chore(perf-issues): Setup N+1 DB query experiment (#90530)

Similar to #90070 but for the
N+1 DB query detector.

This PR has no functional changes, and the new group type is not
released, so will not produce any results.
A follow up PR will have the detector changes, and roll them out
internally to test.

Author: leeandher

src/sentry/issues/grouptype.py

@@ -332,6 +332,17 @@ class PerformanceNPlusOneGroupType(PerformanceGroupTypeDefaults, GroupType):
     released = True
 
 
+@dataclass(frozen=True)
+class PerformanceNPlusOneExperimentalGroupType(PerformanceGroupTypeDefaults, GroupType):
+    type_id = 1906
+    slug = "performance_n_plus_one_db_queries_experimental"
+    description = "N+1 Query (Experimental)"
+    category = GroupCategory.PERFORMANCE.value
+    category_v2 = GroupCategory.PERFORMANCE_BEST_PRACTICE.value
+    default_priority = PriorityLevel.LOW
+    released = False
+
+
 @dataclass(frozen=True)
 class PerformanceConsecutiveDBQueriesGroupType(PerformanceGroupTypeDefaults, GroupType):
     type_id = 1007

src/sentry/utils/performance_issues/base.py

@@ -30,6 +30,7 @@ class DetectorType(Enum):
     DB_MAIN_THREAD = "db_main_thread"
     HTTP_OVERHEAD = "http_overhead"
     EXPERIMENTAL_N_PLUS_ONE_API_CALLS = "experimental_n_plus_one_api_calls"
+    EXPERIMENTAL_N_PLUS_ONE_DB_QUERIES = "experimental_n_plus_one_db_queries"
 
 
 # Detector and the corresponding system option must be added to this list to have issues created.

src/sentry/utils/performance_issues/detectors/experiments/n_plus_one_db_span_detector.py

@@ -0,0 +1,279 @@
+from __future__ import annotations
+
+import hashlib
+from typing import Any
+
+from sentry.issues.grouptype import PerformanceNPlusOneExperimentalGroupType
+from sentry.issues.issue_occurrence import IssueEvidence
+from sentry.models.organization import Organization
+from sentry.models.project import Project
+from sentry.utils import metrics
+from sentry.utils.performance_issues.base import (
+    DetectorType,
+    PerformanceDetector,
+    get_notification_attachment_body,
+    get_span_evidence_value,
+    total_span_time,
+)
+from sentry.utils.performance_issues.performance_problem import PerformanceProblem
+from sentry.utils.performance_issues.types import Span
+from sentry.utils.safe import get_path
+
+
+class NPlusOneDBSpanExperimentalDetector(PerformanceDetector):
+    """
+    Detector goals:
+      - identify a database N+1 query with high accuracy
+      - collect enough information to create a good fingerprint (see below)
+      - only return issues with good fingerprints
+
+    A good fingerprint is one that gives us confidence that, if two fingerprints
+    match, then they correspond to the same issue location in code (and
+    therefore, the same fix).
+
+    To do this we look for a specific structure:
+
+      [-------- transaction span -----------]
+         [-------- parent span -----------]
+            [source query]
+                          [n0]
+                              [n1]
+                                  [n2]
+                                      ...
+
+    If we detect two different N+1 problems, and both have matching parents,
+    source queries, and repeated (n) queries, then we can be fairly confident
+    they are the same issue.
+    """
+
+    __slots__ = (
+        "stored_problems",
+        "potential_parents",
+        "source_span",
+        "n_hash",
+        "n_spans",
+        "transaction",
+    )
+
+    type = DetectorType.EXPERIMENTAL_N_PLUS_ONE_DB_QUERIES
+    settings_key = DetectorType.EXPERIMENTAL_N_PLUS_ONE_DB_QUERIES
+
+    def __init__(self, settings: dict[DetectorType, Any], event: dict[str, Any]) -> None:
+        super().__init__(settings, event)
+
+        self.stored_problems = {}
+        self.potential_parents = {}
+        self.n_hash: str | None = None
+        self.n_spans: list[Span] = []
+        self.source_span: Span | None = None
+        root_span = get_path(self._event, "contexts", "trace")
+        if root_span:
+            self.potential_parents[root_span.get("span_id")] = root_span
+
+    def is_creation_allowed_for_system(self) -> bool:
+        # Defer to the issue platform for whether to create issues
+        # See https://develop.sentry.dev/backend/issue-platform/#releasing-your-issue-type
+        return True
+
+    def is_creation_allowed_for_organization(self, organization: Organization | None) -> bool:
+        return True  # This detector is fully rolled out
+
+    def is_creation_allowed_for_project(self, project: Project | None) -> bool:
+        return self.settings["detection_enabled"]
+
+    def visit_span(self, span: Span) -> None:
+        span_id = span.get("span_id", None)
+        op = span.get("op", None)
+        if not span_id or not op:
+            return
+
+        if not self._is_db_op(op):
+            # This breaks up the N+1 we're currently tracking.
+            self._maybe_store_problem()
+            self._reset_detection()
+            # Treat it as a potential parent as long as it isn't the root span.
+            if span.get("parent_span_id", None):
+                self.potential_parents[span_id] = span
+            return
+
+        if not self.source_span:
+            # We aren't currently tracking an N+1. Maybe this span triggers one!
+            self._maybe_use_as_source(span)
+            return
+
+        # If we got this far, we know we're a DB span and we're looking for a
+        # sequence of N identical DB spans.
+        if self._continues_n_plus_1(span):
+            self.n_spans.append(span)
+        else:
+            previous_span = self.n_spans[-1] if self.n_spans else None
+            self._maybe_store_problem()
+            self._reset_detection()
+
+            # Maybe this DB span starts a whole new N+1!
+            if previous_span:
+                self._maybe_use_as_source(previous_span)
+            if self.source_span and self._continues_n_plus_1(span):
+                self.n_spans.append(span)
+            else:
+                self.source_span = None
+                self._maybe_use_as_source(span)
+
+    def on_complete(self) -> None:
+        self._maybe_store_problem()
+
+    def _is_db_op(self, op: str) -> bool:
+        return op.startswith("db") and not op.startswith("db.redis")
+
+    def _maybe_use_as_source(self, span: Span) -> None:
+        parent_span_id = span.get("parent_span_id", None)
+        if not parent_span_id or parent_span_id not in self.potential_parents:
+            return
+
+        self.source_span = span
+
+    def _continues_n_plus_1(self, span: Span) -> bool:
+        if self.source_span is None:
+            return False
+
+        expected_parent_id = self.source_span.get("parent_span_id", None)
+        parent_id = span.get("parent_span_id", None)
+        if not parent_id or parent_id != expected_parent_id:
+            return False
+
+        span_hash = span.get("hash", None)
+        if not span_hash:
+            return False
+
+        if span_hash == self.source_span.get("hash", None):
+            # The source span and n repeating spans must have different queries.
+            return False
+
+        if not self.n_hash:
+            self.n_hash = span_hash
+            return True
+
+        return span_hash == self.n_hash
+
+    def _maybe_store_problem(self) -> None:
+        if not self.source_span or not self.n_spans:
+            return
+
+        # Do we have enough spans?
+        count = self.settings.get("count")
+        if len(self.n_spans) < count:
+            return
+
+        # Do the spans take enough total time?
+        if not self._is_slower_than_threshold():
+            return
+
+        # We require a parent span in order to improve our fingerprint accuracy.
+        parent_span_id = self.source_span.get("parent_span_id", None)
+        if not parent_span_id:
+            return
+        parent_span = self.potential_parents[parent_span_id]
+        if not parent_span:
+            return
+
+        # Track how many N+1-looking problems we found but dropped because we
+        # couldn't be sure (maybe the truncated part of the query differs).
+        if not contains_complete_query(
+            self.source_span, is_source=True
+        ) or not contains_complete_query(self.n_spans[0]):
+            metrics.incr("performance.performance_issue.truncated_np1_db")
+            return
+
+        if not self._contains_valid_repeating_query(self.n_spans[0]):
+            metrics.incr("performance.performance_issue.unparametrized_first_span")
+            return
+
+        fingerprint = self._fingerprint(
+            parent_span.get("op", None),
+            parent_span.get("hash", None),
+            self.source_span.get("hash", None),
+            self.n_spans[0].get("hash", None),
+        )
+        if fingerprint not in self.stored_problems:
+            self._metrics_for_extra_matching_spans()
+
+            offender_span_ids = [span.get("span_id", None) for span in self.n_spans]
+
+            self.stored_problems[fingerprint] = PerformanceProblem(
+                fingerprint=fingerprint,
+                op="db",
+                desc=self.n_spans[0].get("description", ""),
+                type=PerformanceNPlusOneExperimentalGroupType,
+                parent_span_ids=[parent_span_id],
+                cause_span_ids=[self.source_span.get("span_id", None)],
+                offender_span_ids=offender_span_ids,
+                evidence_display=[
+                    IssueEvidence(
+                        name="Offending Spans",
+                        value=get_notification_attachment_body(
+                            "db",
+                            self.n_spans[0].get("description", ""),
+                        ),
+                        # Has to be marked important to be displayed in the notifications
+                        important=True,
+                    )
+                ],
+                evidence_data={
+                    "transaction_name": self._event.get("transaction", ""),
+                    "op": "db",
+                    "parent_span_ids": [parent_span_id],
+                    "parent_span": get_span_evidence_value(parent_span),
+                    "cause_span_ids": [self.source_span.get("span_id", None)],
+                    "offender_span_ids": offender_span_ids,
+                    "repeating_spans": get_span_evidence_value(self.n_spans[0]),
+                    "repeating_spans_compact": get_span_evidence_value(
+                        self.n_spans[0], include_op=False
+                    ),
+                    "num_repeating_spans": str(len(offender_span_ids)),
+                },
+            )
+
+    def _is_slower_than_threshold(self) -> bool:
+        duration_threshold = self.settings.get("duration_threshold")
+        return total_span_time(self.n_spans) >= duration_threshold
+
+    def _contains_valid_repeating_query(self, span: Span) -> bool:
+        # Make sure we at least have a space, to exclude e.g. MongoDB and
+        # Prisma's `rawQuery`.
+        query = span.get("description", None)
+        return bool(query) and " " in query
+
+    def _metrics_for_extra_matching_spans(self) -> None:
+        # Checks for any extra spans that match the detected problem but are not part of affected spans.
+        # Temporary check since we eventually want to capture extra perf problems on the initial pass while walking spans.
+        n_count = len(self.n_spans)
+        all_matching_spans = [
+            span
+            for span in self._event.get("spans", [])
+            if span.get("span_id", None) == self.n_hash
+        ]
+        all_count = len(all_matching_spans)
+        if n_count > 0 and n_count != all_count:
+            metrics.incr("performance.performance_issue.np1_db.extra_spans")
+
+    def _reset_detection(self) -> None:
+        self.source_span = None
+        self.n_hash = None
+        self.n_spans = []
+
+    def _fingerprint(self, parent_op: str, parent_hash: str, source_hash: str, n_hash: str) -> str:
+        # XXX: this has to be a hardcoded string otherwise grouping will break
+        problem_class = "GroupType.PERFORMANCE_N_PLUS_ONE_DB_QUERIES"
+        full_fingerprint = hashlib.sha1(
+            (str(parent_op) + str(parent_hash) + str(source_hash) + str(n_hash)).encode("utf8"),
+        ).hexdigest()
+        return f"1-{problem_class}-{full_fingerprint}"
+
+
+def contains_complete_query(span: Span, is_source: bool | None = False) -> bool:
+    # Remove the truncation check from the n_plus_one db detector.
+    query = span.get("description", None)
+    if is_source and query:
+        return True
+    else:
+        return query and not query.endswith("...")

src/sentry/utils/performance_issues/performance_detection.py

@@ -243,6 +243,11 @@ def get_detection_settings(project_id: int | None = None) -> dict[DetectorType,
             "duration_threshold": settings["n_plus_one_db_duration_threshold"],  # ms
             "detection_enabled": settings["n_plus_one_db_queries_detection_enabled"],
         },
+        DetectorType.EXPERIMENTAL_N_PLUS_ONE_DB_QUERIES: {
+            "count": settings["n_plus_one_db_count"],
+            "duration_threshold": settings["n_plus_one_db_duration_threshold"],  # ms
+            "detection_enabled": settings["n_plus_one_db_queries_detection_enabled"],
+        },
         DetectorType.CONSECUTIVE_DB_OP: {
             # time saved by running all queries in parallel
             "min_time_saved": settings["consecutive_db_min_time_saved_threshold"],  # ms