fix(profiling): Cast profile chunk timestamps to DateTime64 (#74348)

By default snuba-sdk casts timestamps to a DateTime but here we are
using a DateTime64 column which is causing a bug due to the truncated
timestamp.

src/sentry/profiles/profile_chunks.py

@@ -1,4 +1,6 @@
-from snuba_sdk import Column, Condition, Direction, Op, OrderBy, Query, Request, Storage
+from datetime import datetime
+
+from snuba_sdk import Column, Condition, Direction, Function, Op, OrderBy, Query, Request, Storage
 
 from sentry import options
 from sentry.search.events.types import ParamsType
@@ -20,8 +22,16 @@ def get_chunk_ids(
  Column("chunk_id"),
  ],
  where=[
- Condition(Column("end_timestamp"), Op.GTE, params.get("start")),
- Condition(Column("start_timestamp"), Op.LT, params.get("end")),
+ Condition(
+ Column("end_timestamp"),
+ Op.GTE,
+ resolve_datetime64(params.get("start")),
+ ),
+ Condition(
+ Column("start_timestamp"),
+ Op.LT,
+ resolve_datetime64(params.get("end")),
+ ),
  Condition(Column("project_id"), Op.EQ, project_id),
  Condition(Column("profiler_id"), Op.EQ, profiler_id),
  ],
@@ -44,3 +54,23 @@ def get_chunk_ids(
  )
 
  return [row["chunk_id"] for row in result["data"]]
+
+
+def resolve_datetime64(value: datetime | None, precision: int = 6) -> Function | None:
+ # This is normally handled by the snuba-sdk but it assumes that the underlying
+ # table uses DateTime. Because we use DateTime64(6) as the underlying column,
+ # we need to cast to the same type or we risk truncating the timestamp which
+ # can lead to subtle errors.
+
+ if value is None:
+ return None
+
+ if isinstance(value, datetime) and value.tzinfo is not None:
+ # This is adapted from snuba-sdk
+ # See https://github.com/getsentry/snuba-sdk/blob/2f7f014920b4f527a87f18c05b6aa818212bec6e/snuba_sdk/visitors.py#L168-L172
+ delta = value.utcoffset()
+ assert delta is not None
+ value -= delta
+ value.replace(tzinfo=None)
+
+ return Function("toDateTime64", [value.isoformat(), precision])