TerarkDB 是字節(jié)跳動(dòng)開(kāi)源的?RocksDB 的替代品，具有優(yōu)化的尾部延遲、吞吐量和壓縮等。在大多數(shù)情況下，你可以將現(xiàn)有的 RocksDB 實(shí)例遷移到 TerarkDB，沒(méi)有任何問(wèn)題。

• TerarkDB 僅在 Linux 平臺(tái)下測(cè)試和生產(chǎn)就緒
• 除 C/C++ 之外的語(yǔ)言綁定尚未完全測(cè)試。
• 現(xiàn)有數(shù)據(jù)可以直接從 RocksDB 遷移到 TerarkDB，但不能遷移回 RocksDB。
• TerarkDB 是從 RocksDB v5.18.3 分支出來(lái)的

性能概覽

• RocksDB v6.12
• 服務(wù)器
  • Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz（2 插槽，32 核 64 線程）
  • 376 GB 動(dòng)態(tài)內(nèi)存
  • NVMe TLC 固態(tài)硬盤（3.5 TB）
• 工作臺(tái)工具和工作負(fù)載
  • 利用?db_bench
  • 10 個(gè)客戶端線程，每個(gè)線程 20GB 請(qǐng)求
  • 鍵 = 24 字節(jié)，值 = 2000 字節(jié)
  • heavy_write?意味著 90% 的寫操作
  • heavy_read?意味著 90% 的讀取操作

用法

• 基于塊的表（BlockBasedTable）

#include 
#include "rocksdb/db.h"

rocksdb::DB* db;
rocksdb::Options options;

// Your options here
options.create_if_missing = true;
options.wal_bytes_per_sync = 32768;
options.bytes_per_sync = 32768;

// Open DB
auto status = rocksdb::DB::Open(options, "/tmp/testdb", &db);

// Operations
std::string value;
auto s = db->Put(rocksdb::WriteOptions(), "key1", "value1");
s = db->Get(rocksdb::ReadOptions(), "key1", &value);
assert(s.ok());
assert("value1" == value);

s = db->Delete(rocksdb::WriteOptions(), "key1");
assert(s.ok());

或手動(dòng)設(shè)置表格格式和表格選項(xiàng):

#include 
#include "rocksdb/db.h"
#include "rocksdb/options.h"
#include "rocksdb/table.h"

rocksdb::DB* db;
rocksdb::Options options;

// Your db options here
options.create_if_missing = true;
options.wal_bytes_per_sync = 32768;
options.bytes_per_sync = 32768;

// Manually specify target table and table options
rocksdb::BlockBasedTableOptions table_options;
table_options.block_cache =
    rocksdb::NewLRUCache(32ULL << 30, 8, false);
table_options.block_size = 8ULL << 10;
options.table_factory = std::shared_ptr
                          (NewBlockBasedTableFactory(table_options));

// Open DB
auto status = rocksdb::DB::Open(options, "/tmp/testdb2", &db);

// Operations
std::string value;
auto s = db->Put(rocksdb::WriteOptions(), "key1", "value1");
s = db->Get(rocksdb::ReadOptions(), "key1", &value);
assert(s.ok());
assert("value1" == value);

s = db->Delete(rocksdb::WriteOptions(), "key1");
assert(s.ok());

• Terark 壓縮表格（TerarkZipTable）

#include 
#include "rocksdb/db.h"
#include "rocksdb/options.h"
#include "rocksdb/table.h"
#include "table/terark_zip_table.h"

rocksdb::DB* db;
rocksdb::Options options;

// Your db options here
options.create_if_missing = true;
options.wal_bytes_per_sync = 32768;
options.bytes_per_sync = 32768;

// TerarkZipTable need a `fallback` options because you can indicate which LSM level you want to start using TerarkZipTable
// For example, by setting tzt_options.terarkZipMinLevel = 2, TerarkDB will use your fallback Table on level 0 and 1.
std::shared_ptr table_factory;
rocksdb::BlockBasedTableOptions blockbased_options;
blockbased_options.block_size = 8ULL << 10;
table_factory.reset(NewBlockBasedTableFactory(blockbased_options));

rocksdb::TerarkZipTableOptions tzt_options;
// TerarkZipTable requires a temp directory other than data directory, a slow device is acceptable
tzt_options.localTempDir = "/tmp";
tzt_options.indexNestLevel = 3;
tzt_options.sampleRatio = 0.01;
tzt_options.terarkZipMinLevel = 2; // Start using TerarkZipTable from level 2

table_factory.reset(
    rocksdb::NewTerarkZipTableFactory(tzt_options, table_factory));

options.table_factory = table_factory;

// Open DB
auto status = rocksdb::DB::Open(options, "/tmp/testdb2", &db);

// Operations
std::string value;
auto s = db->Put(rocksdb::WriteOptions(), "key1", "value1");
s = db->Get(rocksdb::ReadOptions(), "key1", &value);
assert(s.ok());
assert("value1" == value);

s = db->Delete(rocksdb::WriteOptions(), "key1");
assert(s.ok());

實(shí)際性能改進(jìn)

TerarkDB 已經(jīng)部署在字節(jié)跳動(dòng)的許多應(yīng)用程序中，在大多數(shù)情況下，TerarkDB 可以幫助減少延遲峰值并極大地提高吞吐量。

磁盤寫入（Disk Write）

獲取延遲(Get Latency?)