Hybrid-storage instances of Tair store data in both memory and disks to support data persistence and high read and write performance at the same time.

Note Hybrid-storage instances are phrased out. For more information, see End-of-Sale for ApsaraDB for Redis hybrid-storage instances. We recommend that you choose Tair persistent memory-optimized instances. For more information, see Persistent memory-optimized instances.


Figure 1. Architecture
Architecture of Tair hybrid-storage instances

Hybrid-storage instances of Tair are developed by Alibaba Cloud and are compatible with the Redis protocol. These instances store all data in disks and hot data in memory to ensure high read and write performance. Hybrid-storage instances deliver a cost-effective solution that supports high-performance queries for frequently accessed data without the limit of memory capacity.

  • In-memory storage: Keys and values of hot data are stored in memory. The information of all keys is also cached in memory. This way, you can check whether a specified key exists in a fast and convenient manner.
  • Disk storage: All keys and values are stored in disks. Redis data structures such as hashes are also stored in disks in a specific format.


Scenario Description
Livestreaming Livestreams generate large amounts of hot data. Most of the data comes from popular live channels. You can use hybrid-storage instances to store data from popular live channels in memory and data from less popular live channels in disks. This allows you to make full use of the limited memory capacity.
E-commerce E-commerce applications need to access large amounts of commodity data. You can use hybrid-storage instances to store data and optimize memory usage. The data of popular commodities is stored in memory, and the data of less popular commodities is stored in disks.
Online education

Online education applications need to access large amounts of data. Such data includes online courses, question libraries, and messages between teachers and students. Only popular courses and the latest question libraries are frequently accessed. You can use hybrid-storage instances to store the data of online courses in disks, and store the data of popular courses and question libraries in memory. Tair provides a cost-effective solution to ensure high read and write performance for frequently accessed data.

The following examples show the benefits of hybrid-storage instances:

  • Example 1: An open source Redis cluster is used to store 100 GB of data. The number of queries per second (QPS) at peak hours is less than 20,000, and 80% of the data is not frequently accessed.

    In this case, you can use a Tair hybrid-storage instance that has 32 GB memory and 128 GB disk capacity. This saves about 70 GB of memory resources and reduces storage costs by more than 50%.

  • Example 2: An on-premises Pika instance is used to reduce the storage costs of an open source Redis deployment. The total size of the data is about 400 GB, and only about 10% of the data is frequently accessed. In addition, high costs are incurred for cluster O&M.

    In this case, you can use a hybrid-storage instance that has 64 GB memory and 512 GB disk capacity. This reduces your O&M costs and ensures high availability.

Select instances based on scenarios

Hybrid-storage instances of Tair and memory-only instances of ApsaraDB for Redis Community Edition are applicable to different scenarios. The following table describes the scenarios to which the two types of instances are applicable.
Item Hybrid-storage instance Community Edition instance
Data size and budget
  • You have large amounts of data to store and want to reduce the storage costs.
  • You are using the Pika, SSDB, or Antibiotic Resistance Genes Database (ARDB) services.
  • You are using master-replica or cluster instances of ApsaraDB for Redis to store large amounts of data.
You have a small amount of data, or your budget is sufficient for storage costs.
Separation of hot and cold data Your business data can be divided into hot data and cold data.
Note Cold data is the data that is not frequently accessed.
Your business data is accessed at random.
QPS performance and latency requirements for hot and cold data
  • High QPS performance is required for hot data.
  • Access to cold data is not sensitive to latency.
  • High QPS performance is required for all data.
  • Access to all data is sensitive to latency.
Access frequency of large keys
  • No large key exists in your business data.
  • Large keys are frequently accessed and need to be stored in memory.
  • Large keys are not frequently accessed and are not sensitive to latency.
Large keys are accessed at random and are sensitive to latency.

Select instances based on specifications

When you create a hybrid-storage instance of Tair, you must specify the memory capacity and the disk capacity. The memory capacity determines the maximum amount of hot data that can be stored, and the disk capacity determines the total amount of data that can be stored. Tair creates an instance with the corresponding CPU resources based on the instance specifications that you choose. For example, if you select a master-replica instance that has 64 GB of memory and 256 GB of disk capacity, the instance can store up to 256 GB of data and can cache 64 GB of data in memory.

For more information, see Overview.

  • You cannot specify the number of CPU cores.
  • Hybrid-storage instances store the metadata of keys in memory and disks. The metadata includes the time-to-live (TTL) value, the least recently used (LRU) clock value, and the type of each key. When you specify the memory capacity and the disk capacity, you must reserve storage capacity for metadata. For more information, see the following description in this topic.
  • Memory capacity

    To support more open source Redis features, hybrid-storage instances store all keys in memory. These instances store the values of frequently accessed keys in memory and the values of all keys in disks. Therefore, make sure that the specified memory capacity is sufficient to store all keys and related metadata. The following table lists the recommended memory capacity based on the total number of keys.

    Total number of keys Recommended memory capacity
    Less than 20 million 64 GB, 32 GB, and 16 GB
    20 million to 50 million 64 GB and 32 GB
    50 million to 100 million 128 GB, 64 GB, and 32 GB
    More than 100 million 128 GB and 64 GB
    Note The memory capacity that you specify when you create a Tair instance determines the CPU resources that are allocated to the instance. Instances that have a higher memory capacity provide higher performance.
  • Disk capacity

    Hybrid-storage instances store all data in disks, including the metadata generated for each key. The metadata occupies extra storage. Therefore, we recommend that you select a disk capacity that is 20% to 50% more than the required storage.

Instance performance

The performance of a hybrid-storage instance depends on the instance type and memory hit ratio. The hit ratio indicates the probability that the requested data is found in memory. Higher specifications and higher hit ratios indicate higher performance. If all requested data is found in the memory of the hybrid-storage instance, the performance is the same as that of a Community Edition instance. A lower memory hit ratio indicates the lower performance of the hybrid-storage instance. If all requested data is found in disks, the hybrid-storage instance delivers the lowest performance.

In the following performance tests, each Tair instance stores 20 million keys and the size of each value is 1 KB. The GET command is used to read values. Different types of keys are accessed in the following scenarios.

Table 1. Performance test results
Testing scenario QPS of a Community Edition instance QPS of a hybrid-storage instance
Keys are accessed at random. 123,000 15,000
Keys are accessed based on the Gaussian distribution, and 20% of the keys are accessed at a probability of 80%. 120,000 54,000
Keys are accessed based on the Gaussian distribution, and 1% of the keys are accessed at a probability of 99%. 135,000 114,000