/terarkdb

A RocksDB compatible KV storage engine with better performance

Primary LanguageC++Apache License 2.0Apache-2.0

About TerarkDB

TerarkDB is a RocksDB replacement with optimized tail latency, throughput and compression etc. In most cases you can migrate your existing RocksDB instance to TerarkDB without any drawbacks.

NOTES

  • TerarkDB was only tested and production ready under Linux platform
  • Language bindings except C/C++ are not fully tested yet.
  • Existing data can be migrated from RocksDB directly to TerarkDB, but cannot migrate back to RocksDB.
  • TerarkDB was forked from RocksDB v5.18.3.

Performance Overview

  • RocksDB v6.12
  • Server
    • Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz (2 Sockets, 32 cores 64 threads)
    • 376 GB DRAM
    • NVMe TLC SSD (3.5 TB)
  • Bench Tools & Workloads
    • use db_bench
    • 10 client threads, 20GB requests per thread
    • key = 24 bytes, value = 2000 bytes
    • heavy_write means 90% write operations
    • heavy_read means 90% read operations

1. Use TerarkDB

Prerequisite

If you enabled TerarkZipTable support (-DWITH_TERARK_ZIP=ON), you should install libaio before compile TerarkDB:

sudo apt-get install libaio-dev

If this is your first time using TerarkDB, we recommend you to use without TerarkZipTable by changing -DWITH_TERARK_ZIP to OFF in build.sh.

Method 1: Use CMake subdirectory (Recommend)

  1. Clone
cd {YOUR_PROJECT_DIR}
git submodule add https://github.com/bytedance/terarkdb.git

cd terarkdb && git submodule update --init --recursive
  1. Edit your Top Project's CMakeLists.txt
add_subdirectory(terarkdb)
target_link_libraries({YOUR_TARGET} terarkdb)
  1. Important Default Options
  • CMAKE_BUILD_TYPE: RelWithDebInfo
  • WITH_JEMALLOC: ON
    • Use Jemalloc or Not (If you are using a different malloc library, change to OFF)
  • WITH_TESTS: OFF
    • Build test cases
  • WITH_TOOLS: OFF
    • Build with TerarkDB tools (e.g. db_bench, ldb etc)
  • WITH_TERARK_ZIP: OFF
    • Build with TerarkZipTable
  • WITH_ZNS: OFF
    • Build with ZNS device support

Notes

  • TerarkDB is built with zstd, lz4, snappy, zlib, gtest, boost by default, if you need these libraries, you can remove them from your higher level application.

Method 2: Link as static library

  1. clone & build
git clone https://github.com/bytedance/terarkdb.git

cd terarkdb && git submodule update --init --recursive

WITH_TESTS=OFF WITH_ZNS=OFF ./build.sh
  1. linking

Directory:

  terarkdb/
        \___ output/
                \_____ include/
                \_____ lib/
                         \___ libterarkdb.a
                         \___ libzstd.a
                         \___ ...

We didn't archieve all static libraries together yet, so you have to pack all libraries to your target:

-Wl,-Bstatic \
-lterarkdb -lbz2 -ljemalloc -llz4 -lsnappy -lz -lzstd \
-Wl,-Bdynamic -pthread -lgomp -lrt -ldl -laio

2. Usage

2.1. BlockBasedTable

#include <cassert>
#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());

Or manually set table format and table options:

#include <cassert>
#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<rocksdb::TableFactory>
                          (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());

2.2. TerarkZipTable

#include <cassert>
#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<rocksdb::TableFactory> 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());

3. Real-world Performance Improvement

TerarkDB has been deployed in lots of applications in Bytedance, in most cases TerarkDB can help to reduce latency spike and improve throughput tremendously.

Disk Write

Get Latency (us)

4. Contributing

  • TerarkDB uses Github issues and pull requests to manage features and bug fixes.
  • All PRs are welcome including code formating and refactoring.

5. License

  • Apache 2.0

6. Users

Please let us know if you are using TerarkDB, thanks! (guokuankuan@bytedance.com)

  • ByteDance (core online services)