Filter Pushdown on indexed columns about qbeast-spark HOT 6 CLOSED

From recent findings, we can skip the push down filters step, since the filters needed are already present in OTreeIndex matchingFiles method

A result from doing:

val df = spark.read.format("qbeast").load("/tmp/qbeast_table")
df.filter("user_id > 537764969").explain(true)

Shows the user_id filter as part of the DataFilters on the last step:

== Physical Plan ==
*(1) Filter (isnotnull(user_id#1762) AND (user_id#1762 >= 537764969))
+- *(1) ColumnarToRow
   +- FileScan parquet [event_time#1755,event_type#1756,product_id#1757,category_id#1758L,category_code#1759,brand#1760,price#1761,user_id#1762,user_session#1763] Batched: true, ...
DataFilters: [isnotnull(user_id#1762), (user_id#1762 >= 537764969)],...
Format: Parquet, Location: OTreeIndex[file:/tmp/qb-testing6614606061063411331], PartitionFilters: [], PushedFilters: [IsNotNull(user_id), GreaterThanOrEqual(user_id,537764969)], ReadSchema: struct<event_time:string,event_type:string,product_id:int,category_id:bigint,category_code:string...

from qbeast-spark.

Just curiosity, what will be in the DataFilters collection if the original filter uses OR instead of AND?

from qbeast-spark.

Good question! My guess is that they would not appear in DataFilters, but I will check it.

from qbeast-spark.

They are all present as a single Expression:

DataFilters: [(((user_id#1762 >= 537764969) OR ((user_id#1762 < 666666666) AND (product_id#1757 >= 6789009)))

from qbeast-spark.

So if we want a precise filtering we still need to work with AST, right?

from qbeast-spark.

Correct. For now, I think we should work only with conjunctions. There are functions in Spark that we can reproduce for splitting predicates, as Delta and other partition-aware formats do.
But maybe @cugni has other hints

Edited: In fact, for a query with only conjunctive predicates, Spark itself already separates them in different Expressions.

from qbeast-spark.