Hi all! As you may know there is an activity on integration of Apache Calcite query optimizer into Ignite codebase is being carried out [1],[2].
One of a bunch of problems in this integration is the absence of out-of-the-box support for secondary indexes in Apache Calcite. After some research I came to conclusion that this problem has a couple of workarounds. Let's name them 1. Phoenix-style approach - representing secondary indexes as materialized views which are natively supported by Calcite engine [3] 2. Drill-style approach - pushing filters into the table scans and choose appropriate index for lookups when possible [4] Both these approaches have advantages and disadvantages: Phoenix style pros: - natural way of adding indexes as an alternative source of rows: index can be considered as a kind of sorted materialized view. - possibility of using index sortedness for stream aggregates, deduplication (DISTINCT operator), merge joins, etc. - ability to support other types of indexes (i.e. functional indexes). Phoenix style cons: - polluting optimizer's search space extra table scans hence increasing the planning time. Drill style pros: - easier to implement (although it's questionable). - search space is not inflated. Drill style cons: - missed opportunity to exploit sortedness. There is a good discussion about using both approaches can be found in [5]. I made a small sketch [6] in order to demonstrate the applicability of the Phoenix approach to Ignite. Key design concepts are: 1. On creating indexes are registered as tables in Calcite schema. This step is needed for internal Calcite's routines. 2. On planner initialization we register these indexes as materialized views in Calcite's optimizer using VolcanoPlanner#addMaterialization method. 3. Right before the query execution Calcite selects all materialized views (indexes) which can be potentially used in query. 4. During the query optimization indexes are registered by planner as usual TableScans and hence can be chosen by optimizer if they have lower cost. This sketch shows the ability to exploit index sortedness only. So the future work in this direction should be focused on using indexes for fast index lookups. At first glance FilterableTable and FilterTableScanRule are good points to start. We can push Filter into the TableScan and then use FilterableTable for fast index lookups avoiding reading the whole index on TableScan step and then filtering its output on the Filter step. What do you think? [1] http://apache-ignite-developers.2346864.n4.nabble.com/New-SQL-execution-engine-tt43724.html#none [2] https://cwiki.apache.org/confluence/display/IGNITE/IEP-37%3A+New+query+execution+engine [3] https://issues.apache.org/jira/browse/PHOENIX-2047 [4] https://issues.apache.org/jira/browse/DRILL-6381 [5] https://issues.apache.org/jira/browse/DRILL-3929 [6] https://github.com/apache/ignite/pull/7115 -- Kind Regards Roman Kondakov
