[jira] [Commented] (HIVE-26221) Add histogram-based column statistics

["Alessandro Solimando (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-26221?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17646588#comment-17646588 ] 

Alessandro Solimando commented on HIVE-26221:
---------------------------------------------

Thank you [~dengzh] and [~amansinha100] for the great review! I would also like to acknowledge the work of Ryan Johnson for all the benchmarking he did which shaped the final decision of directly using CDF function without needing intermediate (binned) histogram representation, and [~kgyrtkirk] and [~amansinha100] for their inputs in the design phase of the proposal.

> Add histogram-based column statistics
> -------------------------------------
>
>                 Key: HIVE-26221
>                 URL: https://issues.apache.org/jira/browse/HIVE-26221
>             Project: Hive
>          Issue Type: Improvement
>          Components: CBO, Metastore, Statistics
>    Affects Versions: 4.0.0-alpha-2
>            Reporter: Alessandro Solimando
>            Assignee: Alessandro Solimando
>            Priority: Major
>              Labels: pull-request-available
>             Fix For: 4.0.0
>
>          Time Spent: 11.5h
>  Remaining Estimate: 0h
>
> Hive does not support histogram statistics, which are particularly useful for skewed data (which is very common in practice) and range predicates.
> Hive's current selectivity estimation for range predicates is based on a hard-coded value of 1/3 (see [FilterSelectivityEstimator.java#L138-L144|https://github.com/apache/hive/blob/56c336268ea8c281d23c22d89271af37cb7e2572/ql/src/java/org/apache/hadoop/hive/ql/optimizer/calcite/stats/FilterSelectivityEstimator.java#L138-L144]).])
> The current proposal aims at integrating histogram as an additional column statistics, stored into the Hive metastore at the table (or partition) level.
> The main requirements for histogram integration are the following:
>  * efficiency: the approach must scale and support billions of rows
>  * merge-ability: partition-level histograms have to be merged to form table-level histograms
>  * explicit and configurable trade-off between memory footprint and accuracy
> Hive already integrates [KLL data sketches|https://datasketches.apache.org/docs/KLL/KLLSketch.html] UDAF. Datasketches are small, stateful programs that process massive data-streams and can provide approximate answers, with mathematical guarantees, to computationally difficult queries orders-of-magnitude faster than traditional, exact methods.
> We propose to use KLL, and more specifically the cumulative distribution function (CDF), as the underlying data structure for our histogram statistics.
> The current proposal targets numeric data types (float, integer and numeric families) and temporal data types (date and timestamp).



--
This message was sent by Atlassian Jira
(v8.20.10#820010)