Skip to content

Pika is a NoSQL database compatible with redis which is developed by Qihoo's infrastructure team.

License

Notifications You must be signed in to change notification settings

chenbt-hz/pika

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

YnbjQf.png

Build Status Downloads

Stargazers Over Time Contributors Over Time
Stargazers over time Contributor over time

Introduction中文

Pika is a high-performance, large-capacity, multi-tenant, data-persistent elastic KV data storage system using RocksDB as the storage engine. It is fully compatible with the Redis protocol and supports its commonly used data structures, such as string/hash/list/zset/set/geo/hyperloglog/pubsub/bitmap/stream, etc. Redis Interface.

When Redis's in-memory usage exceeds 16GiB, it faces problems such as limited memory capacity, single-threaded blocking, long startup recovery time, high memory hardware costs, easily filled buffers, and high switching costs when one master and multiple replicas fail. The emergence of Pika is not to replace Redis but to complement it. Pika strives to completely comply with the Redis protocol, inherit Redis's convenient operation and maintenance design, and solve the bottleneck problem of Redis running out of memory capacity once the data volume becomes huge by using persistent storage. Additionally, Pika can support master-slave mode using the slaveof command, and it also supports full and incremental data synchronization.

Pika can be deployed in a single-machine master-slave mode (slaveof) or in a Codis cluster mode, allowing for simple scaling and shrinking. Migration from Redis to Pika can be smoothly executed by tools.

Pika Features

  • Protocol Compatibility: Fully compatible with the Redis protocol, emphasizing high performance, large capacity, low cost, and scalability.
  • Data Structures: Supports Redis's common data structures, including String, Hash, List, Zset, Set, Geo, Hyperloglog, Pubsub, Bitmap, Stream, ACL, etc.
  • Cold and Hot Data: Caches hot data and persistently stores the full data in RocksDB, implementing a hierarchical storage of cold and hot data.
  • High Capacity: Compared to Redis's in-memory storage, Pika supports data volumes in the hundreds of gigabytes, significantly reducing server resource consumption and enhancing data reliability.
  • Deployment Modes: Supports single-machine master-slave mode (slaveof) and Codis cluster mode, making scaling and shrinking simple.
  • Easy Migration: Smooth migration from Redis to Pika without modifying code.
  • Convenient Operation and Maintenance: Comprehensive operation and maintenance command documentation.

Pika Storage Engine Architecture

  • Supports multiple platforms: CentOS, Ubuntu, macOS
  • Multi-threaded model
  • Based on the RocksDB storage engine
  • Multiple granularity data caching model

Deployment Modes

1. Master-Slave Mode

  • Architecture similar to Redis
  • Good compatibility with Redis protocol and data structures
  • Each data structure uses a separate RocksDB instance
  • Master-slave adopts binlog asynchronous replication

Pika-Master-Slave

2. Distributed Cluster Mode

  • Adopts Codis architecture, supports multiple groups
  • Each group forms a master-slave set
  • Elastic scaling based on groups

Pika-Cluster

Pika User Showcase

Qihoo 360game Weibo Garena
Apus Ffan Meituan XES
HX XL GWD DYD
YM XM XL YM
MM VIP LK KS

Pika has been widely adopted by various companies for internal deployments, demonstrating its scalability and reliability. Some notable usage instances include:

  • 360 Company: Internal deployment with a scale of 10,000+ instances, each having a data volume of 1.8TB.
  • Weibo: Internal deployment with 10,000+ instances.
  • Ximalaya(Xcache): 6,000+ instances with a massive data volume exceeding 120TB.
  • Getui (个推) Company: Internal deployment involving 300+ instances, with a cumulative data volume surpassing 30TB.

Additionally, Pika is utilized by companies such as Xunlei, Xiaomi, Zhihu, New Oriental Education & Technology Group (好未来), Kuaishou, Sohu, Meituan, Maimai, and more. For a comprehensive list of users, you can refer to the official list provided by the Pika project.

These deployments across a diverse range of companies and industries underscore Pika's adaptability and effectiveness in handling large-scale, high-volume data storage requirements.

More

Getting Started with Pika

1. Binary Package Installation

Users can directly download the latest binary version package from releases.

2. Compilation from Source

  • 2.1 Supported Platforms

    • Linux - CentOS
    • Linux - Ubuntu
    • macOS(Darwin)
  • 2.2 Required Library Software

    • gcc g++ supporting C++17 (version >= 9)
    • make
    • cmake (version >= 3.18)
    • autoconf
    • tar
  • 2.3 Compilation Process

    • 2.3.1. Get the source code

        git clone https://github.com/OpenAtomFoundation/pika.git
    • 2.3.2. Switch to the latest release version

        git tag          # Check the latest release tag (e.g., v3.4.1)
        git checkout TAG # Switch to the latest version (e.g., git checkout v3.4.1)
    • 2.3.3. Execute compilation

      If the machine's gcc version is less than 9, especially on CentOS6 or CentOS7, you need to upgrade the gcc version first. Execute the following commands:

        sudo yum -y install centos-release-scl
        sudo yum -y install devtoolset-9-gcc devtoolset-9-gcc-c++
        scl enable devtoolset-9 bash

      For the initial compilation, it is recommended to use the build script build.sh, which checks if the required software is available on the local machine.

        ./build.sh

      Note: The compiled files will be saved in the output directory.

      Pika is compiled by default in release mode, which does not support debugging. If debugging is needed, compile in debug mode.

        rm -rf output/
        cmake -B output -DCMAKE_BUILD_TYPE=Debug
        cd output && make

      Other components, such as codis, can also be compiled using build.sh.

        # Compile codis, default target, build-all
        ./build.sh codis
      
        # Compile codis, but only build codis-proxy
        ./build.sh codis codis-proxy
      
  • 2.4 Start Pika

      ./output/pika -c ./conf/pika.conf
  • 2.5 Clear Compiled Results

    If you need to clear the compilation content, you can choose one of the following methods based on the situation:

    Method 1: Clean only the current compilation content

      cd output && make clean

    Method 2: Completely recompile

      rm -rf output # regenerate cmake
  • 2.6 Pika Development Debugging

    Setting up Pika Development Environment with CLion

3. Containerization

  • 3.1 Running with Docker

    docker run -d \
      --restart=always \
      -p 9221:9221 \
      -v <log_dir>:/pika/log \
      -v <db_dir>:/pika/db \
      -v <dump_dir>:/pika/dump \
      -v <dbsync_dir>:/pika/dbsync \
      pikadb/pika:v3.3.6
    
    redis-cli -p 9221 "info"
  • 3.2 Build Custom Image

    If you want to build your own image, we provide a script build_docker.sh to simplify the process.

    This script accepts several optional parameters:

    • -t tag: Specify the Docker tag for the image. By default, the tag is pikadb/pika:.
    • -p platform: Specify the platform for the Docker image. Options include all, linux/amd64, linux/arm, linux/arm64. By default, it uses the current docker platform setting.
    • --proxy: Use a proxy to download packages to speed up the build process. The build will use Alibaba Cloud's image source.
    • --help: Display help information.

    Here is an example usage:

    ./build_docker.sh -p linux/amd64 -t private_registry/pika:latest
  • 3.3 Running with docker-compose

docker-compose.yaml

  pikadb:
    image: pikadb/pika:lastest
    container_name: pikadb
    ports:
      - "6379:9221"
    volumes:
      - ./data/pika:/pika/log
      # Specify the configuration file path. If you need to specify a configuration file, specify it here.
      # Note: pika.conf should be in the ./deploy/pika directory
      #- ./deploy/pika:/pika/conf
      - ./data/pika/db:/pika/db
      - ./data/pika/dump:/pika/dump
      - ./data/pika/dbsync:/pika/dbsync
    privileged: true
    restart: always

Performance test

  • Thanks deep011 for providing performance test results.

Note: The test results were obtained under specific conditions and scenarios, and may not represent the performance in all environments and scenarios. They are for reference only.

We recommend that you conduct detailed testing of Pika in your own environment based on the usage scenario to assess whether Pika meets your requirements.

1. Test environment

  • CPU Model: Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz
  • CPU Threads: 56
  • Memory: 256GB
  • Disk: 3TB Flash
  • Network: 10GBase-T/Full * 2
  • Operating System: CentOS 6.6
  • Pika Version: 2.2.4

2. Benchmarking Tool

vire-benchmark

3. Test Cases

3.1 Case 1

  • Test Objective

Evaluate the upper limit of QPS for Pika under different worker thread counts.

  • Test Conditions
    • Pika Data Size: 800GB
    • Value: 128 bytes
    • CPU not bound
  • Test Results
    1

    Note: The x-axis represents Pika thread count, and the y-axis represents QPS with a value size of 128 bytes. "set3/get7" indicates 30% set and 70% get operations.

  • Case One Conclusion

    From the above graph, it can be observed that setting Pika's worker thread count to 20-24 is more cost-effective.

3.2 Case 2

  • Test Objective

    Evaluate the RTT performance of Pika with the optimal worker thread count (20 threads).

  • Test Conditions
    • Pika Data Size: 800GB
    • Value: 128 bytes
  • Test Results
    ====== GET ======
      10000000 requests completed in 23.10 seconds
      200 parallel clients
      3 bytes payload
      keep alive: 1
    99.89% <= 1 milliseconds
    100.00% <= 2 milliseconds
    100.00% <= 3 milliseconds
    100.00% <= 5 milliseconds
    100.00% <= 6 milliseconds
    100.00% <= 7 milliseconds
    100.00% <= 7 milliseconds
    432862.97 requests per second
    ====== SET ======
      10000000 requests completed in 36.15 seconds
      200 parallel clients
      3 bytes payload
      keep alive: 1
    91.97% <= 1 milliseconds
    99.98% <= 2 milliseconds
    99.98% <= 3 milliseconds
    99.98% <= 4 milliseconds
    99.98% <= 5 milliseconds
    99.98% <= 6 milliseconds
    99.98% <= 7 milliseconds
    99.98% <= 9 milliseconds
    99.98% <= 10 milliseconds
    99.98% <= 11 milliseconds
    99.98% <= 12 milliseconds
    99.98% <= 13 milliseconds
    99.98% <= 16 milliseconds
    99.98% <= 18 milliseconds
    99.99% <= 19 milliseconds
    99.99% <= 23 milliseconds
    99.99% <= 24 milliseconds
    99.99% <= 25 milliseconds
    99.99% <= 27 milliseconds
    99.99% <= 28 milliseconds
    99.99% <= 34 milliseconds
    99.99% <= 37 milliseconds
    99.99% <= 39 milliseconds
    99.99% <= 40 milliseconds
    99.99% <= 46 milliseconds
    99.99% <= 48 milliseconds
    99.99% <= 49 milliseconds
    99.99% <= 50 milliseconds
    99.99% <= 51 milliseconds
    99.99% <= 52 milliseconds
    99.99% <= 61 milliseconds
    99.99% <= 63 milliseconds
    99.99% <= 72 milliseconds
    99.99% <= 73 milliseconds
    99.99% <= 74 milliseconds
    99.99% <= 76 milliseconds
    99.99% <= 83 milliseconds
    99.99% <= 84 milliseconds
    99.99% <= 88 milliseconds
    99.99% <= 89 milliseconds
    99.99% <= 133 milliseconds
    99.99% <= 134 milliseconds
    99.99% <= 146 milliseconds
    99.99% <= 147 milliseconds
    100.00% <= 203 milliseconds
    100.00% <= 204 milliseconds
    100.00% <= 208 milliseconds
    100.00% <= 217 milliseconds
    100.00% <= 218 milliseconds
    100.00% <= 219 milliseconds
    100.00% <= 220 milliseconds
    100.00% <= 229 milliseconds
    100.00% <= 229 milliseconds
    276617.50 requests per second
  • Case 2 Conclusion

    The response time for 99.9% of get/set operations is within 2ms.

3.3 Case 3

  • Test Objective

    Evaluate the maximum QPS for each command in Pika with the optimal worker thread count.

  • Test Conditions
    • Pika Worker Thread Count: 20
    • Number of Keys: 10,000
    • Number of Fields: 100 (excluding lists)
    • Value: 128 bytes
    • Number of Command Executions: 10 million (except for lrange)
  • Test Results
    PING_INLINE: 548606.50 requests per second
    PING_BULK: 544573.31 requests per second
    SET: 231830.31 requests per second
    GET: 512163.91 requests per second
    INCR: 230861.56 requests per second
    MSET (10 keys): 94991.12 requests per second
    LPUSH: 196093.81 requests per second
    RPUSH: 195186.69 requests per second
    LPOP: 131156.14 requests per second
    RPOP: 152292.77 requests per second
    LPUSH (needed to benchmark LRANGE): 196734.20 requests per second
    LRANGE_10 (first 10 elements): 334448.16 requests per second
    LRANGE_100 (first 100 elements): 50705.12 requests per second
    LRANGE_300 (first 300 elements): 16745.16 requests per second
    LRANGE_450 (first 450 elements): 6787.94 requests per second
    LRANGE_600 (first 600 elements): 3170.38 requests per second
    SADD: 160885.52 requests per second
    SPOP: 128920.80 requests per second
    HSET: 180209.41 requests per second
    HINCRBY: 153364.81 requests per second
    HINCRBYFLOAT: 141095.47 requests per second
    HGET: 506791.00 requests per second
    HMSET (10 fields): 27777.31 requests per second
    HMGET (10 fields): 38998.52 requests per second
    HGETALL: 109059.58 requests per second
    ZADD: 120583.62 requests per second
    ZREM: 161689.33 requests per second
    PFADD: 6153.47 requests per second
    PFCOUNT: 28312.57 requests per second
    PFADD (needed to benchmark PFMERGE): 6166.37 requests per second
    PFMERGE: 6007.09 requests per second
  • Conclusion

    Overall performance is excellent, but some commands exhibit weaker performance (LRANGE, PFADD, PFMERGE).

3.4 Case 4

  • Test Objective

    Compare the maximum QPS between Pika and Redis.

  • Test Conditions
    • Pika Worker Thread Count: 20
    • Number of Keys: 10,000
    • Number of Fields: 100 (excluding lists)
    • Value: 128 bytes
    • Number of Command Executions: 10 million (except for LRANGE)
    • Redis Version: 3.2.0
  • Test Result

1

Observability

Metrics

  1. Pika Server Info: system, ip, port, run_id, config file etc.
  2. Pika Data Info: db size, log size, memory usage etc.
  3. Pika Client Info: The number of connected clients.
  4. Pika Stats Info: status information of compact, slot, etc.
  5. Pika Network Info: Incoming and outgoing traffic and rate of client and master-slave replication.
  6. Pika CPU Info: cpu usage.
  7. Pika Replication Info: Status information of master-slave replication, binlog information.
  8. Pika Keyspace Info: key information of five data types.
  9. Pika Command Exec Count Info: command execution count.
  10. Pika Command Execution Time: Time-consuming command execution.
  11. RocksDB Metrics: RocksDB information of five data types, includes Memtable, Block Cache, Compaction, SST File, Blob File etc.

More details on Metrics.

Documents

Contact Us

About

Pika is a NoSQL database compatible with redis which is developed by Qihoo's infrastructure team.

Resources

License

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 68.9%
  • Go 17.8%
  • Tcl 7.0%
  • Python 1.7%
  • Shell 1.0%
  • JavaScript 0.9%
  • Other 2.7%