RocksDB optimization

[ShooterIT]

Compression type

• Support multiple compression types

Parameters

• enable enable_pipelined_write if having more CPUs
• use index_type: kTwoLevelIndexSearch to make index and filter blocks smaller
• enable pin_top_level_index_and_filter, pin_l0_filter_and_index_blocks_in_cache
• for cache_index_and_filter_blocks_with_high_priority, to avoid too high
• NewBloomFilterPolicy(bits_per_key, false) i.e. BloomFilterPolicy::kAutoBloom
• data_block_index_type: BlockBasedTableOptions::kDataBlockBinaryAndHash
• Set BlockBasedTableOptions::optimize_filters_for_memory to true, jemalloc_usable_size is safe if no bugs, redis adopts it
• memtable_prefix_bloom_size_ratio and memtable_whole_key_filtering, to enable memtable filter.
• CompressionOptions::max_dict_bytes for lz4 and zstd

Data partition by ColumnFamily

• One type one ColumnFamily

[dongdongwcpp]

memtable_prefix_bloom_size_ratio seems need a prefix_extractor, can it work for redis like system?

[ShooterIT]

@dongdongwcpp you mean we should set prefix_extractor if we want to enable memtable_prefix_bloom_size_ratio?
i think in redis system, there are many lookup reads, so i want to enable this feature, make sense? WDYT?