Synthesizing Object State Transformers for Dynamic Software Updates Zelin Zhao*,Yanyan Jiang*,Chang Xu*,Tianxiao Gut,Xiaoxing Ma* *State Key Laboratory for Novel Software Technology and Department of Computer Science and Technology,Nanjing University,China TAlibaba Group,Sunnyvale,CA,USA Abstract-There is an increasing demand for evolving software 1)Almost all (187,or 98.4%)class changes can be updated systems to deliver continuous services of no restart.Dynamic dynamically,indicating that DSU is broadly applicable. software update (DSU)aims to achieve this goal by patching the Even for a few cases (3,or 1.6%)that DSU is impossible system state on the fly but is currently hindered from practice due to non-trivial cross-version object state transformations.This over existing programs,proper refactoring could still make paper revisits this problem through an in-depth empirical study them updatable [4,29.30]. of over 190 class changes from Tomcat 8.The study produced an 2)Most(166,or 87.4%)updates involve trivial object trans- important finding that most non-trivial object state transformers formations over simple predefined rules.Existing DSU can be constructed by reassembling existing old/new version code snippets.This paper presents a domain-specific language systems [15,17,18]are already capable of automatically and an efficient algorithm for synthesizing non-trivial object updating these changes without developers'intervention. transformers over code reuse.We experimentally evaluated our 3)The rest,not many but a non-negligible portion of class tool implementation PASTA with real-world software systems, changes (21,or 11.1%)require non-trivial object transfor- reporting PASTA's effectiveness in succeeding in 7.5x non-trivial mations.Software developers without a DSU background object transformation tasks compared with the best existing DSU would have substantial difficulties in specifying them,as techniques. Index Terms-Software maintenance and evolution,dynamic the required transformers have to carefully manipulate software update,object transformation,program synthesis. two versions of program states simultaneously. The empirical study results suggest that the key obstacle I.INTRODUCTION that hinders the continuous and automatic deployment of DSU in practice is probably how to obtain non-trivial object Dynamic software update (DSU,updating software at transformers.Unfortunately,this circumstance has not been runtime without restarting)[1]is a trending feature in modern seriously recognized by existing research.In fact,our later software systems.DSU keeps systems up-to-date with security experiments show that state-of-the-art techniques,like TOS [27] patches,bug fixes,and feature upgrades without hurting the and AOTES [28],could only succeed in 0 and 2 out of 25 non- systems'availability.DSU has become increasingly practical and compelling [1-8]:Linux Kernel [9-12]and Microsoft trivial object transformation tasks.Their apparent high success rates in past experiments might be due to mixing non-trivial Windows [13]are already dynamically updatable to some extent;Java Virtual Machine (JVM)has been modified to transformers with many trivial ones. In this paper,we leverage another key empirical finding that partially support application updates [14-18];live-upgradable object transformers can be constructed by reassembling existing components are also emerging in databases [19-21],servers [22- old/new version code to establish an algorithm for synthesizing 24],and even mission-critical systems [25]. object transformers in the DSU of Java applications.The Despite that code can be hot upgraded in emerging algorithm exhaustively and heuristically enumerates all possible systems [24,26],automatically updating runtime states for combinations of extracted code snippets,producing both test- seamless system evolution remains a major research chal- passing and developer-readable object transformers.A key lenge [15,27,28].Software updates may include a field of a advantage over existing techniques [27,28]is that an appli- class being added,removed,or semantically changed in a new cation developer can easily verify synthesized transformers' version.In DSU,such field values (if not removed)should be correctness because application code is their major constructs. (re)computed to match the new version code's semantics.This We implemented our algorithm as the PASTA (PATCH is known as the object transformation problem,whose solution STATES)tool for DSU of Java programs.The evaluation typically relies on a key mini-program (a.k.a.a transformer) results over a set of non-trivial class changes (including for computing these field values at update time. those in the empirical study and more)were encouraging: To understand the challenges in object transformation,this PASTA synthesized 7.5x correct non-trivial object transformers paper empirically studied 100 uniform-randomly sampled (60.0%)compared to the best existing techniques TOS [27] commits (consisting of 190 class changes)from Apache and AOTES [28](0.0%and 8.0%,respectively). Tomcat 8 [23],one of the most popular Web backend systems. In summary,this paper's major contributions are recognizing Our major findings include: the non-trivial object transformer synthesis as a critical problemSynthesizing Object State Transformers for Dynamic Software Updates Zelin Zhao∗ , Yanyan Jiang∗ , Chang Xu∗ , Tianxiao Gu† , Xiaoxing Ma∗ ∗State Key Laboratory for Novel Software Technology and Department of Computer Science and Technology, Nanjing University, China †Alibaba Group, Sunnyvale, CA, USA Abstract—There is an increasing demand for evolving software systems to deliver continuous services of no restart. Dynamic software update (DSU) aims to achieve this goal by patching the system state on the fly but is currently hindered from practice due to non-trivial cross-version object state transformations. This paper revisits this problem through an in-depth empirical study of over 190 class changes from Tomcat 8. The study produced an important finding that most non-trivial object state transformers can be constructed by reassembling existing old/new version code snippets. This paper presents a domain-specific language and an efficient algorithm for synthesizing non-trivial object transformers over code reuse. We experimentally evaluated our tool implementation PASTA with real-world software systems, reporting PASTA’s effectiveness in succeeding in 7.5× non-trivial object transformation tasks compared with the best existing DSU techniques. Index Terms—Software maintenance and evolution, dynamic software update, object transformation, program synthesis. I. INTRODUCTION Dynamic software update (DSU, updating software at runtime without restarting) [1] is a trending feature in modern software systems. DSU keeps systems up-to-date with security patches, bug fixes, and feature upgrades without hurting the systems’ availability. DSU has become increasingly practical and compelling [1–8]: Linux Kernel [9–12] and Microsoft Windows [13] are already dynamically updatable to some extent; Java Virtual Machine (JVM) has been modified to partially support application updates [14–18]; live-upgradable components are also emerging in databases [19–21], servers [22– 24], and even mission-critical systems [25]. Despite that code can be hot upgraded in emerging systems [24, 26], automatically updating runtime states for seamless system evolution remains a major research chal￾lenge [15, 27, 28]. Software updates may include a field of a class being added, removed, or semantically changed in a new version. In DSU, such field values (if not removed) should be (re)computed to match the new version code’s semantics. This is known as the object transformation problem, whose solution typically relies on a key mini-program (a.k.a. a transformer) for computing these field values at update time. To understand the challenges in object transformation, this paper empirically studied 100 uniform-randomly sampled commits (consisting of 190 class changes) from Apache Tomcat 8 [23], one of the most popular Web backend systems. Our major findings include: 1) Almost all (187, or 98.4%) class changes can be updated dynamically, indicating that DSU is broadly applicable. Even for a few cases (3, or 1.6%) that DSU is impossible over existing programs, proper refactoring could still make them updatable [4, 29, 30]. 2) Most (166, or 87.4%) updates involve trivial object trans￾formations over simple predefined rules. Existing DSU systems [15, 17, 18] are already capable of automatically updating these changes without developers’ intervention. 3) The rest, not many but a non-negligible portion of class changes (21, or 11.1%) require non-trivial object transfor￾mations. Software developers without a DSU background would have substantial difficulties in specifying them, as the required transformers have to carefully manipulate two versions of program states simultaneously. The empirical study results suggest that the key obstacle that hinders the continuous and automatic deployment of DSU in practice is probably how to obtain non-trivial object transformers. Unfortunately, this circumstance has not been seriously recognized by existing research. In fact, our later experiments show that state-of-the-art techniques, like TOS [27] and AOTES [28], could only succeed in 0 and 2 out of 25 non￾trivial object transformation tasks. Their apparent high success rates in past experiments might be due to mixing non-trivial transformers with many trivial ones. In this paper, we leverage another key empirical finding that object transformers can be constructed by reassembling existing old/new version code to establish an algorithm for synthesizing object transformers in the DSU of Java applications. The algorithm exhaustively and heuristically enumerates all possible combinations of extracted code snippets, producing both test￾passing and developer-readable object transformers. A key advantage over existing techniques [27, 28] is that an appli￾cation developer can easily verify synthesized transformers’ correctness because application code is their major constructs. We implemented our algorithm as the PASTA (PATCH STATES) tool for DSU of Java programs. The evaluation results over a set of non-trivial class changes (including those in the empirical study and more) were encouraging: PASTA synthesized 7.5× correct non-trivial object transformers (60.0%) compared to the best existing techniques TOS [27] and AOTES [28] (0.0% and 8.0%, respectively). In summary, this paper’s major contributions are recognizing the non-trivial object transformer synthesis as a critical problem