Skip to content

Parquet Pushdown Filters: Way to keep the I/O pattern the same when pushdown filters are enabled #24393

Description

@alamb

Is your feature request related to a problem or challenge?

Namely, the idea is to avoid regressions when we turn on filter pushdown

We haven't been able to turn on predicate pushdown because it caused slowdowns. The slowdowns were caused by two reasons:

  1. predicate pushdown changes the I/O patterns
  2. Evaluating predicates is computationally slower in some cases (figuring out bitmasks, rows that passed, etc)

I think after the great work from @hhhizzz @haohuaijin and others upstream in arrow-rs/parquet we are close (if not already) at the point where it is computationally the same or faster

All that remains is the different I/O

The I/O pattern today is

  1. Default: read all column data needed for both filtering and projections when reading a row group
  2. Predicate Pushdown (pushdown_filters): progressive I/O: sequentially fetch the columns required to evaluate filter expressions, and then after filtering is complete, fetches columns needed to evaluate the projections. This stratgey (can) fetch fewer pages but results in more I/O requests (which can increase latency)

@Dandandan / @adriangb also noted on #23492 (comment) that we can't prune I/O for files without a page index (notably the clickbench dataset) anyways , so other than peak buffer needs, there is no benefit to changing the I/O to read columns sequentially in that case

Describe the solution you'd like

I would like is some way to turn on filter pushdown but keep the I/O pattern the same as today (aka read all required data pages for a Row Gorup) up front in one request

That way, once we sort out any lingering CPU slowdowns, we can turn on predicate pushdown by default and leave the I/O pattern as a policy choice (as the I/O policy will likely different for different I/O systems -- e.g. NVMe vs ObjectStore.

Note a small variation could be to automatically use one-shot I/O when the file has no offset/page index (since progressive fetch can't prune I/O there anyway), and use progressive I/O only when the index exists.

Describe alternatives you've considered

So I think we should separate out the notions of "filter evaluation mechanics" from the I/O pattern.

One idea is to add a parquet configuration setting like 'progressive_io' (defaults to false, the behavior today). When true it would fetch data progressively, as happens today when filter_pushdown is enabled

So we would have a feature matrix like

filter_pushdown progressive_io description
false false Same as todays behavior -- one shot I/O
true false filter pushdown, but still keep same I/O (my proposed new default)
false true N/A (just a decode, so no progressive filtering)
true true Progressive filter evaluation and I/O (what happens today when filter_pushdown) is enabled

Additional context

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions