER-17141 + + +3.0-RC9, Feb 18 2015 +-------------------- + +commit bf3ee2cd064b46cf0175d75950c825aa1f42c694 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Wed Feb 18 12:53:41 2015 +1100 + + Flip cache overhead to apply to the allocated bytes rather than the total size. Include the overhead in stats so that tools (e.g., mongostat) report accurate cache full and dirty percentages. This also makes eviction triggers and targets meaningful: with the default trigger of 95% and overhead 8%, eviction was previously never triggered until the cache was completely full. + +commit f9e6f942cf73c8a53aaadbc587c1b7efad6cc832 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Tue Feb 17 11:33:13 2015 -0500 + + Coverity notes the TXN_API_END_RETRY macro has an unnecessary test for "(ret == 0)" at the end of the do {} while loop. + +commit c34a56f357e21d134a2d9d0fefc032544069d8d7 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Tue Feb 17 21:56:34 2015 +1100 + + Allow the maximum number of eviction threads to be reconfigured. This was previously permitted by the API, but the array of thread contexts was not correctly resized, leading to segfaults. + + refs SERVER-17293 + +commit 67527fc235406469e69dbaec3dcd571469e660c0 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Tue Feb 17 21:50:53 2015 +1100 + + Make the eviction walk incremental: don't spend too long in any one file, fix tracking of whether we are making progress. + +commit 788265ed273c63183053e6325a9aa03c89c02860 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Tue Feb 17 21:48:44 2015 +1100 + + Combine the various checks for whether a page can be evicted into one place. + +commit 748e7b0c58b358b14340bacae41f9c46f3c06f7e +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Tue Feb 17 21:47:57 2015 +1100 + + Skip hot pages during write leaves: checkpoint will have to visit them anyway. + +commit a9de0f7ac8ad373d7aef6a480c69a2a7e0b55c59 +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Tue Feb 17 21:18:41 2015 +1100 + + Run recovery after crashing test/format in the recovery test. + +commit 4733961a3c1fa37988178d1b1dd4eb44d83b63f6 +Author: Thomas Rueckstiess <thomas@rueckstiess.net> +Date: Mon Feb 16 12:50:26 2015 +1100 + + fixes and improvements for wtperf parsing + + - convert wtperf microsec to millisec + - don't skip monitor* files when parsing directory + - parsing code reorganization + - renamed wtperf stats fixture + - added tests + +commit dc396e1cd64871219b9e5a1b6558707feb70706e +Author: Keith Bostic <keith@wiredtiger.com> +Date: Sun Feb 15 13:44:30 2015 -0500 + + Clear the btree object statistics we're about to count, otherwise each cursor gets a cumulative value. + +commit 195b144bb37814b31cfa413029cda0b28f13f261 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Fri Feb 13 12:06:36 2015 -0500 + + Don't map WT_NOTFOUND to ENOENT unless a uri was specified, that's the only interesting case. Reference SERVER-17141. + +commit e9d7fee2c2c08985b8e2d2716e899853c5198290 +Author: Thomas Rueckstiess <thomas@rueckstiess.net> +Date: Fri Feb 13 17:22:05 2015 +1100 + + added support to parse wtperf files. + + they go into a separate stats section named "wtperf". + +commit 29d0d26fd1cd76392ea8225c1c4022ca54443737 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Thu Feb 12 18:05:33 2015 -0500 + + Ignore unexpected information in the metadata entry, the metadata entry might have been created by a future release, with unknown options. + +commit 05f07753059a4fa7f0f1bab7a107a9e6d17bf4af +Author: Keith Bostic <keith@wiredtiger.com> +Date: Thu Feb 12 15:42:35 2015 -0500 + + Remove the requirement of a HAVE_DIAGNOSTIC build for the verify commands to work (except for dump_offsets, that requires the btree debugging code and so won't work anywhere but a HAVE_DIAGNOSTIC build). + +commit 006ed9f17c7fc0fe65dc43717ed0239b3bac564c +Author: Keith Bostic <keith@wiredtiger.com> +Date: Thu Feb 12 15:31:36 2015 -0500 + + Add support for a new verify debug option, "dump_shape", that reports the levels of the tree. + + __wt_config_gets() returns WT_NOTFOUND when there's no entry, don't fail every command when DIAGNOSTIC #defined and debug options aren't set. + +commit 46b7721215856d08ca3a37f7ffc27c57b1d4c1d7 +Author: Susan LoVerso <sue@wiredtiger.com> +Date: Thu Feb 12 13:27:32 2015 -0500 + + Add recover config setting and use it in the wt command. #1651 + +commit 0305a51ffba383af13d6078d409a03b249c502c5 +Author: Don Anderson <dda@ddanderson.com> +Date: Wed Feb 11 10:21:04 2015 -0500 + + Add test to detect file ID problems in recovery. Refs #1622. + +commit fc0ff5a9ea09e54512353d2275126cb54dbc5451 +Author: Susan LoVerso <sue@wiredtiger.com> +Date: Tue Feb 10 13:02:28 2015 -0500 + + Allow 'wt' command to run with or without recovery. #1651 + +commit a26d87a53eb2ac2dcae9312b7979499c34c11613 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Mon Feb 2 19:20:33 2015 -0500 + + Replace wiredtiger_strerror_r with WT_SESSION.strerror, reference #1516. + +commit 33c146b51fdac86999e2eaa67f5636490eb441fb +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Thu Feb 12 13:44:35 2015 +1100 + + Disable aggregation across all open checkpoints if statistics cursors don't specify a checkpoint. + +commit 04ec3d021d2f8b08b69d3ea5d0f243f468c71f2e +Author: Michael Cahill <michael.cahill@wiredtiger.com> +Date: Thu Feb 12 13:00:49 2015 +1100 + + Move server thread waits to the beginning of their loops: check that we're still running before waiting. This makes more sense to me, but also fixes a problem introduced recently where the checkpoint server could hang on shutdown if the signal from the closing thread got lost. + +commit 85aae87cb1e019b0bcac4854e6508f11104f5339 +Author: Keith Bostic <keith@wiredtiger.com> +Date: Wed Feb 11 12:34:03 2015 -0500 + + Mimic Alex's fix in 152a0ef, to fsync created files to disk, for truncated + files, never surprise the upper-layer. + +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. + |