ci(bench): keep compact-only SQL benchmark results in their own PR comment

Sharing one comment per benchmark id across presets let a `pr-compact`
run (Parquet + Vortex Compact, no Vortex) replace the Vortex results
posted by `action/bench-sql`, and the same happens for `appian-nvme` and
`tpch-s3-10` under `pr-all`, which are compact-only there too.

The matrix now emits a `comment_key` per entry: the benchmark id when
the run includes regular Vortex, otherwise `<id>-compact`. The workflow
tags PR comments by it, so every label upserts the same comment except
runs whose formats differ.

Signed-off-by: Joe Isaacs <joe.isaacs@live.co.uk>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018ixWKUArMHQ19JTDMdggWV

Author: joseph-isaacs

.github/workflows/sql-bench-matrix.yml

@@ -258,10 +258,10 @@ jobs:
         with:
           file-path: comment.md
           # There is exactly one comment per comment-tag. If a comment with this tag already exists,
-          # this action will *update* the comment instead of posting a new comment. The preset is
-          # deliberately not part of the tag, so every preset that runs a benchmark updates the same
-          # comment and the latest run wins.
-          comment-tag: bench-pr-comment-${{ matrix.id }}
+          # this action will *update* the comment instead of posting a new comment. `comment_key`
+          # ignores the preset, so every label that runs a benchmark updates the same comment, but
+          # separates compact-only runs, whose formats differ (see bench-orchestrator ci_matrix).
+          comment-tag: bench-pr-comment-${{ matrix.comment_key }}
 
       - name: Comment PR on failure
         if: failure() && inputs.mode == 'pr' && github.event.pull_request.head.repo.fork == false
@@ -271,7 +271,7 @@ jobs:
             # 🚨🚨🚨❌❌❌ SQL BENCHMARK FAILED ❌❌❌🚨🚨🚨
 
             Benchmark `${{ matrix.name }}` (${{ inputs.matrix_preset }}) failed! Check the [workflow run](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}) for details.
-          comment-tag: bench-pr-comment-${{ matrix.id }}
+          comment-tag: bench-pr-comment-${{ matrix.comment_key }}
 
       - name: Upload Benchmark Results
         if: inputs.mode == 'develop'

bench-orchestrator/bench_orchestrator/ci_matrix/render.py

@@ -26,9 +26,19 @@ def _data_formats(coverage: Coverage) -> list[Format]:
     return [fmt for fmt in _FORMAT_ORDER if fmt in present and fmt not in _NOT_GENERATED]
 
 
+def _comment_key(benchmark: BenchmarkCase, coverage: Coverage) -> str:
+    # PR benchmark comments are upserted by this key, so every preset that runs a benchmark updates
+    # one comment. Runs without regular Vortex (Vortex Compact plus Parquet controls) report
+    # different formats, so they get their own comment instead of replacing the Vortex results.
+    if any(target.format == Format.VORTEX for target in coverage.targets):
+        return benchmark.id
+    return f"{benchmark.id}-compact"
+
+
 def _matrix_entry(benchmark: BenchmarkCase, coverage: Coverage) -> dict[str, object]:
     entry: dict[str, object] = {
         "id": benchmark.id,
+        "comment_key": _comment_key(benchmark, coverage),
         "subcommand": benchmark.subcommand,
         "name": benchmark.name,
         "targets": [target.to_dict() for target in coverage.targets],

bench-orchestrator/tests/test_matrix.py

@@ -125,6 +125,20 @@ def test_pr_full_covers_pr() -> None:
         assert set(cast("list[str]", entry["data_formats"])) <= set(cast("list[str]", full_entry["data_formats"]))
 
 
+def test_comment_keys_share_pr_comments_by_format() -> None:
+    keys: dict[str, set[str]] = {}
+    for preset in ("pr", "pr-compact", "pr-all", "pr-full"):
+        for entry in _entries(preset):
+            formats = {file_format for _engine, file_format in _targets(entry)}
+            benchmark_id = cast("str", entry["id"])
+            expected = benchmark_id if "vortex" in formats else f"{benchmark_id}-compact"
+            assert entry["comment_key"] == expected
+            keys.setdefault(benchmark_id, set()).add(cast("str", entry["comment_key"]))
+
+    assert keys["tpch-nvme"] == {"tpch-nvme", "tpch-nvme-compact"}
+    assert keys["polarsignals"] == {"polarsignals"}
+
+
 def test_resolver_rejects_empty_targets() -> None:
     benchmark = replace(
         CATALOG.benchmarks[0],