Speed Up Query Execution Using Vectorization

Normal vs Vectorized Instructions

Vectorized Queries

Vectorized execution operates on batches of rows at a time instead of individual rows.

It allows queries to be sped up by reducing the number of method calls, improving cache efficiency, and potentially enabling CPU SIMD instructions.

As of Druid 0.20.2, the following query types can be vectorized:

Vectorization has a few important requirements:

Query-level filters must either run on bitmap indexes or offer vectorized row matchers.
Supported filters include selector, bound, in, like, regex, search, and, or, and not.
Filters in filtered aggregators must offer vectorized row matchers.
Aggregators must offer vectorized implementations, including count, numeric sum/min/max, any-value aggregators, hyperUnique, filtered, approxHistogram, approxHistogramFold, and fixedBucketsHistogram.
Virtual columns must offer vectorized implementations.
GroupBy dimension specs must be default specs, without extraction functions, filtered dimension specs, or multi-value dimensions.
Timeseries queries cannot use descending order.
Only immutable segments and table datasources are supported.

As of Druid 0.20.2, these query types cannot be vectorized:

Property	Default	Description
`vectorize`	`true`	Enables or disables vectorized query execution.
`vectorSize`	`512`	Sets the row batching size for a query and overrides `druid.query.default.context.vectorSize` when set.
`vectorizeVirtualColumns`	`false`	Controls vectorized processing for virtual columns. Values are `false`, `true`, and `force`.

Set vectorization defaults in common config:

druid.query.default.context.vectorize=true
druid.query.default.context.vectorSize=1024
druid.query.default.context.vectorizeVirtualColumns=true

Or pass vectorization parameters in query context:

{
  "context": {
    "vectorize": true,
    "vectorSize": 512,
    "vectorizeVirtualColumns": false
  }
}