andygrove opened a new issue, #2101: URL: https://github.com/apache/datafusion-ballista/issues/2101
### Is your feature request related to a problem or challenge? The `tpch` benchmark binary already writes a machine-readable JSON summary with `--output`, but two gaps make it awkward to use for tracking performance across refs: 1. **Results are lost when a query fails.** Both the `ballista` and `datafusion` run loops only write the summary *after* the whole loop, and a failing query aborts the run (`?`/`.unwrap()` on `ctx.sql`/`collect`). So a single failure — or an out-of-memory `SIGKILL` at a large scale factor — discards every result collected so far. This is easy to hit at SF1000. 2. **There is no built-in way to compare two runs.** Comparing a baseline against a candidate (per-query delta, totals, row-count agreement) has to be done by hand or with an out-of-tree script. ### Describe the solution you'd like - Persist the summary JSON after **every** query (atomically), so a killed run keeps completed results. - Record a failed query with an `error` message plus whatever iterations completed, and continue to the remaining queries; still exit non-zero if any query failed. - Add a `tpch compare <baseline.json> <candidate.json>` subcommand that diffs two runs query-by-query: fastest iteration per side, delta, delta %, suite totals, and row-count agreement. ### Additional context Complements the benchmarking guide added in #2075 by making the numbers that page records reproducible and comparable. Related to #1630. -- 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]
