bugbug 0.0.650__tar.gz → 0.0.651__tar.gz

{bugbug-0.0.650 → bugbug-0.0.651}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: bugbug
-Version: 0.0.650
+Version: 0.0.651
 Summary: ML tools for Mozilla projects
 Author-email: Marco Castelluccio <mcastelluccio@mozilla.com>
 License-Expression: MPL-2.0

bugbug-0.0.651/VERSION

@@ -0,0 +1 @@
+0.0.651

{bugbug-0.0.650 → bugbug-0.0.651}/bugbug/models/testconfigselect.py

@@ -369,6 +369,8 @@ class GroupCandidates(NamedTuple):
 class PushCandidates(NamedTuple):
     groups: Sequence[GroupCandidates]
     patch: PatchInfo = NO_PATCH_INFO
+    # Configs already selected for the push (e.g. for other groups), whose setup cost is paid.
+    used_configs: frozenset[str] = frozenset()
 
 
 def _plainest_cheapest(configs: Iterable[str]) -> str:
@@ -441,7 +443,7 @@ class ConfigSelector:
         """Select configs for the groups of each push (see select_configs)."""
         selected: list[list[list[str]]] = [[[] for _ in push.groups] for push in pushes]
         coverage = [[0.0 for _ in push.groups] for push in pushes]
-        used: list[set[str]] = [set() for _ in pushes]
+        used: list[set[str]] = [set(push.used_configs) for push in pushes]
         active = {
             (p, g) for p, push in enumerate(pushes) for g in range(len(push.groups))
         }
@@ -520,6 +522,7 @@ class ConfigSelector:
         patch_files: Collection[str] = (),
         allowed_families: Collection[str] | None = None,
         very_likely_groups: Collection[str] = (),
+        used_configs: Collection[str] = (),
     ) -> dict[str, list[str]]:
         """Select the configs to run the groups of a push on.
 
@@ -533,6 +536,9 @@ class ConfigSelector:
         families building the modified files, see build_platforms), configs of other families are
         not candidates, unless a group has no candidate of the allowed families.
 
+        `used_configs` are configs already selected for other groups of the push (e.g. by a previous
+        call): their setup cost is considered paid.
+
         Groups without candidates are not in the result (callers should not restrict them).
         """
         patch = PatchInfo.from_files(patch_files) if patch_files else NO_PATCH_INFO
@@ -564,7 +570,9 @@ class ConfigSelector:
                 )
             )
 
-        selected = self._select([PushCandidates(groups, patch)], threshold)[0]
+        selected = self._select(
+            [PushCandidates(groups, patch, frozenset(used_configs))], threshold
+        )[0]
         return {g.group: sorted(configs) for g, configs in zip(groups, selected)}
 
     def _print_evaluation(

{bugbug-0.0.650 → bugbug-0.0.651}/bugbug/models/testselect.py

@@ -409,13 +409,20 @@ TUNED_XGBOOST_PARAMS = {"n_estimators": 400, "learning_rate": 0.03, "max_depth":
 
 # Number of runnables per push that the model should select at each confidence level used by the
 # Firefox taskgraph (e.g. bugbug-*-low/medium/high): roughly what the models selected at the fixed
-# 0.7/0.8/0.9 thresholds before the sample weighting and tuning changed their scale. Groups at the
-# "very_high" level run on a config of each OS family (see the HTTP service's config selection):
-# in offline evaluations, failing pushes caught at the same cost were ~2.3 points higher with one
-# such group per push, and about as high with 0.5 to 3.
+# 0.7/0.8/0.9 thresholds before the sample weighting and tuning changed their scale. The
+# "very_low" level, used by the taskgraph on expanded pushes, selects three times as many as "low".
+# Groups at the "very_high" level run on a config of each OS family (see the HTTP service's config
+# selection): in offline evaluations, failing pushes caught at the same cost were ~2.3 points
+# higher with one such group per push, and about as high with 0.5 to 3.
 CONFIDENCE_LEVEL_TARGETS = {
-    "label": {"low": 26.0, "medium": 15.0, "high": 5.5},
-    "group": {"low": 103.0, "medium": 36.0, "high": 11.0, "very_high": 1.0},
+    "label": {"very_low": 78.0, "low": 26.0, "medium": 15.0, "high": 5.5},
+    "group": {
+        "very_low": 309.0,
+        "low": 103.0,
+        "medium": 36.0,
+        "high": 11.0,
+        "very_high": 1.0,
+    },
 }
 
 # Minimum share of the failing test pushes where the model selects a failure at the "low"

bugbug-0.0.650/VERSION

@@ -1 +0,0 @@
-0.0.650