Include result serialization in query timings for dataframe entries - #1585
Include result serialization in query timings for dataframe entries#1585alexey-milovidov wants to merge 10 commits into
Conversation
ClickBench requires back-to-back runtimes that include sending the query result back to the client (README, "Output Suppression"). The polars entry (and the sibling in-process dataframe entries that share its server.py template) stopped the timer right after query execution and rendered the result afterwards. Move result rendering inside the timed section — matching what duckdb-dataframe and chdb-dataframe already do — and disable display truncation so the full result (every row, column, and complete cell contents) is rendered: - polars, polars-dataframe: pl.Config tbl_rows/tbl_cols/tbl_width_chars set to -1 and fmt_str_lengths raised; str()/repr() moved inside the timing. - pandas, dask: pandas display options set to unlimited; str() moved inside the timing (dask results are pandas objects). - daft-parquet, daft-parquet-partitioned: collect(num_preview_rows=None) so __str__ renders every row, plus to_pydict() inside the timing as the fetchall() equivalent (daft's table repr truncates long cells). All ClickBench query results are small (LIMIT 10/25), so the rendering cost is negligible; this is a methodology fix, not a perf change. Closes #1397 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
GitHub Actions
Error
…ars (c6a.2xlarge, c6a.4xlarge, c6a.metal, c7a.metal-48xl, c8g.4xlarge, c8g.metal-48xl)
|
Results for Logs:
|
….2xlarge, c6a.4xlarge, c6a.xlarge, c8g.4xlarge)
|
Results for Logs:
|
…, c8g.metal-48xl)
|
Results for Logs:
|
…xlarge, c7a.metal-48xl, c8g.metal-48xl)
|
Results for Logs:
|
|
Results for Logs:
|
|
Results for Logs:
|
|
Results for Logs:
|
|
Results for Logs:
|
|
Results for Logs:
|
|
The run of Logs:
|
|
The run of Logs:
|
1 participant
Fixes #1397.
ClickBench requires back-to-back runtimes that include sending the query result back to the client (README, "Output Suppression"). As @vogelsgesang pointed out, the polars entry stopped the timer right after query execution and rendered the result afterwards. The sibling in-process dataframe entries built on the same
server.pytemplate (polars-dataframe,pandas,dask,daft-parquet,daft-parquet-partitioned) had the same flaw, so this fixes all six.The fix moves result rendering inside the timed section — matching what
duckdb-dataframe(str(rel)in the timing) andchdb-dataframe(Prettyformat in the timing) already do — and additionally disables display truncation so the full result is rendered, since a truncated render would itself be partial output suppression:pl.Config.set_tbl_rows(-1),set_tbl_cols(-1),set_tbl_width_chars(-1),set_fmt_str_lengths(65535)at startup;str()/repr()of the collected result moved inside the timing. Verified against a live server: 25-row results render all rows, Q30-style 90-column results render all columns, long strings are not truncated.display.max_rows/max_columns/width/max_colwidthset to unlimited;str()moved inside the timing (dask results are pandas objects afterdask.compute()).collect(num_preview_rows=None)so__str__renders every row (default previews only 8), plusto_pydict()inside the timing as thefetchall()equivalent, because daft's table repr always truncates long cell contents.Entries already compliant and untouched:
duckdb-dataframe,chdb-dataframe,duckdb-memory(fetchall()in the timing),sirius(CLI output captured in the timing).All ClickBench query results are small (LIMIT 10/25), so the added rendering cost is well below the 1 ms timing resolution — this is a methodology fix and existing results don't need to be re-run.
🤖 Generated with Claude Code