ݺߣ

ݺߣShare a Scribd company logo

jDMN - DecisionCamp 2024.pptx presentación presentación

•Download as PPTX, PDF•
•
Diego Martín
Diego Martín

Jdmn

1 of 16
Download to read offline
Optimizing Performance and Scalability in Low-Code / No-Code DMN Platforms September 2024
Content 2  Problem Statement  Problem Solution  Overall structure of jDMN  Transpiler to Java / Kotlin / Python  Code optimisation  Q&As
Problem Statement 3  Challenging BDM environment  High risk decisions  Fast-release cycles  High Tech Risk  High number of executions  Quality  functionality, reliability, usability, flexibility, efficiency, maintainability  LCNCP platform  Attributes in scope  Efficiency  performance  scalability
jDMN: Overall Structure 4  DMN Processors  Reader / Writer  Validators  Transformers  Interpreter  Translator  Dialects  DMN 1.1 – 1.5  Signavio
jDMN: Overall Structure 5 Semantic Analyzer Semantic Tree AST Lexical Analyzer Source Text Parser Source Tokens Syntax Analyzer Code Generator Interpreter Error Manager & Logger
Translation 6
jDMN: Translation 7
jDMN: Translation 8  How did we built it?  Syntax-Driven Translation Schematas (SDTS)  Based on Knuth’s attributed grammars  Synthesized attributes  Inherited Attributes Expr1 → Expr2 + Term { Expr1.value = Expr2.value + Term.value } Expr → Term { Expr.value = Term.value } Term1 → Term2 * Factor { Term1.value = Term2.value * Factor.value } Term → Factor { Term.value = Factor.value } Factor → "(" Expr ") { Factor.value = Expr.value } Factor → integer { Factor.value = strToInt(integer.str) }
jDMN: Translation 9  Advantages include  Performance: we have seen runtimes 4-10 times faster than previous engine execution in complex decision tests  Stability: given that we now control the code generation, we are able to resolve issues without relying on the vendor  Functionality: the fact that we control the code generation means that we are also able to enable more advanced functionality for DMN/Java models
Code Optimisation – Model Level 10  AWS Lambda  gRPC / protobuf Users Message Bus AWS Stack
Code Optimisation – DRG Element Level 11  Caching  Map-reduce for  MID
Code Optimisation – DRG Element Level 12  Lazy evaluation  Inner nodes (Decisions, BKMs, Decision Services)  Leaves (Input Data)
Code Optimisation – DRG Element Level 13 Performance Before DRG Optimization Decision Name Request Count Response Time 90th (ms) Min Response Time (ms) Max Response Time (ms) D1 12899 40 4 5562 D2 361 6676 21 9653 D3 28731 15 1 933 D4 86165 39 4 6367 Performance After DRG Optimization Decision Name Request Count Response Time 90th (ms) Min Response Time (ms) Max Response Time (ms) D1 12899 35 5 7150 D2 361 315 153 4603 D3 28745 12 1 606 D4 86204 37 4 7736
Code Optimisation at FEEL level 14  Native Types  Built-in functions  Native Compiler / Interpreter (e.g. JIT compiler)
What next? 15  Enhance lazy evaluation  Three Address Code optimization
Questions? 16 https://github.com/goldmansachs/jdmn

More Related Content

jDMN - DecisionCamp 2024.pptx presentación presentación

  • 1. Optimizing Performance and Scalability in Low-Code / No-Code DMN Platforms September 2024
  • 2. Content 2  Problem Statement  Problem Solution  Overall structure of jDMN  Transpiler to Java / Kotlin / Python  Code optimisation  Q&As
  • 3. Problem Statement 3  Challenging BDM environment  High risk decisions  Fast-release cycles  High Tech Risk  High number of executions  Quality  functionality, reliability, usability, flexibility, efficiency, maintainability  LCNCP platform  Attributes in scope  Efficiency  performance  scalability
  • 4. jDMN: Overall Structure 4  DMN Processors  Reader / Writer  Validators  Transformers  Interpreter  Translator  Dialects  DMN 1.1 – 1.5  Signavio
  • 5. jDMN: Overall Structure 5 Semantic Analyzer Semantic Tree AST Lexical Analyzer Source Text Parser Source Tokens Syntax Analyzer Code Generator Interpreter Error Manager & Logger
  • 8. jDMN: Translation 8  How did we built it?  Syntax-Driven Translation Schematas (SDTS)  Based on Knuth’s attributed grammars  Synthesized attributes  Inherited Attributes Expr1 → Expr2 + Term { Expr1.value = Expr2.value + Term.value } Expr → Term { Expr.value = Term.value } Term1 → Term2 * Factor { Term1.value = Term2.value * Factor.value } Term → Factor { Term.value = Factor.value } Factor → "(" Expr ") { Factor.value = Expr.value } Factor → integer { Factor.value = strToInt(integer.str) }
  • 9. jDMN: Translation 9  Advantages include  Performance: we have seen runtimes 4-10 times faster than previous engine execution in complex decision tests  Stability: given that we now control the code generation, we are able to resolve issues without relying on the vendor  Functionality: the fact that we control the code generation means that we are also able to enable more advanced functionality for DMN/Java models
  • 10. Code Optimisation – Model Level 10  AWS Lambda  gRPC / protobuf Users Message Bus AWS Stack
  • 11. Code Optimisation – DRG Element Level 11  Caching  Map-reduce for  MID
  • 12. Code Optimisation – DRG Element Level 12  Lazy evaluation  Inner nodes (Decisions, BKMs, Decision Services)  Leaves (Input Data)
  • 13. Code Optimisation – DRG Element Level 13 Performance Before DRG Optimization Decision Name Request Count Response Time 90th (ms) Min Response Time (ms) Max Response Time (ms) D1 12899 40 4 5562 D2 361 6676 21 9653 D3 28731 15 1 933 D4 86165 39 4 6367 Performance After DRG Optimization Decision Name Request Count Response Time 90th (ms) Min Response Time (ms) Max Response Time (ms) D1 12899 35 5 7150 D2 361 315 153 4603 D3 28745 12 1 606 D4 86204 37 4 7736
  • 14. Code Optimisation at FEEL level 14  Native Types  Built-in functions  Native Compiler / Interpreter (e.g. JIT compiler)
  • 15. What next? 15  Enhance lazy evaluation  Three Address Code optimization