Dynamic filter pushdown into IcebergTableScan

Open
  • Is something wrong?
#68 3 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
5/5
Estimated time
Over a week
Newbie friendliness
35/100
Issue type
Feature
Clarity
Mostly clear
Activity status
Active
Tech stack
rust
Domain
backend, databases, performance

Research direction

Start with crates/datafusion/src/read/scan.rs and trace how IcebergTableScan gathers and handles filter pushdown, then follow the existing expr_to_predicate path. The issue describes two layers: per-batch filtering here and a per-task dynamic predicate hook in iceberg-rust; the latter needs changes outside this repository. Done means dynamic filters reach the scan and results remain correct, with row-group pruning requiring the iceberg-rust hook.

Written by the indexing model from the issue text.

Description

Moved from apache/iceberg-rust#2376 at the suggestion of @alexanderbianchi, now that the DataFusion integration lives here.

Is your feature request related to a problem or challenge?

IcebergTableScan (crates/datafusion/src/read/scan.rs) does not take part in DataFusion's dynamic filter pushdown. The filters that HashJoinExec builds at execution time never reach the Iceberg scan. Probe-side scans like lineitem therefore can't prune row groups using the build side's join keys.

Example: TPC-H q14 (lineitem JOIN part ON l_partkey = p_partkey) at SF10. Even with enable_dynamic_filter_pushdown, parquet.pushdown_filters, parquet.reorder_filters and parquet.bloom_filter_on_read all on, lineitem is scanned in full. The dynamic filter is only evaluated row by row in the FilterExec above the scan, which doesn't skip any row groups. On the same data, Trino's Iceberg connector runs this query about 5x faster. Trino samples its DynamicFilter once per split and feeds it into row-group statistics and bloom filters before reading any data.

The scan takes filters: &[Expr] once and converts them into a single static Predicate. Nothing feeds it a predicate that only becomes useful mid-execution.

Describe the solution you'd like

Two layers:

1. In this repo: IcebergTableScan implements gather_filters_for_pushdown / handle_child_pushdown_result, accepts DynamicFilterPhysicalExprs from parent joins, and translates them to Iceberg Predicates through the existing expr_to_predicate path. The current soundness rule still applies: the predicate only has to be implied by the filter, and unconvertible parts are dropped.

2. In iceberg-rust: a hook that samples the dynamic predicate once per FileScanTask, along these lines:

pub trait DynamicPredicate: Send + Sync {
    /// Called right before a FileScanTask opens its Parquet file.
    /// None = not selective yet (e.g. build side not finished).
    fn current(&self) -> Option<Predicate>;
}

impl TableScanBuilder {
    pub fn with_dynamic_predicate(self, dp: Arc<dyn DynamicPredicate>) -> Self;
}

ArrowReader::process_file_scan_task would AND the result with the static predicate. From there it flows into the row filter, row-group pruning and page-index selection that the reader already supports.

Until (2) lands, (1) can apply the filters per batch after decode. That is already a win, but it doesn't reach row-group pruning.

Results from a downstream prototype
  • Per-batch only (layer 1): -21% total TPC-H time at SF1, -9.4% at SF10.
  • Per-task DynamicPredicate (both layers): -12.4% at SF10, all 22 queries returning matching results.
Open problem: caching the bound predicate across tasks

Each FileScanTask pays for conversion plus Predicate::bind. For a hash-join probe set of about 580K keys (q04 at SF10), bind takes about 61 ms and conversion about 19 ms. With about 88 lineitem files and parallelism 12, that's roughly 560 ms of wall-clock per scan.

I tried five cache designs and all of them regressed: Mutex + Arc::ptr_eq, OnceLock, IN-list caps of 4096 and 200, and per-filter Mutex<CachedBound>. Each failed in one of three ways:

  • DynamicFilterPhysicalExpr::current() calls remap_children, which returns a new Arc on the join pushdown path. Pointer-equality caches never hit.
  • In multi-join scans the filters seal at different times. A first-non-None cache locks in a partial predicate. That cost q04 +34%, q06 +74% and q16 +42%.
  • Locking on every per-task call adds contention across concurrent scan tasks.

A correct cache needs DataFusion to tell consumers when a filter's state has changed and when it is final. That could be an is_sealed() or a monotonic generation() on DynamicFilterPhysicalExpr. If that's the right direction, I can open a matching issue in apache/datafusion.

Willingness to contribute

Yes. I can open a PR for layer 1 here, and a PR in iceberg-rust for the DynamicPredicate hook.

Dominant language
Rust
Stars
30
Forks
30
Avg merge
3d 19h
Merged PRs (30d)
15

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from apache/datafusion-iceberg

All issues in apache/datafusion-iceberg

Similar issues

More Rust issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.