Generating SAA Projects from Dify Workflows

Operation Instructions

For export operation instructions, please refer to the spring-ai-alibaba-graph-studio module to quickly generate a Spring AI Alibaba project.

After starting Graph Studio, you can currently generate a project by calling the following HTTP request, which will automatically generate a zip package of the Spring AI Alibaba project.

curl --location --request POST 'http://localhost:8080/starter.zip' \
--header 'Content-Type: application/x-www-form-urlencoded' \
--data-urlencode 'dependencies=spring-ai-alibaba-graph,web' \
--data-urlencode 'appMode=workflow' \
--data-urlencode 'type=maven-project' \
--data-urlencode 'language=java' \
--data-urlencode 'bootVersion=3.5.0' \
--data-urlencode 'baseDir=demo' \
--data-urlencode 'groupId=com.example' \
--data-urlencode 'artifactId=demo' \
--data-urlencode 'name=demo' \
--data-urlencode 'description=Spring AI Alibaba Project Exported from Dify.' \
--data-urlencode 'packageName=com.example.demo' \
--data-urlencode 'packaging=jar' \
--data-urlencode 'javaVersion=17' \
--data-urlencode '{put-your-dify-dsl-here}'

Please note that this supporting functionality is still under continuous development. Please pay attention to documentation updates, as we will add more detailed explanations.

Performance Test Data

Test Cluster Specifications

1. Spring AI Alibaba project, independently deployed container, maintaining default thread pool and other configuration parameters, 2 PODs, POD specification 2C4G
2. Dify platform, official deployment method, maintaining default configuration parameters, with 2 PODs for each component, POD specification 2C4G

Effective Concurrent Processing Limit

• Test Method: For each scenario, starting from 10 RPS (Requests Per Second) and gradually increasing until increasing the RPS value no longer improves TPS and the success rate decreases.
• Conclusion: Dify’s maximum processing limit is < 10 RPS; Spring AI Alibaba’s maximum processing limit is approximately 150 RPS.

Dify performance test screenshot:

Spring AI Alibaba performance test screenshot:

Throughput in Extreme Scenarios

• Test Method: Sending test request volumes far exceeding reasonable concurrency to the cluster (test scenario: 1000 RPS), observing changes in the cluster’s throughput and success rate.
• Conclusion: In this scenario, Dify’s success rate is less than 10%, with an average RT approaching 60s, and most requests timing out (response greater than 60s); Spring AI Alibaba’s success rate remains largely unchanged, maintaining over 99%, with an average RT around 18s.

Dify performance test screenshot:

Spring AI Alibaba performance test screenshot:

Optimization Potential

• For Dify, subsequent optimization could involve adjusting related component configurations to see if there is performance improvement.
• For SAA, it is expected that simple adjustments to the thread pool strategy could significantly increase current concurrency and throughput.