This page is about Data Skipping capability, which has been available in beta since Databricks Runtime 3.0 and is deprecated. An enhanced version of Data Skipping is available as part of Delta Lake on Databricks. We recommend that you switch to using Delta Lake on Databricks to continue taking advantage of this feature. See Delta Lake on Databricks Data skipping for details.
In addition to Partition pruning, Databricks Runtime includes another feature that is meant to avoid scanning irrelevant data, namely the Data Skipping Index. It uses file-level statistics in order to perform additional skipping at file granularity. This works with, but does not depend on, Hive-style partitioning.
The effectiveness of data skipping depends on the characteristics of your data and its physical layout. As skipping is done at file granularity, it is important that your data is horizontally partitioned across multiple files. This will typically happen as a consequence of having multiple append jobs, (shuffle) partitioning, bucketing, and/or the use of
spark.sql.files.maxRecordsPerFile. It works best on tables with sorted buckets (
CREATE TABLE ... CLUSTERED BY ... SORTED BY ...), or with columns that are correlated with partition keys (for example,
brandName - modelName,
companyID - stockPrice), but also when your data just happens to exhibit some sortedness / clusteredness (for example,
This beta feature has a number of important limitations:
- It’s Opt-In: needs to be enabled manually, on a per-table basis.
- It’s SQL only: there is no DataFrame API for it.
- Once a table is indexed, the effects of subsequent
ADD PARTITIONoperations are not guaranteed to be visible until the index is explicitly REFRESHed.
CREATE DATASKIPPING INDEX ON [TABLE] [db_name.]table_name
Enables Data Skipping on the given table for the first (i.e. left-most) N supported columns, where N is controlled by
spark.databricks.io.skipping.defaultNumIndexedCols (default: 32)
partitionBy columns are always indexed and do not count towards this N.
CREATE DATASKIPPING INDEX ON [TABLE] [db_name.]table_name FOR COLUMNS (col1, ...)
Enables Data Skipping on the given table for the specified list of columns. Same as above, all
partitionBy columns will always be indexed in addition to the ones specified.
DESCRIBE DATASKIPPING INDEX [EXTENDED] ON [TABLE] [db_name.]table_name
Displays which columns of the given table are indexed, along with the corresponding types of file-level statistic that are collected.
EXTENDED is specified, a third column called “effectiveness_score” is displayed that gives an approximate measure of how beneficial we expect DataSkipping to be for filters on the corresponding columns.
REFRESH DATASKIPPING INDEX ON [TABLE] [db_name.]table_name
Rebuilds the whole index. I.e. all the table’s partitions will be re-indexed.
REFRESH DATASKIPPING INDEX ON [TABLE] [db_name.]table_name PARTITION (part_col_name1[=val1], part_col_name2[=val2], ...)
Re-indexes the specified partitions only. This operation should generally be faster than full index refresh.