Summary: readhotwriterandom now generates a certain number of keys randomly. This number is a percentage (determined by hotkeyspercentage) of --num. Then when a key is needed one of the keys in that list is chosen randomly.
Test Plan: Ran db_bench with small --num parameter to see if the keys are random and that we only use a subset of the "hot" keys.
Removed Cache from metrics flow for Block::MetricsIter to DBImpl
Summary:
Before this patch we had to send metrics to `Cache`, which then sent the metrics to `DBImpl`. Now we directly send them to `DBImpl`, by the use of a pointer to a pointer of a `BlockMetrics` instance. We can do this safely and without additional locks because:
* `Block::MetricsIter` runs in the same thread as `DBImpl::Get()` (in which the `BlockMetrics` pointer resides)
* `DBImpl::Get()` locks `DBImpl` again after the `Block::MetricsIter` is destroyed, so we don't have to do an additional lock.
Hotcold: Added ability to select which levels record metrics
Summary: Added an option to `Options` to specify which levels have metrics recorded and used that option in `Version::Get()` to stop metrics being recorded for wrong levels and in `DBImpl` to stop `DoCompactionWork()` trying to split on hotcold for levels where no metrics are available.
Hotcold: Removed Table from metrics flow path and now only create BlockMetrics in Block::MetricsIter
Summary: Metrics no longer pass through `Table` to the cache and instead go directly from `Block::MetricsIter`. Additionally `BlockMetrics` is now only created when a `Seek()` is performed.
Hotcold: Short circuiting behaviour for file lookups when finding key in hot file
Summary:
Before this update both hot and cold file had to be queried in order to get the latest value for a given key. Now if the key is found in the hot file then we know the latest value is in the hot file.
Changed compaction to not allow a user key to have a record with a smaller sequence number in the hot file if there is a record with a larger sequence number in the cold file.
Added short circuiting to `Version::Get()` to assume that if a key is found in a newer file then it has a newer sequence number.
Test Plan:
`make check`
ran `db_bench` compiled with assertions on and in hot cold mode.
Hotcold: Changing compaction based on hotness of record.
Summary:
Changed `DoCompactionWork()` to output multiple files and imported https://reviews.facebook.net/D8025 (Ideally all issues with version_set.h, version_set.cc and version_set_test.cc should be brought up there so I can keep that patch up to date and we can decide whether to mainline it separately) so we can handle overlapping files in level 1+.
Added code to determine whether a record is hot or cold given the metrics DB and an iterator on the database (which allows us to get all the necessary data in order to lookup the hotness)
projects / rocksdb.git / log