Viktor Holmqvist Jonathan Nilsfors Philipp Leitner Cachematic – Automatic Invalidation in Application-Level Caching Systems Industry Track Presentation 

Chalmers !2 Company Partner - Bison https://bisonanalytics.com 

Chalmers !3 Application-Level Caching https://medium.com/@MatthewFTech/spring-boot-cache-with-redis-56026f7da83a 

Chalmers !4 Cache Invalidation When to purge / update cached data? 

Chalmers !5 Cache Invalidation When to purge / update cached data? • Time-Based • Explicit 

Chalmers !6 Cache Invalidation When to purge / update cached data? • Time-Based • Explicit (company requirement) 

Chalmers !7 Current State Manual Cache Management 

Chalmers !8 Current State Manual Cache Management Consistent with state of industry, see for example Mertz and Nunes (TSE 2017) https://ieeexplore.ieee.org/stamp/stamp.jsp? tp=&arnumber=7762909 

Chalmers !9 Challenges of Manual Cache Management • Easy to miss cacheable queries • Development / maintenance complexity • Potential for bugs 

Chalmers !10 Goal Develop a library for application-level caching • Automated • Reusable • Minimal developer input “Cachematic” 

Chalmers !11 Cachematic 

Chalmers !12 Cachematic 

Chalmers !13 Application Code …. End Users Statement Parser Dependency Tree Hash Function Read Request Processor Write Request Processor SQL Database Key-Value Store Cachematic Code Annotations Serialization Function Cachematic Architecture 

Chalmers !14 Cache Management Algorithm parse DB statement process based on statement type (query or write) 

Chalmers !15 If Query: if result in cache then: return cached val else: query DB save to val to cache Cache Management Algorithm parse DB statement process based on statement type (query or write) 

Chalmers !16 If Query: if result in cache then: return cached val else: query DB save to val to cache Cache Management Algorithm If Write: decide which tables / columns / rows are touched foreach cached result: check for invalidation (hierarchically) write update to DB parse DB statement process based on statement type (query or write) 

Chalmers !17 Implementation Application Code …. End Users Statement Parser Dependency Tree Hash Function Read Request Processor Write Request Processor SQL Database Key-Value Store Cachematic Code Annotations Serialization Function pyparsing Open Source release planned 

Chalmers !18 Experimental Evaluation 

Chalmers !19 Experiment Goal • Evaluate Cachematic based on Bison’s production application 

Chalmers !20 Experiment Goal • Evaluate Cachematic based on Bison’s production application Main RQ: • Can we improve cache hit rate / performance for read requests? 

Chalmers !21 Experiment Setup Load Testing Runs 6 user profiles AWS using Locust (https://locust.io) 30-minute test runs 

Chalmers !22 Read Case (median) 

Chalmers !23 Read Case Cache Hit Rate 69% (manual: 42%) 

Chalmers !24 Read Case (90th percentile) 

Chalmers !25 Lessons Learned • Can we improve cache hit rate / performance for read requests? Yes (by about a factor of 1.6 in our test setup) 

Chalmers !26 Lessons Learned • Can we improve cache hit rate / performance for read requests? Yes (by about a factor of 1.6 in our test setup) • But: significant overhead in write requests To some extent engineering limitation Needs to be improved for OSS release 

Chalmers !27 Conclusions Cachematic: Automatic application cache management for Python Improves cache hit rate and read performance (w.r.t. manual cache management solution) However: negative impact on write requests Although acceptable for Bison, Cachematic is not (yet) a general solution for all applications 

Chalmers !28 I Am Hiring! Talk to me about opportunities in Sweden! (PhD / postdoc)