Stop Thinking, Just Do!

Sungsoo Kim's Blog

Main Compensation and Delta Compensation

tagsTags

15 January 2016


Main Compensation

While updates and deletes of records in the delta storage are handled transparently and do not affect our caching algorithm, an aggregate cache entry can become inconsistent with respect to a record invalidation in the main storage, including deletes and updates of the current transaction. Instead of recalculating an aggregate cache entry with every record invalidation in the main storage, we employ an approach that uses bit vector comparison to efficiently detect invalidated records and apply them to aggregate cache entries in a process called main compensation. As illustrated in Fig. 3, we use the consistent view manager to retrieve current record visibilities during aggregate cache entry usage.

The record invalidation is handled through the consistent view manager (see Fig. 1) that creates a bit vector representing the visibility of records of a table for an incoming query based on its transaction token. When an aggregate query is cached, the current snapshot is captured using this visibility vector. When a query is executed using an aggregate cache entry, an efficient bit vector comparison of the current snapshot with the snapshot at the cache creation time is used, thereby detect invalidated records, and apply them for main compensation. The details of aggregate cache main compensation can be found in [19] and are omitted in this paper for simplification reasons.

Delta Compensation

As the last step in query execution (Fig. 3), any query using the aggregate cache must apply delta compensation operation which accounts for records in the delta storage visible to the current transaction. When the aggregate cache is based on multiple tables joins, the complexity of answering a query using the aggregate cache increases as the aggregate cache is computed on the main partitions only, and the query must be compensated with all subjoins on deltas and mains. As a result, the profit of caching an aggregate query based on many tables may be very low because their performance using the aggregate cache is not superior to not using it. The techniques proposed in this paper have the main goal of extending the class of aggregate queries which qualify to be admitted into the SAP HANA aggregate cache.


comments powered by Disqus