RTA for Ad Placement

Scenarios

Advertisement (ad) placement is indispensable for enterprise promotion and marketing, especially for new media facing fierce competition. Now there are diversified advertising channels and detailed tailored ad placement.

Customer's increased awareness of ad ROI demands precise audience selection. For example, on a short video platform, advertisers have to configure rules on placing ads such as age, gender, and education background of audiences. The inflexibility hinders advertisers from making decisions on ad placement, and hundreds of millions or even billions of advertising fees need to be paid every year. However, it is still difficult to accurately reach target audiences. To let advertisers have autonomy to deliver or reject each ad request, Real-Time API (RTA) rose to the challenge.

RTA is used to meet the real-time personalized delivery needs of advertisers.

Challenges

Advertisers' RTA system read data from core profile databases, helping advertisers make placement decisions. The newer the data, the better the placement effect. Therefore, the latest data generated by a big data platform needs to be written into profile databases in a timely manner. To address high concurrency, ultra-low latency, and ultra-large volumes of RTA requests, core profile databases must have the following features:

• Quick import of vast amount of data; accurate decision-making

  Hundreds of GB or even several TB of all profile data needs to be periodically imported to profile databases. The faster the data is imported, the more accurate the model is, and the better ad placement effect is.

• High-concurrency access

  The RTA system needs to handle a large number of real-time bidding requests. For example, hundreds of thousands to millions of QPS are sent by RTA systems of e-commerce and financial customers.

• Stable low latency

  The media asks advertisers to provide decisions within 40 ms to 100 ms. Databases need to execute requests within single-digit milliseconds.

• Low cost

  To achieve ultimate performance, open-source self-hosted Redis needs to be installed for handling RTA requests. However, expensive TB-level data storage devices give advertisers a dilemma in selecting suitable models.

In an RTA system, common profile databases have the following problems:

• MySQL: It is difficult to handle hundreds of thousands to millions of concurrent QPS at low latency.
• MongoDB/HBase: It is inexpensive to store TB-level data, but MongoDB and HBase instances cannot maintain a stable low latency. High timeout rate may cause project suspension, injurious to commercial interests.
• In-memory database: For example, open-source self-hosted Redis is widely used in the industry. In-memory databases provide ultimate performance with high concurrency and low latency. However, there are risks such as poor stability and data loss. It costs too much and takes long to import TB-level user profile data.

GeminiDB Redis API features stability, low latency, and cost-effectiveness, and data can be imported extremely fast using FastLoad.

Solution Overview

In an RTA system, storage of GeminiDB Redis instances costs less than that of open-source self-hosted Redis instances. FastLoad enables quick offline data import. GeminiDB Redis API ensures stability and low latency and has rich practice cases of online advertising and recommendation services.

Advantages

• 5 to 10 times faster data import by FastLoad

  Traditional databases can only write data one by one using standard protocols. Data is crunched by the compute layer through a complex process and then written to the storage layer. Therefore, it usually takes hours or days for a big data platform to periodically import hundreds of GB or even several TB of profile data, which has a great impact on online services.

  FastLoad, an enterprise-level feature of GeminiDB Redis API, improves the data import speed by 5 to 10 times and reduces the impact on online services. To address RTA requests, the big data platform handles highly-concurrent workloads, FastLoad directly transfers massive volumes of data into a storage engine of the GeminiDB Redis instance via a dedicated high-speed persistent channel, and the storage engine on the GeminiDB Redis database orchestrates the data.

• Millions of concurrent requests and latency in sub-milliseconds

  GeminiDB Redis API uses a separated storage and compute architecture. Three copies of data are stored in a distributed shared storage pool. All nodes support efficient reads and writes, and compute power can be scaled up and out, making it easy to cope with workload spikes.

  With a multi-thread architecture, high-performance storage pool, and in-depth optimization of the memory data structure and access algorithm, GeminiDB Redis API can respond to requests in sub-milliseconds. Such an ultra-low latency is critical to real-time data processing and analysis, especially in scenarios such as online games, financial technologies, advertising systems, and real-time recommendation systems. Therefore, GeminiDB Redis API is ideal for large-scale real-time interaction and high-frequency transactions.

  With vast experience on the live network, GeminiDB Redis API can handle more than one million QPS while ensuring average latency of 1 ms and p99 latency of 2 ms.

• Efficient data compression and storage at low costs

  By compressing both logical and block data, GeminiDB Redis API greatly reduces storage resource consumption while maintaining ultimate performance. Compute and storage resources are decoupled from each other, so they can be flexibly scaled out. Achieving optimized resource utilization means a range of benefits for enterprises, most notably saving costs in storing data.

  With vast experience on the live network, GeminiDB Redis API has a compression ratio of 4:1. That is, only about 3 TB out of 12 TB of data is occupied on a GeminiDB Redis instance.