数据库摘要sybg分析中物化策略μ验教训计算机科学.pdfVIP

Materialization Strategies in the Vertica ytic Database?Lessons Learned ZhuJie SCST,Soochow University : 20134227002@suda.edu.cn ?Column-store databases allow for various tuple reconstruction strategies (also called mate- rialization strategies).Early materialization (EM) is easy to implement while late materialization(LM) is hard.But LM usually performs much better than EM,except that situation revolving around join- s,such as neither input fits in memory (also called spilling joins).Sideways information passing(SIP) techniques can help you get the best of both worlds. EM combined with SIP allows us to get the benefits of LM,without the bookkee complexity or worse performance for spilling joins.This also provides some other benefits due to positive in ction with compression and sort orders of the data. Keywords?Column-store database?Materialization Strategies?EM ?LM ?SIP 1 Introduction All six authors are from HP,Vertica Group.Vertica is a column-store database system developped by HP. olumn-store databases,different columns are stored separa y on disk.And hence to produce results for any query, the necessary columns need to be stitched together during query execution.This process is referred to as tuple reconstruction or mate-rialization. Abadi et al. has raised two strategies ?early materialization(EM) and late materialization(LM).EM refers to stitching columns into partial tuples as early as possible,i.e., during column scan in query processing.LM refers to fetching columns on demand for each operator in the query n.For example, in a query with multiple joins, while evaluating any particular join, only the columns ne