Dremio maintains physically optimized representations of source data known as Data Reflections. The query optimizer can accelerate a query by utilizing one or more Data Reflections to partially or entirely satisfy that query, rather than processing the raw data in the underlying data source.
Dremio supports two fundamental types of Data Reflections: Raw Reflections and Aggregation Reflections. Many of the options for configuring and managing both types of Data Reflections are the same, but they each optimize different types of query patterns.
Raw Reflections
Raw Reflections preserve row-level fidelity of the anchor dataset. A Raw Reflection includes one or more fields from the anchor dataset, and is sorted and partitioned by specific fields in the dataset retrieved from the database on Dremio. You can use Raw Reflections to perform a number of optimizations.
Aggregation Reflections
Aggregation Reflections maintain summary data about the anchor dataset. An Aggregation Reflection includes one or more dimensions and measures fields from the anchor dataset, sorted, partitioned and distributed by specified columns. Some of these columns are configured as dimensions that will be used in GROUP BY (or DISTINCT) statements, and other columns are configured as measures that will be used in calculations such as MAX, MIN, AVG, SUM, and COUNT.
Query Acceleration
Dremio uses Data Reflection for query acceleration.
When Dremio receives a user query, it first determines whether any Data Reflections have at least one physical dataset in common with the query after both have undergone dataset expansion. All Data Reflections that pass this step are then evaluated to determine if they cover the query.
For Data Reflections that cover the query, Dremio will determine the cost of using the Data Reflection to execute the query. These costs are then compared to the cost of executing the query against the physical datasets, and the lowest cost query plan is selected for physical plan generation. Typically using one or more Data Reflections will be less expensive than executing the query against the raw physical data.
