Uploaded image for project: 'IMPALA'
  1. IMPALA
  2. IMPALA-1755

Disabling codegen on parquet scanner causes some queries' performance to regress

    XMLWordPrintableJSON

Details

    • Bug
    • Status: Resolved
    • Blocker
    • Resolution: Fixed
    • Impala 2.2
    • None
    • Backend

    Description

      The regression was caused by https://github.com/cloudera/Impala/commit/0686cd9c3ed7ae48d5bd4fe602266034ef871ffc, which disables codegen on any plan fragments with no codegen-enabled operators (e.g., a scan over a parquet table feeding directly into an exchange). This speeds up some queries by removing codegen overhead, but can slow down queries where the predicates on a parquet scanner benefit from codegen.

      TPCDS-Q8 is definitely affected by this, and TPCDS-Q47 and TPCH-Q16 may be as well (I haven't dug into it).

      Attachments

        1. slow_q8.txt
          64 kB
          Martin Grund
        2. fast_q8.txt
          64 kB
          Martin Grund

        Activity

          People

            mmokhtar Mostafa Mokhtar
            mgrund_impala_bb91 Martin Grund
            Votes:
            0 Vote for this issue
            Watchers:
            9 Start watching this issue

            Dates

              Created:
              Updated:
              Resolved: