Introduction

• Skew can cause a lot of problems
  
• Sufficient skew in the underlying data causes
  load imbalances in the resulting parallel join
  swamping the gains due to parallelism
• Earlier Skew handling algos perform worse
  than simple algos in the absence of skew
• There is no evidence that extreme degrees of
  skew occur commonly in practice
• Requirement to not penalize normal cases
  much and also handle extreme cases reasonably
• Paper presents several algos to do this
• Sampling is the main new ingredient