feat(presto/slurm): immutable result dirs, srun tracking, port preflight

[misiugodfrey]

Summary

Hardens the NVL72 Slurm benchmark scripts: results are written to a stable per-job directory, process teardown is targeted rather than broad, and a new coordinator query collector enables post-run analysis.

Based on kjmph's bench33 POC patch.

• Immutable result directories: benchmark output now lands in result_dir_<SLURM_JOB_ID>/ rather than a fixed path that gets overwritten on every run; latest_result_dir.txt is written alongside as a convenience pointer
• srun PID tracking: COORD_SRUN_PID and WORKER_SRUN_PIDS are captured and used for targeted kill on exit instead of broad pkill patterns that could terminate unrelated jobs
• on_exit trap improvements: both run-presto-benchmarks.sh and run-analyze-tables.sh now use consistent on_exit traps that clean up in the right order
• Port preflight infrastructure: canonicalize_file_path, preflight_file, preflight_image_roles helpers for validating image paths and roles before container launch
• UCX TLS defaults per variant: CPU gets sysv,tcp,self; GPU keeps existing ^ib,ud:aux,sm; both remain shell-overridable via UCX_TLS
• collect-coordinator-queries.sh (new): fetches the coordinator's retained /v1/query list and full per-query JSON for post-mortem analysis; works by Slurm job ID or direct URL; uses srun --overlap for zero-extra-allocation Slurm access
• summarize_coordinator_queries.py (new): stdlib-only TSV report writer that reads the collected JSON and produces summary.tsv + per-query operator_summaries/

Dependencies

No functional dependencies on PRs #1–#2. However, the split files (functions.sh, launch-run.sh, run-presto-benchmarks.sh) contain the full staging diff — rebase PRs #4 and #5 on top of this once merged to drop the duplicated hunks.

Verification

• Run a benchmark job and verify output is written to result_dir_<SLURM_JOB_ID>/ with a latest_result_dir.txt pointing to it
• Run a second benchmark job and confirm the first job's result directory is not overwritten
• Kill a benchmark job mid-run and verify the on_exit trap cleanly terminates the coordinator and worker srun processes without leaving orphaned containers
• Run collect-coordinator-queries.sh <job_id> against a running job; confirm queries.json, per-query JSON files, summary.tsv, and operator_summaries/ are written
• Run collect-coordinator-queries.sh --server http://host:9200 (direct URL mode) and confirm it works without a Slurm job ID
• Verify summarize_coordinator_queries.py produces a correctly-formatted TSV with one row per query and truncates long SQL previews at 180 characters
• Run ./ci/check_style.sh on changed Python and shell files

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.