Background:
LevelDB uses a cache (util/cache.h, util/cache.cc) of (key,value)
pairs for two purposes:
- a cache of (table, file handle) pairs
- a cache of blocks
The cache places the (key,value) pairs in a reference-counted
wrapper. When it returns a value, it returns a reference to this
wrapper. When the client has finished using the reference and
its enclosed (key,value), it calls Release() to decrement the
reference count.
Each (key,value) pair has an associated resource usage. The
cache maintains the sum of the usages of the elements it holds,
and removes values as needed to keep the sum below a capacity
threshold. It maintains an LRU list so that it will remove the
least-recently used elements first.
The max_open_files option to LevelDB sets the size of the cache
of (table, file handle) pairs. The option is not used in any
other way.
The observed behaviour:
If LevelDB at any time used more file handles concurrently than
the cache size set via max_open_files, it attempted to reduce the
number by evicting entries from the table cache. This could
happen most easily during compaction, and if max_open_files was
low. Because the handles were in use, their reference count did
not drop to zero, and so the usage sum in the cache was not
modified by the evictions. Subsequent Insert() calls returned
valid handles, but their entries were immediately evicted from
the cache, which though empty still acted as though full. As a
result, there was effectively no caching, and the number of open
file handles rose []ly until it hit system-imposed limits and
the process died.
If one set max_open_files lower, the cache was more likely to
exhibit this beahviour, and cause the process to run out of file
descriptors. That is, max_open_files acted in almost exactly the
opposite manner from what was intended.
The problems:
1. The cache kept all elements on its LRU list eligible for capacity
eviction---even those with outstanding references from clients. This was
ineffective in reducing resource consumption because there was an
outstanding reference, guaranteeing that the items remained. A secondary
issue was that there is no guarantee that these in-use items will be the
last things reached in the LRU chain, which actually recorded
"least-recently requested" rather than "least-recently used".
2. The sum of usages was decremented not when a (key,value) was evicted from
the cache, but when its reference count went to zero. Thus, when things
were removed from the cache, either by garbage collection or via Erase(),
the usage sum was not necessarily decreased. This allowed the cache to act
as though full when it was in fact not, reducing caching effectiveness, and
leading to more resources being consumed---the opposite of what the
evictions were intended to achieve.
3. (minor) The cache's clients insert items into it by first looking up the
key, and inserting only if no value is found. Although the cache has an
internal lock, the clients use no locking to ensure atomicity of the
Lookup/Insert pair. (see table/table.cc: block_cache->Insert() and
db/table_cache.cc: cache_->Insert()). Thus, if two threads Insert() at
about the same time, they can both Lookup(), find nothing, and both
Insert(). The second Insert() would evict the first value, leaving each
thread with a handle on its own version of the data, and with the second
version in the cache. It would be better if both threads ended up with a
handle on the same (key,value) pair, which implies it must be the first item
inserted. This suggests that Insert() should not replace an existing value.
This can be made safe with current usage inside LeveDB itself, but this is
not easy to change first because Cache is a public interface, so to change
the semantics of an existing call might break things, second because Cache
is an abstract virtual class, so adding a new abstract virtual method may
break other implementations, and third, the new method "insert without
replacing" cannot be implemented in terms of the existing methods, so cannot
be implemented with a non-abstract default. But fortunately, the effects
of this issue are minor, so this issue is not fixed by this change.
The changes:
The assumption in the fixes is that it is always better to cache
entries unless removal from the cache would lead to deallocation.
Cache entries now have an "in_cache" boolean indicating whether
the cache has a reference on the entry. The only ways that this can
become false without the entry being passed to its "deleter" are via
Erase(), via Insert() when an element with a duplicate key is inserted,
or on destruction of the cache.
The cache now keeps two linked lists instead of one. All items
in the cache are in one list or the other, and never both. Items
still referenced by clients but erased from the cache are in
neither list. The lists are:
- in-use: contains the items currently referenced by clients, in no particular
order. (This list is used for invariant checking. If we removed the check,
elements that would otherwise be on this list could be left as disconnected
singleton lists.)
- LRU: contains the items not currently referenced by clients, in LRU order
A new internal Ref() method increments the reference count. If
incrementing from 1 to 2 for an item in the cache, it is moved
from the LRU list to the in-use list.
The Unref() call now moves things from the in-use list to the LRU
list if the reference count falls to 1, and the item is in the
cache. It no longer adjusts the usage sum. The usage sum now
reflects only what is in the cache, rather than including
still-referenced items that have been evicted.
The LRU_Append() now takes a "list" parameter so that it can be
used to append either to the LRU list or the in-use list.
Lookup() is modified to use the new Ref() call, rather than
adjusting the reference count and LRU chain directly.
Insert() eviction code is also modified to adjust the usage sum and the
in_cache boolean of the evicted elements. Some LevelDB tests assume that there
will be no caching whatsoever if the cache size is set to zero, so this is
handled as a special case.
A new private method FinishErase() is factored out
with the common code from where items are removed from the cache.
Erase() is modified to adjust the usage sum and the in_cache
boolean of the erased elements, and to use FinishErase().
Prune() is modified to use FinishErase() also, and to make use of the fact that
the lru_ list now contains only items with reference count 1.
- EvictionPolicy is modified to test that an entry with an
outstanding handle is not evicted. This test fails with the old cache.cc.
- A new test case UseExceedsCacheSize verifies that even when the
cache is overfull of entries with outstanding handles, none are
evicted. This test fails with the old cache.cc, and is the key
issue that causes file descriptors to run out when the cache
size is set too small.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=123247237
Detailed description:
Suppose an input split is generated between two leveldb record blocks and the preceding block ends with null padding.
A reader that previously read at least 1 record within the first block (before encountering the padding) upon trying to read the next record, will successfully and correctly read the next logical record from the subsequent block, but will return a last record offset pointing to the padding in the first block.
When this happened in a [], it resulted in duplicate records being handled at what appeared to be different offsets that were separated by only a few bytes.
This behavior is only observed when at least 1 record was read from the first block before encountering the padding. If the initial offset for a reader was within the padding, the correct record offset would be reported, namely the offset within the second block.
The tests failed to catch this scenario/bug, because each read test only read a single record with an initial offset. This CL adds an explicit test case for this scenario, and modifies the test structure to read all remaining records in the test case after an initial offset is specified. Thus an initial offset that jumps to record #3, with 5 total records in the test file, will result in reading 2 records, and validating the offset of each of them in order to pass successfully.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=115338487
1. Object files, libraries, and compiled executables are put
into subdirectories.
2. The shared library is linked from individual object files.
This provides for greater parallelism on large desktops
while at the same time making for easier builds on small
(i.e. embedded) systems. Fixes issue #279.
3. One program, db_bench, is compiled using the shared library.
4. The source file for "leveldbutil" was renamed from
leveldb_main.cc to leveldbutil.cc. This provides for simpler
makefile rules.
5. Because all targets placed the library (libleveldb.a) at the top
level, the last platform built (desktop/device) always overwrote
any prior artifact.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=113407013
In preparation for accepting GitHub pull requests this new README
section outlines the general criteria that the leveldb project owners
will use when accepting external (and internal) project contributions.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=111349899
The write operations in the table happens without holding the mutex
lock, but concurrent writes are avoided using "writers_" queue.
The Arena::MemoryUsage could access the blocks when write happens.
So, the memory usage is cached in atomic word and can be loaded
from any thread safely.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=107573379
A recent CL (104348226) created the port_posix library, but omitted: port/atomic_pointer.h.
And when:
[] test third_party/leveldb:all
was run this error was reported:
//third_party/leveldb:port_posix does not depend on a
module exporting 'third_party/leveldb/port/atomic_pointer.h'
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=105243399
This allows build agents to select from multiple Xcode installations.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=104859097
The approximate RAM usage of the database is calculated from the memory
allocated for write buffers and the block cache. This is to give an
estimate of memory usage to leveldb clients.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=104222307
Prune() drops on-memory read cache of the database, so that the client can
relief its memory shortage.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=101335710
The create function took |num_keys| as an int, but callers and implementers wanted it to function as a size_t (e.g. passing std::vector::size() in, passing it to vector constructors as a size arg, indexing containers by it, etc.). This resulted in implicit conversions between the two types as well as warnings (found with Chromium's external copy of these sources, built with MSVC) about signed vs. unsigned comparisons.
The leveldb sources were already widely using size_t elsewhere, e.g. for key and filter lengths, so using size_t here is not inconsistent with the existing code. However, it does change the public C API.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=101074871
All other Status::Code enum values have an Is**() method with the one
exception of InvalidArgument.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=97166441
Prior implementation would always try to reuse the manifest, even if reuse_logs
was false (the default). This was missed because the stock
Env::NewAppendableFile implementation returns false forcing the creation of a
new log.
(Based on a suggestion by cmumford.)
"open" benchmark on my workstation speeds up significantly since we
can now avoid three fdatasync calls and a compaction per open:
Before: ~80000 microseconds
After: ~130 microseconds
Details:
(1) Added Options::reuse_logs (currently defaults to false) to control
new behavior. The intention is to change the default to true after some
baking.
(2) Added Env::NewAppendableFile() whose default implementation returns
a not-supported error.
(3) VersionSet::Recovery attempts to reuse the MANIFEST from which
it is recovering.
(4) DBImpl recovery attempts to reuse the last log file and memtable.
(5) db_test.cc now tests a new configuration that sets reuse_logs to true.
(6) fault_injection_test also tests a reuse_logs==true config.
(7) Added a new recovery_test.
This test is intended to ensure leveldb properly detects and recovers from
faults - specifically unwritten file data lost as a result of a system reset.
Changes are:
* Update version number to 1.18
* Replace the basic fprintf call with a call to fwrite in order to
work around the apparent compiler optimization/rewrite failure that we are
seeing with the new toolchain/iOS SDKs provided with Xcode6 and iOS8.
* Fix ALL the header guards.
* Createed a README.md with the LevelDB project description.
* A new CONTRIBUTING file.
* Don't implicitly convert uint64_t to size_t or int. Either preserve it as
uint64_t, or explicitly cast. This fixes MSVC warnings about possible value
truncation when compiling this code in Chromium.
* Added a DumpFile() library function that encapsulates the guts of the
"leveldbutil dump" command. This will allow clients to dump
data to their log files instead of stdout. It will also allow clients to
supply their own environment.
* leveldb: Remove unused function 'ConsumeChar'.
* leveldbutil: Remove unused member variables from WriteBatchItemPrinter.
* OpenBSD, NetBSD and DragonflyBSD have _LITTLE_ENDIAN, so define
PLATFORM_IS_LITTLE_ENDIAN like on FreeBSD. This fixes:
* issue #143
* issue #198
* issue #249
* Switch from <cstdatomic> to <atomic>. The former never made it into the
standard and doesn't exist in modern gcc versions at all. The later contains
everything that leveldb was using from the former.
This problem was noticed when porting to Portable Native Client where no memory
barrier is defined. The fact that <cstdatomic> is missing normally goes
unnoticed since memory barriers are defined for most architectures.
* Make Hash() treat its input as unsigned. Before this change LevelDB files
from platforms with different signedness of char were not compatible. This
change fixes: issue #243
* Verify checksums of index/meta/filter blocks when paranoid_checks set.
* Invoke all tools for iOS with xcrun. (This was causing problems with the new
XCode 5.1.1 image on pulse.)
* include <sys/stat.h> only once, and fix the following linter warning:
"Found C system header after C++ system header"
* When encountering a corrupted table file, return Status::Corruption instead of
Status::InvalidArgument.
* Support cygwin as build platform, patch is from https://code.google.com/p/leveldb/issues/detail?id=188
* Fix typo, merge patch from https://code.google.com/p/leveldb/issues/detail?id=159
* Fix typos and comments, and address the following two issues:
* issue #166
* issue #241
* Add missing db synchronize after "fillseq" in the benchmark.
* Removed unused variable in SeekRandom: value (issue #201)
- Cleanup: delete unused IntSetToString
It was added in http://cr/19491949 (and was referenced at the time).
The last reference was removed in http://cr/19507363.
This fixes warning/error with pre-release crosstoolv18:
'std::string leveldb::{anonymous}::IntSetToString(const std::set<long unsigned int>&)' defined but not used [-Werror=unused-function]
- Added arm64 and and armv7s to IOS build as suggested on leveldb mailing list.
- Changed local variable type from int to size_t
This eliminates compiler warning/error and resolves
https://code.google.com/p/leveldb/issues/detail?id=140
- Make Log::Reader not report a corruption when the last record in a
log file is truncated.
- Fix issue 224: variable created but not utilized.
- Remove comment that referenced a removed feature.
- switched from mmap based writing to simpler stdio based writing. Has a
minor impact (0.5 microseconds) on microbenchmarks for asynchronous
writes. Synchronous writes speed up from 30ms to 10ms on linux/ext4.
Should be much more reliable on diverse platforms.
- compaction errors now immediately put the database into a read-only
mode (until it is re-opened). As a downside, a disk going out of
space and then space being created will require a re-open to recover
from, whereas previously that would happen automatically. On the
plus side, many corruption possibilities go away.
- force the DB to enter an error-state so that all future writes fail
when a synchronous log write succeeds but the sync fails.
- repair now regenerates sstables that exhibit problems
- fix issue 218 - Use native memory barriers on OSX
- fix issue 212 - QNX build is broken
- fix build on iOS with xcode 5
- make tests compile and pass on windows
Fix issues 200, 201
Also,
* Fix link to bigtable paper in docs.
* New sstables will have the file extension .ldb. .sst files will
continue to be recognized.
* When building for iOS, use xcrun to execute the compiler. This may
affect issue 177.
Fix issues 77, 87, 182, 190.
Additionally, fix the bug described in
https://groups.google.com/d/msg/leveldb/yL6h1mAOc20/vLU64RylIdMJ
where a large contiguous keyspace of deleted data was not getting
compacted.
Also fix a bug where options.max_open_files was not getting clamped
properly.
sanjay
m3b
John Abd-El-Malek
Nicholas Westlake
mjwiacek
cmumford
Chris Mumford
Bruce Dawson
Lars-Magnus Skog
Chris Mumford
ssid
ndmatthews
tzik
pkasting
Venilton FalvoJr
ideawu
Paul Irish
David Turner
David Grogan