Summary:
This is first in a series of diffs that fixes data races detected by thread sanitizer.
Here the problem is that we call Ref() on a column family during a single-threaded write, without holding a mutex.
Test Plan: TSAN is no longer complaining about LevelLimitReopen.
Reviewers: yhchiang, rven, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D32121
Summary:
This patch changes concurrency guarantees around ColumnFamilySet::column_families_ and ColumnFamilySet::column_families_data_.
Before:
* When mutating: lock DB mutex and spin lock
* When reading: lock DB mutex OR spin lock
After:
* When mutating: lock DB mutex and be in write thread
* When reading: lock DB mutex or be in write thread
That way, we eliminate the spin lock that protects these hash maps and simplify concurrency. That means we don't need to lock the spin lock during writing, since writing is mutually exclusive with column family create/drop (the only operations that mutate those hash maps).
With these new restrictions, I also needed to move column family create to the write thread (column family drop was already in the write thread).
Even though we don't need to lock the spin lock during write, impact on performance should be minimal -- the spin lock is almost never busy, so locking it is almost free.
This addresses task t5116919.
Test Plan:
make check
Stress test with lots and lots of column family drop and create:
time ./db_stress --threads=30 --ops_per_thread=5000000 --max_key=5000 --column_families=200 --clear_column_family_one_in=100000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress/
Reviewers: yhchiang, rven, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D30651
Summary:
When scaling to higher number of column families, the worst bottleneck was MaybeScheduleFlushOrCompaction(), which did a for loop over all column families while holding a mutex. This patch addresses the issue.
The approach is similar to our earlier efforts: instead of a pull-model, where we do something for every column family, we can do a push-based model -- when we detect that column family is ready to be flushed/compacted, we add it to the flush_queue_/compaction_queue_. That way we don't need to loop over every column family in MaybeScheduleFlushOrCompaction.
Here are the performance results:
Command:
./db_bench --write_buffer_size=268435456 --db_write_buffer_size=268435456 --db=/fast-rocksdb-tmp/rocks_lots_of_cf --use_existing_db=0 --open_files=55000 --statistics=1 --histogram=1 --disable_data_sync=1 --max_write_buffer_number=2 --sync=0 --benchmarks=fillrandom --threads=16 --num_column_families=5000 --disable_wal=1 --max_background_flushes=16 --max_background_compactions=16 --level0_file_num_compaction_trigger=2 --level0_slowdown_writes_trigger=2 --level0_stop_writes_trigger=3 --hard_rate_limit=1 --num=33333333 --writes=33333333
Before the patch:
fillrandom : 26.950 micros/op 37105 ops/sec; 4.1 MB/s
After the patch:
fillrandom : 17.404 micros/op 57456 ops/sec; 6.4 MB/s
Next bottleneck is VersionSet::AddLiveFiles, which is painfully slow when we have a lot of files. This is coming in the next patch, but when I removed that code, here's what I got:
fillrandom : 7.590 micros/op 131758 ops/sec; 14.6 MB/s
Test Plan:
make check
two stress tests:
Big number of compactions and flushes:
./db_stress --threads=30 --ops_per_thread=20000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=15 --max_background_compactions=10 --max_background_flushes=10 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000
max_background_flushes=0, to verify that this case also works correctly
./db_stress --threads=30 --ops_per_thread=2000000 --max_key=10000 --column_families=20 --clear_column_family_one_in=10000000 --verify_before_write=0 --reopen=3 --max_background_compactions=3 --max_background_flushes=0 --db=/fast-rocksdb-tmp/db_stress --prefixpercent=0 --iterpercent=0 --writepercent=75 --db_write_buffer_size=2000000
Reviewers: ljin, rven, yhchiang, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D30123
Summary:
Introduces a new class for managing write buffer memory across column
families. We supplement ColumnFamilyOptions::write_buffer_size with
ColumnFamilyOptions::write_buffer, a shared pointer to a WriteBuffer
instance that enforces memory limits before flushing out to disk.
Test Plan: Added SharedWriteBuffer unit test to db_test.cc
Reviewers: sdong, rven, ljin, igor
Reviewed By: igor
Subscribers: tnovak, yhchiang, dhruba, xjin, MarkCallaghan, yoshinorim
Differential Revision: https://reviews.facebook.net/D22581
Summary:
The very last reference happens in DBImpl::GetOptions()
I built with both DBImpl::GetOptions() and ColumnFamilyData::options() commented out
Test Plan: make all check
Reviewers: sdong, yhchiang, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D29073
Summary:
This diff adds three sets of APIs to RocksDB.
= GetColumnFamilyMetaData =
* This APIs allow users to obtain the current state of a RocksDB instance on one column family.
* See GetColumnFamilyMetaData in include/rocksdb/db.h
= EventListener =
* A virtual class that allows users to implement a set of
call-back functions which will be called when specific
events of a RocksDB instance happens.
* To register EventListener, simply insert an EventListener to ColumnFamilyOptions::listeners
= CompactFiles =
* CompactFiles API inputs a set of file numbers and an output level, and RocksDB
will try to compact those files into the specified level.
= Example =
* Example code can be found in example/compact_files_example.cc, which implements
a simple external compactor using EventListener, GetColumnFamilyMetaData, and
CompactFiles API.
Test Plan:
listener_test
compactor_test
example/compact_files_example
export ROCKSDB_TESTS=CompactFiles
db_test
export ROCKSDB_TESTS=MetaData
db_test
Reviewers: ljin, igor, rven, sdong
Reviewed By: sdong
Subscribers: MarkCallaghan, dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D24705
Summary: It turns out that -Wshadow has different rules for gcc than clang. Previous commit fixed clang. This commits fixes the rest of the warnings for gcc.
Test Plan: compiles
Reviewers: ljin, yhchiang, rven, sdong
Reviewed By: sdong
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D28131
Summary: as title
Test Plan: ./db_test
Reviewers: sdong, yhchiang, rven, igor
Reviewed By: igor
Subscribers: dhruba, leveldb
Differential Revision: https://reviews.facebook.net/D28269
Summary:
Make inplace_update_support and inplace_update_num_locks dynamic.
inplace_callback becomes immutable
We are almost free of references to cfd->options() in db_impl
Test Plan: unit test
Reviewers: igor, yhchiang, rven, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D25293
Summary:
This is not a critical options. Making it dynamic so that we can remove
more reference to cfd->options()
Test Plan: unit test
Reviewers: yhchiang, sdong, igor
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D24957
Summary:
make compaction related options changeable. Most of changes are tedious,
following the same convention: grabs MutableCFOptions at the beginning
of compaction under mutex, then pass it throughout the job and register
it in SuperVersion at the end.
Test Plan: make all check
Reviewers: igor, yhchiang, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D23349
Summary:
Previously, one single column family is given to WriteBatchWithIndex to index keys for all column families. An extra map from column family ID to comparator is maintained which can override the default comparator given in the constructor. A WriteBatchWithIndex::SetComparatorForCF() is added for user to add comparators per column family.
Also move more codes into anonymous namespace.
Test Plan: Add a unit test
Reviewers: ljin, igor
Reviewed By: igor
Subscribers: dhruba, leveldb, yhchiang
Differential Revision: https://reviews.facebook.net/D23355
Summary: as title
Test Plan:
make all check
I will think a way to set up stress test for this
Reviewers: sdong, yhchiang, igor
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D23055
Summary:
When memtable is full it calls the registered callback. That callback then registers column family as needing the flush. Every write checks if there are some column families that need to be flushed. This completely eliminates the need for MakeRoomForWrite() function and simplifies our Write code-path.
There is some complexity with the concurrency when the column family is dropped. I made it a bit less complex by dropping the column family from the write thread in https://reviews.facebook.net/D22965. Let me know if you want to discuss this.
Test Plan: make check works. I'll also run db_stress with creating and dropping column families for a while.
Reviewers: yhchiang, sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D23067
Summary: Correct some comments and typos in RocksDB.
Test Plan: Inspection
Reviewers: sdong, igor
Reviewed By: igor
Differential Revision: https://reviews.facebook.net/D23133
Summary:
Introducing WriteController, which is a source of truth about per-DB write delays. Let's define an DB epoch as a period where there are no flushes and compactions (i.e. new epoch is started when flush or compaction finishes). Each epoch can either:
* proceed with all writes without delay
* delay all writes by fixed time
* stop all writes
The three modes are recomputed at each epoch change (flush, compaction), rather than on every write (which is currently the case).
When we have a lot of column families, our current pull behavior adds a big overhead, since we need to loop over every column family for every write. With new push model, overhead on Write code-path is minimal.
This is just the start. Next step is to also take care of stalls introduced by slow memtable flushes. The final goal is to eliminate function MakeRoomForWrite(), which currently needs to be called for every column family by every write.
Test Plan: make check for now. I'll add some unit tests later. Also, perf test.
Reviewers: dhruba, yhchiang, MarkCallaghan, sdong, ljin
Reviewed By: ljin
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D22791
Summary:
As a preparation to support updating some options dynamically, I'd like
to first introduce ImmutableOptions, which is a subset of Options that
cannot be changed during the course of a DB lifetime without restart.
ColumnFamily will keep both Options and ImmutableOptions. Any component
below ColumnFamily should only take ImmutableOptions in their
constructor. Other options should be taken from APIs, which will be
allowed to adjust dynamically.
I am yet to make changes to memtable and other related classes to take
ImmutableOptions in their ctor. That can be done in a seprate diff as
this one is already pretty big.
Test Plan: make all check
Reviewers: yhchiang, igor, sdong
Reviewed By: sdong
Subscribers: leveldb, dhruba
Differential Revision: https://reviews.facebook.net/D22545
Summary:
I will move compression related options in a separate diff since this
diff is already pretty lengthy.
I guess I will also need to change JNI accordingly :(
Test Plan: make all check
Reviewers: yhchiang, igor, sdong
Reviewed By: igor
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D21915
Summary:
Add WriteBatchWithIndex so that a user can query data out of a WriteBatch, to support MongoDB's read-its-own-write.
WriteBatchWithIndex uses a skiplist to store the binary index. The index stores the offset of the entry in the write batch. When searching for a key, the key for the entry is read by read the entry from the write batch from the offset.
Define a new iterator class for querying data out of WriteBatchWithIndex. A user can create an iterator of the write batch for one column family, seek to a key and keep calling Next() to see next entries.
I will add more unit tests if people are OK about this API.
Test Plan:
make all check
Add unit tests.
Reviewers: yhchiang, igor, MarkCallaghan, ljin
Reviewed By: ljin
Subscribers: dhruba, leveldb, xjin
Differential Revision: https://reviews.facebook.net/D21381
Summary: const string& dbname parameter is not used
Test Plan: make all
Reviewers: sdong, igor
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D20703
Summary: Add a parameter path_id to DB::CompactRange(), to indicate where the output file should be placed to.
Test Plan: add a unit test
Reviewers: yhchiang, ljin
Reviewed By: ljin
Subscribers: xjin, igor, dhruba, MarkCallaghan, leveldb
Differential Revision: https://reviews.facebook.net/D20085
Summary:
Task 4580155. Some conditions in DBImpl::MakeRoomForWrite can be cached in
ColumnFamilyData, because theirs value can be changed only during compaction,
adding new memtable and/or add recalculation of compaction score.
These conditions are:
cfd->imm()->size() == cfd->options()->max_write_buffer_number - 1
cfd->current()->NumLevelFiles(0) >= cfd->options()->level0_stop_writes_trigger
cfd->options()->soft_rate_limit > 0.0 &&
(score = cfd->current()->MaxCompactionScore()) > cfd->options()->soft_rate_limit
cfd->options()->hard_rate_limit > 1.0 &&
(score = cfd->current()->MaxCompactionScore()) > cfd->options()->hard_rate_limit
P.S.
As it's my first diff, Siying suggested to add everybody as a reviewers
for this diff. Sorry, if I forgot someone or add someone by mistake.
Test Plan: make all check
Reviewers: haobo, xjin, dhruba, yhchiang, zagfox, ljin, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D19311
Summary:
We've seen some production issues where column family is detected as stale, although there is only one column family in the system. This is a quick fix that:
1) doesn't flush stale column families if there's only one of them
2) Use 4 as a coefficient instead of 2 for determening when a column family is stale. This will make flushing less aggressive, while still keep a nice dynamic flushing of very stale CFs.
Test Plan: make check
Reviewers: dhruba, haobo, ljin, sdong
Reviewed By: sdong
Subscribers: leveldb
Differential Revision: https://reviews.facebook.net/D18861
Summary:
Forward iterator puts everything together in a flat structure instead of
a hierarchy of nested iterators. this should simplify the code and
provide better performance. It also enables more optimization since all
information are accessiable in one place.
Init evaluation shows about 6% improvement
Test Plan: db_test and db_bench
Reviewers: dhruba, igor, tnovak, sdong, haobo
Reviewed By: haobo
Subscribers: sdong, leveldb
Differential Revision: https://reviews.facebook.net/D18795
Summary:
replace the super version acquisision in tailing itrator with thread
local
Test Plan: will post results
Reviewers: igor, haobo, sdong, yhchiang, dhruba
Reviewed By: igor
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17757
Summary:
Similar to GetImp(), use SuperVersion from thread local instead of acquriing mutex.
I don't expect this change will make a dent on NewIterator() performance
because the bottleneck seems to be on the rest part of the API
Test Plan:
make asan_check
will post perf numbers
Reviewers: haobo, igor, sdong, dhruba, yhchiang
Reviewed By: sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D17643
Summary: Client doesn't need to know anything about ColumnFamily ID. By making WriteBatch take ColumnFamilyHandle as a parameter, we can eliminate method GetID() from ColumnFamilyHandle
Test Plan: column_family_test
Reviewers: haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16887
Summary:
I'm cleaning up some code preparing for the big diff review tomorrow. This is the first part of the cleanup.
Changes are mostly cosmetic. The goal is to decrease amount of code difference between columnfamilies and master branch.
This diff also fixes race condition when dropping column family.
Test Plan: Ran db_stress with variety of parameters
Reviewers: dhruba, haobo
Differential Revision: https://reviews.facebook.net/D16833
Summary:
Column family IDs should be unique, even if column family is dropped. To achieve this, we save max column family in manifest.
Note that the diff is still not ready. I'm only using differential to move the patch to my Mac machine.
Test Plan: added a test to column_family_test
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16581
Summary: Added a function DeleteSuperVersion that can be called in DBImpl destructor before PurgingObsoleteFiles. That way, PurgeObsoleteFiles will be able to delete all files held by alive super versions.
Test Plan: column_family_test with valgrind
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16545
Summary:
I though I might get away with as little changes to LogAndApply() as possible. It turns out this is not the case.
This diff introduces different behavior of LogAndApply() for three cases:
1. column family add
2. column family drop
3. no-column family manipulation
(1) and (2) don't support group commit yet.
There were a lot of problems with old version od LogAndApply, detected by db_stress. The biggest was non-atomicity of manifest writes and metadata changes (i.e. if column family add is in manifest, it also has to be in in-memory data structure).
Test Plan: db_stress
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16491
Summary:
* Add ColumnFamilyHandle::GetID() function. Client needs to know column family's ID to be able to construct WriteBatch
* Handle WriteBatch::Handler failure gracefully. Since WriteBatch is not a very smart function (it takes raw CF id), client can add data to WriteBatch for column family that doesn't exist. In that case, we need to gracefully return failure status from DB::Write(). To do that, I added a return Status to WriteBatch functions PutCF, DeleteCF and MergeCF.
Test Plan: Added test to column_family_test
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16323
Summary:
The change to the public behavior:
* When opening a DB or creating new column family client gets a ColumnFamilyHandle.
* As long as column family handle is alive, client can do whatever he wants with it, even drop it
* Dropped column family can still be read from (using the column family handle)
* Added a new call CloseColumnFamily(). Client has to close all column families that he has opened before deleting the DB
* As soon as column family is closed, any calls to DB using that column family handle will fail (also any outstanding calls)
Internally:
* Ref-counting ColumnFamilyData
* New thread-safety for ColumnFamilySet
* Dropped column families are now completely dropped and their memory cleaned-up
Test Plan: added some tests to column_family_test
Reviewers: dhruba, haobo, kailiu, sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D16101
Summary: Revised thread-safety guarantees and implemented a way to spinlock the object.
Test Plan: make check
Reviewers: dhruba, haobo, sdong, kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15975
Summary:
WriteBatch can have multiple column families in one batch. Every column family has different options. So we have to add a way for write batch to get options for an arbitrary column family.
This required a bit more acrobatics since lots of interfaces had to be changed.
Test Plan: make check
Reviewers: dhruba, haobo, sdong, kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15957
Summary: Replaced most of occurrences of Options with more specific DBOptions. This brings us very close to supporting different configuration options for each column family.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15933
Summary:
Lots of code expects Options on construction/function call. My original idea was to split Options argument into ColumnFamilyOptions and DBOptions (the latter only if needed). However, this will require huge code changes very deep in the stack.
The better idea is to have ColumnFamilyData hold both ColumnFamilyOptions and Options. ColumnFamilyData::Options would be constructed from DBOptions (same for each column family) and ColumnFamilyOptions (different for each column family)
Now when we construct a class or call any method that requires Options, we can just push him ColumnFamilyData::Options and be sure that it's using column-family-specific settings.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15927
Summary:
Adapting table cache to column families is interesting. We want table cache to be global LRU, so if some column families are use not as often as others, we want them to be evicted from cache. However, current TableCache object also constructs tables on its own. If table is not found in the cache, TableCache automatically creates new table. We want each column family to be able to specify different table factory.
To solve the problem, we still have a single LRU, but we provide the LRUCache object to TableCache on construction. We have one TableCache per column family, but the underyling cache is shared by all TableCache objects.
This allows us to have a global LRU, but still be able to support different table factories for different column families. Also, in the future it will also be able to support different directories for different column families.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15915
Summary: InternalStats is a messy thing, keeping both DB data and column family data. However, it's better off living in ColumnFamilyData than in DBImpl. For now, at least.
Test Plan: make check
Reviewers: dhruba, kailiu, haobo, sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15879
Summary:
There are three SanitizeOption-s now : one for DBOptions, one for ColumnFamilyOptions and one for Options (which just calls the other two)
I have also reshuffled some options -- table_cache options and info_log should live in DBOptions, for example.
Test Plan: make check doesn't complain
Reviewers: dhruba, haobo, kailiu, sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15873
Summary: Compaction picker and internal key comparator are different for each column family (not global), so they should live in ColumnFamilyData
Test Plan: make check
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15801
Summary:
Sometimes we iterate through column families, and unlock the mutex in the body of the iteration. While mutex is unlocked, some column family might be created or dropped. We need to be able to continue iterating through column families even though our current column family got dropped.
This diff implements circular linked lists that connect all column families. It then uses the link list to enable iterating through linked lists. Even if the column family is dropped, its next_ pointer still can be used to advance to another alive column family.
Test Plan: make check
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15603
Summary: Making room for write will be the hardest part of the column family implementation. For now, I just iterate through all column families and run MakeRoomForWrite() for every one.
Test Plan: make check does not complain
Reviewers: dhruba, haobo
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15597
Summary: ColumnFamilyData grew a lot, there's much more data that it holds now. It makes more sense to encapsulate it better by making it a class.
Test Plan: make check
Reviewers: dhruba, haobo, kailiu, sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15579
Summary: This one is big. It adds ability to write to and read from different column families (see the unit test). It also supports recovery of different column families from log, which was the hardest part to reason about. We need to make sure to never delete the log file which has unflushed data from any column family. To support that, I added another concept, which is versions_->MinLogNumber()
Test Plan: Added a unit test in column_family_test
Reviewers: dhruba, haobo, sdong, kailiu
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15537
Summary: All memtables and immutable memtables are moved from DBImpl to ColumnFamilyData. For now, they are all referenced from default column family in DBImpl. It shouldn't be hard to get them from custom column family.
Test Plan: make check
Reviewers: dhruba, kailiu, sdong
CC: leveldb
Differential Revision: https://reviews.facebook.net/D15459
Igor Canadi