Hybrid Storage Edition (Cluster Architecture)

Last updated: 2020-09-17 11:36:53

    TencentDB for Redis Hybrid Storage Edition (cluster architecture) is based on Tendis, a KV (key-value) RocksDB storage engine developed by and widely used in Tencent. It is compatible with Redis protocols and features high performance, high compression ratio and high stability. Tencent has extensive experience in Tendis operation.

    Hybrid Storage Edition (cluster architecture) is composed of two components, i.e., Redis (cache) and Tendis (storage engine). It is suitable for KV storage scenarios, as it balances performance and cost, and greatly reduces your business operating costs by 80% in the scenarios where cold data takes up a lot of storage space.

    Hybrid Storage Edition (cluster architecture) is fully compatible with Redis 4.0 Cluster Edition commands. It is easy to use and can make full use of a rich variety of data structures and operational commands of Redis for efficiency.

    Hybrid Storage Edition (cluster architecture) stores all data on disk, and caches all keys and the values of hot keys in the memory. The system architecture is shown as below:


    Low cost

    • Data is automatically cached and automatically degraded to cold data. All data is stored on disk, and hot data is cached in the memory. Hybrid Storage Edition reduces operating costs by 40% to 80% compared with TencentDB for Redis Memory Edition.
    • Hybrid Storage Edition adopts the LZ4 data compression algorithm to automatically compress data once stored on disk, which balances performance and capacity and saves up to 90% of the disk capacity.

    High efficiency

    • With 100% compatibility with Redis protocols, all efficient Redis data structures and APIs can be used in the business.
    • In Hybrid Storage Edition, the business does not need to swap hot and cold data, or deal with the data inconsistency, cache breakdown, cache avalanche and other problems existing in traditional caching schemes. Hybrid Storage Edition reduces the complexity of the business, improves the development efficiency and reduces the OPS cost.

    High performance

    • Up to 3,000,000+ QPS for hot data access, comparable to native Redis
    • Up to 1,000,000 QPS for concurrent writes

    Large capacity

    • Super-large storage capacity of 240 GB to 32 TB
    • The data stored on disk can have 6 replicas, fully ensuring data reliability.


    The core components of TencentDB for Redis Hybrid Storage Edition (cluster architecture) include Proxy, Redis cache, and Tendis engine, as described below:

    • Proxy: it routes and distributes client requests, distributes commands to the correct shard according to their keys, collects part of monitoring data, and disables high-risk commands online, etc.
    • Redis cache: it is based on Redis 4.0 Cluster Edition. In order to achieve automatic degradation to cold data, Hybrid Storage Edition modifies core Redis features, including value eviction, value eviction based on time, synchronization of written data to Tendis, cold data access, and master/slave synchronization of hot data, etc. The modified Redis in Hybrid Storage Edition is 100% compatible with Redis 4.0 Cluster Edition commands.
    • Tendis engine: it is a KV storage engine developed by Tencent and compatible with Redis protocols. Tendis has been used in Tencent for many years with its performance and stability being fully verified. In the hybrid storage system, its key features include the storage and reading of full data, data backup, incremental log backup, etc.

    Application Scenarios

    Storage scenarios

    The data access acceleration scenario where Redis is used as the primary storage engine rather than the cache.

    Latency-insensitive scenario

    Because cold data needs to be read from disks, the latency of accessing cold data in Hybrid Storage Edition will significantly increase to tens or even hundreds of milliseconds, while the access delay in Redis Memory Edition is less than 1 ms.

    Distributed hot/cold data scenario

    Hybrid storage is ideal for scenarios where hot data degrades to cold data over time. Hot data is accessed frequently and requires short response latency. Cold data can have a short response latency after prefetch.

    Caching scenario

    • Insufficient performance: the caching scenario requires high performance, but the Hybrid Storage Edition’s performance in cold data is much inferior to that of Memory Edition. For more information about performance, please see Performance.
    • Cost increase: Hybrid Storage Edition increases the disk storage space and adds more disk storage compute nodes. Since the caching scenario requires memory capacity close to disk capacity, business costs will rise rather than fall.

    Latency-sensitive scenario

    Because cold data needs to be read from disks, the latency of accessing cold data in Hybrid Storage Edition will significantly increase to tens or even hundreds of milliseconds, while the access delay in Redis Memory Edition is less than 1 ms. Therefore, Hybrid Storage Edition is not recommended for a latency-sensitive scenario.



    • The minimum disk capacity must be greater than the memory capacity, otherwise the data may not be written.
    • The memory caches all keys and only evicts values, so the memory may not be able to cache all keys due to too small disk capacity configuration. Please evaluate the disk space.
    • Run the following set command with the 128-byte value to test the maximum write performance:
    redis-benchmark -h -p 6379 -c 100 -n 60000000 -r 1000000000 -d 128 -t set -a passwd
    Shard Quantity Total Cache Capacity (GB) Total Disk Capacity Range (GB) Maximum Write Performance (QPS)
    4 64 240 - 520 60,000
    4 128 480 - 960 60,000
    4 256 1,000 - 2,000 60,000
    8 128 480 - 960 120,000
    8 256 960 - 1,920 120,000
    8 512 2,000 - 4,000 120,000
    16 256 960 - 1,920 240,000
    16 512 1,920 - 3,840 240,000
    16 1,024 4,000 - 8,000 240,000
    32 512 3,840 - 7,680 480,000
    32 1,024 7,680 - 15,360 480,000
    32 2,048 16,000 - 32,000 480,000

    Degradation to Cold Data

    Value eviction policy

    • Value eviction only
      When you set a timeout period for keys or use the expire command, both keys and values will be evicted from the memory; otherwise, only values will be evicted.
    • value-eviction-policy
      • Through value-eviction-policy, you can set how many days a value has not been accessed before it is automatically evicted from the memory.
    • The default value of this parameter is 7 days. It can be modified by users in the console.
    • maxmemory-policy
      • Hybrid Storage Edition only supports allkeys-lru (default) and allkeys-random.
      • When Redis memory usage reaches maxmemory, the system evicts values from the memory according to maxmemory-policy.

    Value cache policy

    • value-cache-policy
      You can use this parameter to configure when disk data is cached into Redis memory: Redis caches a value into the memory if the number of times the value is accessed within 5 minutes is greater than value-cache-policy. This parameter can avoid cache invalidation caused by, such as, data traversal. If this parameter is configured to 1, the cold data is immediately cached.

    • The expire command description
      If you use Expire Time to set a timeout period for keys, Hybrid Storage Edition will follow the original semantics of this command to evict expired keys and values from the memory and disks. The same is true for keys set with EXPIRE, EXPIREAT, PEXPIRE, and PEXPIREAT commands.

    • Big key eviction
      To ensure reading performance, Hybrid Storage Edition currently does not evict a value from the memory, if the value is larger than 8 MB or has complex (non-string) structures with more than 1,000 fields. Therefore, Hybrid Storage Edition does not have an ideal effect on the degradation of complex data structures, such as large Hash structures, which will be continuously optimized in the future.

    Command Compatibility

    Hybrid Storage Edition (cluster architecture) stores data in a distributed manner, and its biggest difference from the Memory Edition (standard architecture) lies in whether a single command supports multikey access. For the cluster architecture, commands can be categorized into supported, custom, and unsupported. For the complete list of compatible commands, please see Command Compatibility.

    • Unsupported commands
      The system will return the following error:
      keys *
      (error) ERR unknown command 'keys'
    • Partially supported commands
      Hybrid Storage Edition (cluster architecture) is compatible with smart clients such as JedisCluster. For compatibility with JedisCluster, TencentDB for Redis modifies the IP list returned by the supported commands, and the IP address of each node in the returned information is the instance's VIP.

      • CONFIG GET
    • Supported cross-slot commands
      Currently, cross-slot access commands supported by Hybrid Storage Edition (cluster architecture) include MGET, MSET, and DEL but not other multikey commands.

    • Custom commands
      Through VIP encapsulation, Hybrid Storage Edition (cluster architecture) provides a user experience in cluster mode comparable to the standalone edition, making it much easier for use in different scenarios. To increase the transparency to OPS, custom commands can be used. Access to each node in the cluster is supported by adding a parameter node ID on the right of the original command parameter list, such as COMMAND arg1 arg2 ... [node ID]. The node ID can be obtained through the cluster nodes command or in the console.> cluster nodes
      25b21f1836026bd49c52b2d10e09fbf8c6aa1fdc slave 36034e645951464098f40d339386e9d51a9d7e77 0 1531471918205 1 connected
      da6041781b5d7fe21404811d430cdffea2bf84de master - 0 1531471916000 2 connected 10923-16383
      36034e645951464098f40d339386e9d51a9d7e77 myself,master - 0 1531471915000 1 connected 0-5460
      53f552fd8e43112ae68b10dada69d3af77c33649 slave da6041781b5d7fe21404811d430cdffea2bf84de 0 1531471917204 3 connected
      18090a0e57cf359f9f8c8c516aa62a811c0f0f0a slave ef3cf5e20e1a7cf5f9cc259ed488c82c4aa17171 0 1531471917000 2 connected
      ef3cf5e20e1a7cf5f9cc259ed488c82c4aa17171 master - 0 1531471916204 0 connected 5461-10922
      Native command:
      info server
      Custom command:
      info server ef3cf5e20e1a7cf5f9cc259ed488c82c4aa17171
      SCAN command examples:
      scan 0 238b45926a528c85f40ae89d6779c802eaa394a2
      scan 0 match a* 238b45926a528c85f40ae89d6779c802eaa394a2
      KEYS command example:
      keys a* 238b45926a528c85f40ae89d6779c802eaa394a2
    • Transactional support
      Hybrid Storage Edition (cluster architecture) supports transactional commands. The keys of a transaction should be stored in the same slot, and the keys of the WATCH command and transaction-related keys should be stored in the same slot too. HashTag is recommended for multikey transactions in cluster mode.

    • Multi-database support
      Hybrid Storage Edition (cluster architecture) supports the SELECT 0 command but not multiple databases.

    • Poor-performance commands

      • linsert and lrem: the linsert and lrem commands in the List command family have poor performance and are not recommended thus. They will traverse the list nodes in the disk with the O(n) execution time complexity. If there are many list nodes, the command execution will time out.
      • append: the append command performs poorly when the character size exceeds 1 MB.

    Was this page helpful?

    Was this page helpful?

    • Not at all
    • Not very helpful
    • Somewhat helpful
    • Very helpful
    • Extremely helpful
    Send Feedback