Report on VLDB 2017 Conference: Insights and Highlights from Database Research

Preface – From August 28 to September 1, 2017, VLDB 2017 was held at the Technical University of Munich, one of the three top conferences in the database field. Alibaba was a gold sponsor, and several Ant Financial colleagues, including members from OceanBase and GeaBase, attended. This article is a learning summary of their experience.

Overall Impression – The week in Munich was packed with sessions and discussions. Besides classic relational database topics such as optimizers, engines, distributed execution, and transaction concurrency control, the program also covered big‑data processing, graph, spatial, text, semi‑structured, streaming data, data mining, crowdsourcing, social‑network analysis, and visualization. Two hot research directions emerged: (1) prototype database systems that exploit new hardware (NVM, flash, GPU, FPGA) and (2) applying traditional relational techniques to big‑data platforms like Spark.

Industry reports were abundant, especially from traditional database vendors. Oracle delivered several workshops, including a talk on product‑failure decisions and another on adaptive statistics. SAP HANA presented its NVM‑based storage practice, highlighting LLVM‑compiled query plans.

Chinese researchers played a strong role, winning several awards (10‑year best paper, young scholar) and contributing many talks and discussions.

Topic Sharing – The author highlights several areas of personal interest:

FADS – “Failed Aspirations in Database Systems” collects valuable lessons from failed attempts in the field.

Cache Evolution – Oracle’s cache solutions from the 8i i‑cache to Timesten and the recent Velocity Scale approach are described, emphasizing cost, transparency, and operational complexity.

Other Topics – Includes Oracle’s acquisition of Berkeley DB, Microsoft’s timestamp‑based concurrency control, and Stream Insight.

Keynote Reports

1. New hardware impact on database design – presented by Wolfgang, covering heterogeneous CPUs, GPUs, FPGAs, and the “Dark Silicon” effect.

2. Spark development history – presented by Michael Franklin, tracing the evolution of big‑data processing platforms and their relationship to relational databases.

Key observations from the second keynote: the database field has undergone massive change, big‑data software is disruptive, database thinking is crucial for adding value, and outdated mindsets must be discarded.

Classic Database Topics

Statistical information and optimizer research, distributed transactions, concurrency control, and distributed query processing are still active. Notable papers include:

Distributed Join Algorithms on Thousands of Cores – MPI‑based radix‑hash and sort‑merge joins on a 4096‑core machine.

HippogriffDB – Balancing I/O and GPU bandwidth for big‑data analytics.

Adaptive Work Placement for Heterogeneous Resources – Self‑adapting operator placement across CPU, GPU, FPGA.

SAP HANA Adoption of Non‑Volatile Memory – Using NVRAM for both memory and storage extensions.

BlueCache – Scalable flash‑based key‑value store.

Caribou – Intelligent distributed storage leveraging new hardware.

Memory Management Techniques for Large‑Scale Persistent‑Main‑Memory Systems – NVM‑aware allocator design.

Transaction and concurrency control papers highlighted include:

The End of a Myth: Distributed Transaction Can Scale.

An Empirical Evaluation of In‑Memory Multi‑Version Concurrency Control.

High Performance Transactions via Early Write Visibility.

Write‑Behind Logging using NVM characteristics.

Partitioning and parallelism papers such as AdaptDB, Clay, and SquirrelJoin were also listed.

Postscript – The author reflects on the vibrant German innovation in databases (SAP HANA, Exa, HyPer) and contrasts Munich’s slower urban pace with the rapid development of internet services in China.