Improved results with running 1 query test only

tpc-h/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "iw-tpch"
-version = "0.2.3"
+version = "0.2.4"
 description = "TPC-H benchmark runner for Snowflake interactive and standard warehouses"
 requires-python = ">=3.13"
 dependencies = [

tpc-h/src/tpch/execution.py

@@ -75,6 +75,7 @@ def best_result(name: str, iteration: int, attempts: list[AttemptResult]) -> Que
             "client_elapsed_s": best["client_elapsed_s"],
             "query_id": best["query_id"],
             "attempt_query_ids": [a["query_id"] for a in ok_attempts],
+            "attempt_client_times": [a["client_elapsed_s"] for a in ok_attempts],
             "result_columns": best["columns"],
             "result_rows": best["rows"],
             "error": None,
@@ -88,6 +89,7 @@ def best_result(name: str, iteration: int, attempts: list[AttemptResult]) -> Que
         "client_elapsed_s": last["client_elapsed_s"],
         "query_id": last["query_id"],
         "attempt_query_ids": [a["query_id"] for a in attempts if a["query_id"]],
+        "attempt_client_times": [a["client_elapsed_s"] for a in attempts],
         "error": last["error"],
     }
 
@@ -108,6 +110,7 @@ def avg_result(name: str, iteration: int, attempts: list[AttemptResult]) -> Quer
             "client_elapsed_s": avg_elapsed,
             "query_id": last_ok["query_id"],
             "attempt_query_ids": [a["query_id"] for a in ok_attempts],
+            "attempt_client_times": [a["client_elapsed_s"] for a in ok_attempts],
             "result_columns": last_ok["columns"],
             "result_rows": last_ok["rows"],
             "error": None,
@@ -121,6 +124,7 @@ def avg_result(name: str, iteration: int, attempts: list[AttemptResult]) -> Quer
         "client_elapsed_s": last["client_elapsed_s"],
         "query_id": last["query_id"],
         "attempt_query_ids": [a["query_id"] for a in attempts if a["query_id"]],
+        "attempt_client_times": [a["client_elapsed_s"] for a in attempts],
         "error": last["error"],
     }
 

tpc-h/src/tpch/results.py

@@ -42,6 +42,7 @@ def enrich_server_elapsed(conn, results: list[QueryResult], *, use_avg: bool = F
             r["server_elapsed_s"] = sum(servers) / len(servers)
         else:
             r["server_elapsed_s"] = min(servers)
+        r["attempt_server_times"] = servers
 
 
 def _stats(times: list[float]) -> dict:
@@ -61,11 +62,21 @@ def summarize(results: list[QueryResult]) -> dict:
     ok = [r for r in results if r["status"] == "OK"]
     failed = [r for r in results if r["status"] != "OK"]
     validation_failed = [r for r in ok if r.get("validation") == "FAIL"]
-    client_times = [r["client_elapsed_s"] for r in ok]
-    server_times = [
-        r["server_elapsed_s"] if r.get("server_elapsed_s") is not None else r["client_elapsed_s"]
-        for r in ok
-    ]
+
+    unique_queries = set(r["query"] for r in results)
+    single_query = len(unique_queries) == 1
+
+    if single_query:
+        client_times = [t for r in ok for t in r.get("attempt_client_times", [r["client_elapsed_s"]])]
+        server_times = [t for r in ok for t in r.get("attempt_server_times", [])]
+        if not server_times:
+            server_times = client_times
+    else:
+        client_times = [r["client_elapsed_s"] for r in ok]
+        server_times = [
+            r["server_elapsed_s"] if r.get("server_elapsed_s") is not None else r["client_elapsed_s"]
+            for r in ok
+        ]
     out = {
         "total_queries": len(results),
         "successful": len(ok),

tpc-h/src/tpch/types.py

@@ -21,9 +21,11 @@ class QueryResult(TypedDict, total=False):
     client_elapsed_s: float
     query_id: str | None
     attempt_query_ids: list[str | None]
+    attempt_client_times: list[float]
     result_columns: list[str]
     result_rows: list[tuple[Any, ...]]
     server_elapsed_s: float | None
+    attempt_server_times: list[float]
     error: str | None
     validation: str
     validation_error: str | None