apache / paimon

Apache Paimon is a lake format that enables building a Realtime Lakehouse Architecture with Flink and Spark for both streaming and batch operations.
https://paimon.apache.org/
Apache License 2.0
2.43k stars 954 forks source link

[Feature] Using bitmap index to accelerate the query #4530

Open Tan-JiaLiang opened 1 week ago

Tan-JiaLiang commented 1 week ago

Search before asking

Motivation

Currently we have introduced bitmap indexes. There are some optimisations that we can do when the user queries only against the bitmap indexed columns.

Suppose we have a table usershop_behavior with a bitmap index on the gender column and a bsi index on the gmv column.

CREATE TABLE usershop_behavior (
    uid BIGINT
    gender STRING,
    gmv BIGINT
) WITH (
    'file-index.bitmap.columns' = 'gender',
    'file-index.bsi.columns' = 'gmv'
);

The bitmap index and bsi index can be used not only for filtering, but also for some simple aggregation like:

SELECT 
    gender,
    COUNT(*) AS total,
    SUM(gmv) AS total_gmv,
    AVG(gmv) AS avg_gmv
FROM usershop_behavior
GROUP BY gender

SELECT 
    gender,
    COUNT(*) AS total,
    SUM(gmv) AS total_gmv,
    AVG(gmv) AS avg_gmv
FROM usershop_behavior
WHERE gender='M'
GROUP BY gender

BSI index can be useful in topk scenarios

SELECT *
FROM usershop_behavior
ORDER BY gmv DESC
LIMIT 10;

SELECT *
FROM usershop_behavior
WHERE gender='M'
ORDER BY gmv DESC
LIMIT 10;

Solution

Apache Flink and Apache Spark are already provides some interfaces. e.g.

Apache Flink: org.apache.flink.table.connector.source.abilities.SupportsAggregatePushDown

Apache Spark: org.apache.spark.sql.connector.read.SupportsPushDownTopN org.apache.spark.sql.connector.read.SupportsPushDownAggregates

When queries hit the bitmap index rules, we can rewrite TableScan to BitmapIndexScan.

Anything else?

Currently our index is designed to be used only for Data Skipping and it is not as reliable as filtering using partitioned keys. (We can't tell the flink&spark engine that filtering with indexes is reliable.)

This is because creating the index is split into several steps.

  1. stop the ingesting task
  2. using alter table to add index options
  3. call rewrite index procedure
  4. restart the ingesting task

We need to find a way to make indexes reliable, like partition keys. (e.g. throw exception when index is empty?) Otherwise it's hard for our index to do its job.

Are you willing to submit a PR?