abstractdog commented on PR #6793:
URL: https://github.com/apache/hive/pull/6793#issuecomment-5926538252

   latest testing results:
   
   ## TPC-DS A/B: `hive.llap.io.parquet.native.enabled`
   
   99 queries × 3 reps × 2 arms = **594 runs, 0 failures**, against 
`tpcds_partitioned_parquet_10`.
   Metric is `Run DAG` from the Tez query summary (pure DAG execution, excluding
   compile/submit/fetch). Cluster restarted before each arm.
   
   **A/B validity:** `false` arm showed LLAP IO counters in 0/297 runs, `true` 
arm in 297/297 —
   the flag really does switch the read path, so the comparison is valid.
   
   ### Correctness: no differences
   
   | | queries |
   |---|---|
   | byte-identical output | 90 |
   | identical, both empty | 6 |
   | identical multiset, row order differs | 3 (q31, q65, q71) |
   | **value or row-count differenc
   
   The 3 ordering differences are read-path defects: for those
   queries the row order also varies between repetitions with the flag held 
constant, in both
   arms (`ORDER BY` on constant/tietie-breaker). q77 is a fourth such
   query, which happened to order stably in the `false` arm.
   
   ### Performance: 9.1% faster warm, 19.5% faster on first execution
   
   | | `false` | `true` | change |
   |---|---:|---:|---:|
   | total `Run DAG`, warm | 995.8s
   | total `Run DAG`, first execution | 1605.4s | 1291.7s | **−19.5%** |
   
   Median per-query change **−7.9%**; 84/99 queries faster by some margin. 32 
robustly faster
   (every warm sample beats every ` noise, 7 material but
   indeterminate. The win is broad rather than a few outliers — the top 8 
queries account for
   only 46% of the total saving.
   
   <details>
   <summary>Caveats</summary>
   
   Single-box run, 10 GB scale factor. Run-to-run spread exceeds 15% between a 
query's own two
   warm reps for 25 (query, arm) pairs, so the aggregate and the robust counts 
are the
   load-bearing numbers — an individual query's percentage is indicative only.
   </details>


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to