Return TableProviderFilterPushDown::Exact when Parquet Pushdown Enabled

[tustvold]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]
(This section helps Arrow developers understand the context and why for this feature, in addition to the what)

Currently even when parquet predicate pushdown is enabled, and the predicate can be fully pushed down, the physical plan still contains a FilterExec when using ListingTable

| physical_plan | ProjectionExec: expr=[service@0 as service, host@1 as host, pod@2 as pod, container@3 as container, image@4 as image, time@5 as time, client_addr@6 as client_addr, request_duration_ns@7 as request_duration_ns, request_user_agent@8 as request_user_agent, request_method@9 as request_method, request_host@10 as request_host, request_bytes@11 as request_bytes, response_bytes@12 as response_bytes, response_status@13 as response_status]                            |
|               |   CoalesceBatchesExec: target_batch_size=4096                                                                                                                                                                                                                                                                                                                                                                                                                                |
|               |     FilterExec: container@3 = backend_container_0 OR pod@2 = aqcathnxqsphdhgjtgvxsfyiwbmhlmg                                                                                                                                                                                                                                                                                                                                                                                 |
|               |       RepartitionExec: partitioning=RoundRobinBatch(8)                                                                                                                                                                                                                                                                                                                                                                                                                       |
|               |         ParquetExec: limit=None, partitions=[home/raphael/Downloads/data.parquet], predicate=container_min@0 <= backend_container_0 AND backend_container_0 <= container_max@1 OR pod_min@2 <= aqcathnxqsphdhgjtgvxsfyiwbmhlmg AND aqcathnxqsphdhgjtgvxsfyiwbmhlmg <= pod_max@3, projection=[service, host, pod, container, image, time, client_addr, request_duration_ns, request_user_agent, request_method, request_host, request_bytes, response_bytes, response_status] |
|               |        

Describe the solution you'd like

ListingTable::supports_filter_pushdown should return TableProviderFilterPushDown::Exact when

• Parquet predicate pushdown is enabled
• The FileFormat is parquet
• The predicate is fully pushed down by ParquetExec (not all predicates are supported)

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

[alamb]

This is a good proposal I think -- it would skip unecessary filtering and likely make plans faster. It will become more useful (maybe even necessary) when predicate pushdown is enabled by default -- #3463

[jychen7]

I think it is a good idea. I confirm it would help improve performance by 30x in #5404 (comment).

---

I am thinking to pick this up but seems the work is not as trivial as I thought. Following are my two cents' thoughts.

Related code to improve is following
https://github.com/apache/arrow-datafusion/blob/7545177001b9dc04951f0c1c2008509f3895de8e/datafusion/core/src/datasource/listing/table.rs#L736-L756

1. The FileFormat is parquet

This is trivial, we can match self.options.format with ParquetFormat

2. Parquet predicate pushdown is enabled

This setting could come from ParquetExec.pushdown_filters or ConfigOptions (more specifically, ConfigOptions.execution.parquet.pushdown_filters)

https://github.com/apache/arrow-datafusion/blob/7545177001b9dc04951f0c1c2008509f3895de8e/datafusion/core/src/physical_plan/file_format/parquet.rs#L393

However, when ListingTable.supports_filter_pushdown is called, ParquetExec is not created yet, nor does SessionState being passed as input.

Idea: shall we add state: &SessionState as input for ListingTable.supports_filter_pushdown, the same as ListingTable.scan ❓

https://github.com/apache/arrow-datafusion/blob/7545177001b9dc04951f0c1c2008509f3895de8e/datafusion/core/src/datasource/listing/table.rs#L668-L670

3. The predicate is fully pushed down by ParquetExec (not all predicates are supported)

is it making sure the row_filter is built from the predicate without error? If so, seems it requires converting a logical Expr to PhysicalExpr first.

https://github.com/apache/arrow-datafusion/blob/7545177001b9dc04951f0c1c2008509f3895de8e/datafusion/core/src/physical_plan/file_format/parquet.rs#L527-L547

[tustvold]

I wonder if it might be possible to always return exact for Parquet files, and to just manually insert a FilterExec for any predicates that can't be pushed down

Tagging @alamb and @crepererum who are more familiar with this part of the codebase

[alamb]

I'll try and look into this in more detail tomorrow

[crepererum]

I wonder if it might be possible to always return exact for Parquet files, and to just manually insert a FilterExec for any predicates that can't be pushed down

That's what we want to do for InfluxDB IOx as well: https://github.com/influxdata/influxdb_iox/issues/7408

[alamb]

I am going to find time sometime this week

[alamb]

Background Reading

Here is a background article about parquet predicate pushdown: https://www.influxdata.com/blog/querying-parquet-millisecond-latency/#heading5

Specifically the section on Late materialization describes what is done in the RowFilter code

Code links for Listing Table

Supports filter pushdown: https://docs.rs/datafusion/latest/datafusion/datasource/listing/struct.ListingTable.html#impl-TableProvider-for-ListingTable

Conditions:

Here are the conditions @tustvold lists above, and some links about what they mean

Parquet predicate pushdown is enabled

This refers to the datafusion.execution.parquet.pushdown_filters configuration setting (docs here)

The FileFormat is parquet

This should be relatively straightforward to determine

The predicate is fully pushed down by ParquetExec (not all predicates are supported)

"pushdown" refers to this code in ParquetExec (there is some background information in the parquet RowFilter API)

datafusion/datafusion/core/src/datasource/physical_plan/parquet/row_filter.rs

Lines 43 to 70 in 16a3557

	/// This module contains utilities for enabling the pushdown of DataFusion filter predicates (which
	/// can be any DataFusion `Expr` that evaluates to a `BooleanArray`) to the parquet decoder level in `arrow-rs`.
	/// DataFusion will use a `ParquetRecordBatchStream` to read data from parquet into arrow `RecordBatch`es.
	/// When constructing the `ParquetRecordBatchStream` you can provide a `RowFilter` which is itself just a vector
	/// of `Box<dyn ArrowPredicate>`. During decoding, the predicates are evaluated to generate a mask which is used
	/// to avoid decoding rows in projected columns which are not selected which can significantly reduce the amount
	/// of compute required for decoding.
	///
	/// Since the predicates are applied serially in the order defined in the `RowFilter`, the optimal ordering
	/// will depend on the exact filters. The best filters to execute first have two properties:
	/// 1. The are relatively inexpensive to evaluate (e.g. they read column chunks which are relatively small)
	/// 2. They filter a lot of rows, reducing the amount of decoding required for subsequent filters and projected columns
	///
	/// Given the metadata exposed by parquet, the selectivity of filters is not easy to estimate so the heuristics we use here primarily
	/// focus on the evaluation cost.
	///
	/// The basic algorithm for constructing the `RowFilter` is as follows
	/// 1. Recursively break conjunctions into separate predicates. An expression like `a = 1 AND (b = 2 AND c = 3)` would be
	/// separated into the expressions `a = 1`, `b = 2`, and `c = 3`.
	/// 2. Determine whether each predicate is suitable as an `ArrowPredicate`. As long as the predicate does not reference any projected columns
	/// or columns with non-primitive types, then it is considered suitable.
	/// 3. Determine, for each predicate, the total compressed size of all columns required to evaluate the predicate.
	/// 4. Determine, for each predicate, whether all columns required to evaluate the expression are sorted.
	/// 5. Re-order the predicate by total size (from step 3).
	/// 6. Partition the predicates according to whether they are sorted (from step 4)
	/// 7. "Compile" each predicate `Expr` to a `DatafusionArrowPredicate`.
	/// 8. Build the `RowFilter` with the sorted predicates followed by the unsorted predicates. Within each partition
	/// the predicates will still be sorted by size.

[alamb]

I think there are likely two things needed for this PR:

1. Some API that ListingTable can use to determine if the predicates can be pushed down
2. Tests that show / confirm that using RowFilter really is an exact filtering

[alamb]

FWIW I think @itsjunetime is looking into this ticket

[itsjunetime]

Yes, sorry - I'm still somewhat new to this codebase, so I'm taking a good second to read through the associated docs and comments to figure it out first.

[crepererum]

No need to rush. I think @alamb's comment was mostly meant as an assignment signal, so that nobody else starts to work on it and we end up wasting resources 🙂

[alamb]

No need to rush. I think @alamb's comment was mostly meant as an assignment signal, so that nobody else starts to work on it and we end up wasting resources 🙂

Yes this is what I meant. Sorry about not being clear.

[alamb]

I filed #12115 to track some additional testing that I would like to do for this feature.

[alamb]

There is a nice PR here for this feature: #12135