[LUCENE-10427] OLAP likewise rollup during segment merge process

• Type: New Feature
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Lucene Fields: New
• Assignee: Unassigned
• Reporter: Suhan Mao
• Created: 18/Feb/22 15:26
• Updated: 28/Aug/22 16:36

Description

Currently, many OLAP engines support rollup feature like clickhouse(AggregateMergeTree)/druid. 

Rollup definition: https://athena.ecs.csus.edu/~mei/olap/OLAPoperations.php

One of the way to do rollup is to merge the same dimension buckets into one and do sum()/min()/max() operation on metric fields during segment compact/merge process. This can significantly reduce the size of the data and speed up the query a lot.

 

Abstraction of how to do

1. Define rollup logic: which is dimensions and metrics.
2. Rollup definition for each metric field: max/min/sum ...
3. index sorting should the the same as dimension fields.
4. We will do rollup calculation during segment merge just like other OLAP engine do.

 

Assume the scenario

We use ES to ingest realtime raw temperature data every minutes of each sensor device along with many dimension information. User may want to query the data like "what is the max temperature of some device within some/latest hour" or "what is the max temperature of some city within some/latest hour"

In that way, we can define such fields and rollup definition:

1. event_hour(round to hour granularity)
2. device_id(dimension)
3. city_id(dimension)
4. temperature(metrics, max/min rollup logic)

The raw data will periodically be rolled up to the hour granularity during segment merge process, which should save 60x storage ideally in the end.

 

How we do rollup in segment merge

bucket: docs should belong to the same bucket if the dimension values are all the same.

1. For docvalues merge, we send the normal mappedDocId if we encounter a new bucket in DocIDMerger.
2. Since the index sorting fields are the same with dimension fields. if we encounter more docs in the same bucket, We emit special mappedDocId from DocIDMerger .
3. In DocValuesConsumer.mergeNumericField, if we meet special mappedDocId, we do a rollup calculation on metric fields and fold the result value to the first doc in the  bucket. The calculation just like a streaming merge sort rollup.
4. We discard all the special mappedDocId docs because the metrics is already folded to the first doc of in the bucket.
5. In BKD/posting structure, we discard all the special mappedDocId docs and only place the first doc id within a bucket in the BKD/posting data. It should be simple.

 

How to define the logic

 

public class RollupMergeConfig {
  private List<String> dimensionNames;
  private List<RollupMergeAggregateField> aggregateFields;
} 

public class RollupMergeAggregateField {
  private String name;
  private RollupMergeAggregateType aggregateType;
}

public enum RollupMergeAggregateType {
  COUNT,
  SUM,
  MIN,
  MAX,
  CARDINALITY // if data sketch is stored in binary doc values, we can do a union logic 
}

 

 

I have written the initial code in a basic level. I can submit the complete PR if you think this feature is good to try.