[IMPALA-5636] Impala writer claims that file uses BIT_PACKED encoding when it doesn't

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: Impala 2.10.0
• Component/s: Backend
• Labels: parquet
• Epic Color: ghx-label-1
• Assignee: Tianyi Wang
• Reporter: Tim Armstrong
• Created: 09/Jul/17 21:08
• Updated: 17/Aug/17 00:23
• Resolved: 31/Jul/17 17:37

Description

The Parquet writer always adds the BIT_PACKED and RLE encodings even though (I'm pretty sure) we never write out the BIT_PACKED encoding for rep or def levels. The BIT_PACKED encoding is deprecated according to the Parquet specification: https://github.com/Parquet/parquet-format/blob/master/Encodings.md and it is not clear that Impala can even read it correctly: IMPALA-3006

One way of seeing that Impala claims to need the BIT_PACKED encoding is to write a parquet file with Impala then inspect it with parquet-tool.

BIT_PACKED is never written by Impala's parquet writer - the definition levels are always written using an RleEncoder and reported as Encoding::RLE. Weirdly, the repetition levels are reported as "BIT_PACKED", but this encoding has no effect since we don't actually write out repetition levels.

    // The code that populates the column chunk metadata's encodings field
    // relies on these specific values for the definition/repetition level
    // encodings.
    header.definition_level_encoding = Encoding::RLE;
    header.repetition_level_encoding = Encoding::BIT_PACKED;
    current_page_->header.__set_data_page_header(header);

Issue Links

• relates to: IMPALA-3006 Impala may use incorrect bit order with BIT_PACKED encoding