arize-phoenix 11.7.0__py3-none-any.whl → 11.9.0__py3-none-any.whl

phoenix/server/api/types/Project.py

@@ -1,30 +1,33 @@
 from __future__ import annotations
 
 import operator
-from datetime import datetime, timedelta
-from typing import TYPE_CHECKING, Annotated, Any, ClassVar, Optional
+from datetime import datetime, timezone
+from typing import TYPE_CHECKING, Annotated, Any, ClassVar, Literal, Optional, cast
 
 import strawberry
-from aioitertools.itertools import islice
+from aioitertools.itertools import groupby, islice
 from openinference.semconv.trace import SpanAttributes
-from sqlalchemy import desc, distinct, func, or_, select
+from sqlalchemy import and_, desc, distinct, exists, func, or_, select
 from sqlalchemy.dialects import postgresql, sqlite
 from sqlalchemy.sql.elements import ColumnElement
 from sqlalchemy.sql.expression import tuple_
+from sqlalchemy.sql.functions import percentile_cont
 from strawberry import ID, UNSET, Private, lazy
-from strawberry.relay import Connection, Node, NodeID
+from strawberry.relay import Connection, Edge, Node, NodeID, PageInfo
 from strawberry.types import Info
 from typing_extensions import assert_never
 
-from phoenix.datetime_utils import right_open_time_range
+from phoenix.datetime_utils import get_timestamp_range, normalize_datetime, right_open_time_range
 from phoenix.db import models
-from phoenix.db.helpers import SupportedSQLDialect
+from phoenix.db.helpers import SupportedSQLDialect, date_trunc
 from phoenix.server.api.context import Context
+from phoenix.server.api.exceptions import BadRequest
 from phoenix.server.api.input_types.ProjectSessionSort import (
     ProjectSessionColumn,
     ProjectSessionSort,
 )
-from phoenix.server.api.input_types.SpanSort import SpanSort, SpanSortConfig
+from phoenix.server.api.input_types.SpanSort import SpanColumn, SpanSort, SpanSortConfig
+from phoenix.server.api.input_types.TimeBinConfig import TimeBinConfig, TimeBinScale
 from phoenix.server.api.input_types.TimeRange import TimeRange
 from phoenix.server.api.types.AnnotationConfig import AnnotationConfig, to_gql_annotation_config
 from phoenix.server.api.types.AnnotationSummary import AnnotationSummary
@@ -34,6 +37,7 @@ from phoenix.server.api.types.pagination import (
     ConnectionArgs,
     Cursor,
     CursorSortColumn,
+    CursorSortColumnDataType,
     CursorString,
     connection_from_cursors_and_nodes,
     connection_from_list,
@@ -45,6 +49,7 @@ from phoenix.server.api.types.SpanCostSummary import SpanCostSummary
 from phoenix.server.api.types.TimeSeries import TimeSeries, TimeSeriesDataPoint
 from phoenix.server.api.types.Trace import Trace
 from phoenix.server.api.types.ValidationResult import ValidationResult
+from phoenix.server.types import DbSessionFactory
 from phoenix.trace.dsl import SpanFilter
 
 DEFAULT_PAGE_SIZE = 30
@@ -262,6 +267,16 @@ class Project(Node):
         filter_condition: Optional[str] = UNSET,
         orphan_span_as_root_span: Optional[bool] = True,
     ) -> Connection[Span]:
+        if root_spans_only and not filter_condition and sort and sort.col is SpanColumn.startTime:
+            return await _paginate_span_by_trace_start_time(
+                db=info.context.db,
+                project_rowid=self.project_rowid,
+                time_range=time_range,
+                first=first,
+                after=after,
+                sort=sort,
+                orphan_span_as_root_span=orphan_span_as_root_span,
+            )
         stmt = (
             select(models.Span.id)
             .select_from(models.Span)
@@ -698,83 +713,603 @@ class Project(Node):
             )
         return updated_at
 
-    @strawberry.field(
-        description="Hourly span count for the project.",
-    )  # type: ignore
+    @strawberry.field
     async def span_count_time_series(
         self,
         info: Info[Context, None],
-        time_range: Optional[TimeRange] = UNSET,
+        time_range: TimeRange,
+        time_bin_config: Optional[TimeBinConfig] = UNSET,
     ) -> SpanCountTimeSeries:
-        """Returns a time series of span counts grouped by hour for the project.
+        if time_range.start is None:
+            raise BadRequest("Start time is required")
 
-        This field provides hourly aggregated span counts, which can be useful for
-        visualizing span activity over time. The data points represent the number
-        of spans that started in each hour.
+        dialect = info.context.db.dialect
+        utc_offset_minutes = 0
+        field: Literal["minute", "hour", "day", "week", "month", "year"] = "hour"
+        if time_bin_config:
+            utc_offset_minutes = time_bin_config.utc_offset_minutes
+            if time_bin_config.scale is TimeBinScale.MINUTE:
+                field = "minute"
+            elif time_bin_config.scale is TimeBinScale.HOUR:
+                field = "hour"
+            elif time_bin_config.scale is TimeBinScale.DAY:
+                field = "day"
+            elif time_bin_config.scale is TimeBinScale.WEEK:
+                field = "week"
+            elif time_bin_config.scale is TimeBinScale.MONTH:
+                field = "month"
+            elif time_bin_config.scale is TimeBinScale.YEAR:
+                field = "year"
+        bucket = date_trunc(dialect, field, models.Span.start_time, utc_offset_minutes)
+        stmt = (
+            select(bucket, func.count(models.Span.id))
+            .join_from(models.Span, models.Trace)
+            .where(models.Trace.project_rowid == self.project_rowid)
+            .group_by(bucket)
+            .order_by(bucket)
+        )
+        if time_range.start:
+            stmt = stmt.where(time_range.start <= models.Span.start_time)
+        if time_range.end:
+            stmt = stmt.where(models.Span.start_time < time_range.end)
 
-        Args:
-            info: The GraphQL info object containing context information.
-            time_range: Optional time range to filter the spans. If provided, only
-                spans that started within this range will be counted.
+        data = {}
+        async with info.context.db() as session:
+            async for t, v in await session.stream(stmt):
+                timestamp = _as_datetime(t)
+                data[timestamp] = TimeSeriesDataPoint(timestamp=timestamp, value=v)
 
-        Returns:
-            A SpanCountTimeSeries object containing data points with timestamps
-            (rounded to the nearest hour) and corresponding span counts.
-
-        Notes:
-            - The timestamps are rounded down to the nearest hour.
-            - If a time range is provided, the start time is rounded down to the
-              nearest hour, and the end time is rounded up to the nearest hour.
-            - The SQL query is optimized for both PostgreSQL and SQLite databases.
-        """
-        # Determine the appropriate SQL function to truncate timestamps to hours
-        # based on the database dialect
-        if info.context.db.dialect is SupportedSQLDialect.POSTGRESQL:
-            # PostgreSQL uses date_trunc for timestamp truncation
-            hour = func.date_trunc("hour", models.Span.start_time)
-        elif info.context.db.dialect is SupportedSQLDialect.SQLITE:
-            # SQLite uses strftime for timestamp formatting
-            hour = func.strftime("%Y-%m-%dT%H:00:00.000+00:00", models.Span.start_time)
+        data_timestamps: list[datetime] = [data_point.timestamp for data_point in data.values()]
+        min_time = min([*data_timestamps, time_range.start])
+        max_time = max(
+            [
+                *data_timestamps,
+                *([time_range.end] if time_range.end else []),
+            ],
+            default=datetime.now(timezone.utc),
+        )
+        for timestamp in get_timestamp_range(
+            start_time=min_time,
+            end_time=max_time,
+            stride=field,
+            utc_offset_minutes=utc_offset_minutes,
+        ):
+            if timestamp not in data:
+                data[timestamp] = TimeSeriesDataPoint(timestamp=timestamp)
+        return SpanCountTimeSeries(data=sorted(data.values(), key=lambda x: x.timestamp))
+
+    @strawberry.field
+    async def trace_count_time_series(
+        self,
+        info: Info[Context, None],
+        time_range: TimeRange,
+        time_bin_config: Optional[TimeBinConfig] = UNSET,
+    ) -> TraceCountTimeSeries:
+        if time_range.start is None:
+            raise BadRequest("Start time is required")
+
+        dialect = info.context.db.dialect
+        utc_offset_minutes = 0
+        field: Literal["minute", "hour", "day", "week", "month", "year"] = "hour"
+        if time_bin_config:
+            utc_offset_minutes = time_bin_config.utc_offset_minutes
+            if time_bin_config.scale is TimeBinScale.MINUTE:
+                field = "minute"
+            elif time_bin_config.scale is TimeBinScale.HOUR:
+                field = "hour"
+            elif time_bin_config.scale is TimeBinScale.DAY:
+                field = "day"
+            elif time_bin_config.scale is TimeBinScale.WEEK:
+                field = "week"
+            elif time_bin_config.scale is TimeBinScale.MONTH:
+                field = "month"
+            elif time_bin_config.scale is TimeBinScale.YEAR:
+                field = "year"
+        bucket = date_trunc(dialect, field, models.Trace.start_time, utc_offset_minutes)
+        stmt = (
+            select(bucket, func.count(models.Trace.id))
+            .where(models.Trace.project_rowid == self.project_rowid)
+            .group_by(bucket)
+            .order_by(bucket)
+        )
+        if time_range:
+            if time_range.start:
+                stmt = stmt.where(time_range.start <= models.Trace.start_time)
+            if time_range.end:
+                stmt = stmt.where(models.Trace.start_time < time_range.end)
+        data = {}
+        async with info.context.db() as session:
+            async for t, v in await session.stream(stmt):
+                timestamp = _as_datetime(t)
+                data[timestamp] = TimeSeriesDataPoint(timestamp=timestamp, value=v)
+
+        data_timestamps: list[datetime] = [data_point.timestamp for data_point in data.values()]
+        min_time = min([*data_timestamps, time_range.start])
+        max_time = max(
+            [
+                *data_timestamps,
+                *([time_range.end] if time_range.end else []),
+            ],
+            default=datetime.now(timezone.utc),
+        )
+        for timestamp in get_timestamp_range(
+            start_time=min_time,
+            end_time=max_time,
+            stride=field,
+            utc_offset_minutes=utc_offset_minutes,
+        ):
+            if timestamp not in data:
+                data[timestamp] = TimeSeriesDataPoint(timestamp=timestamp)
+        return TraceCountTimeSeries(data=sorted(data.values(), key=lambda x: x.timestamp))
+
+    @strawberry.field
+    async def trace_count_by_status_time_series(
+        self,
+        info: Info[Context, None],
+        time_range: TimeRange,
+        time_bin_config: Optional[TimeBinConfig] = UNSET,
+    ) -> TraceCountByStatusTimeSeries:
+        if time_range.start is None:
+            raise BadRequest("Start time is required")
+
+        dialect = info.context.db.dialect
+        utc_offset_minutes = 0
+        field: Literal["minute", "hour", "day", "week", "month", "year"] = "hour"
+        if time_bin_config:
+            utc_offset_minutes = time_bin_config.utc_offset_minutes
+            if time_bin_config.scale is TimeBinScale.MINUTE:
+                field = "minute"
+            elif time_bin_config.scale is TimeBinScale.HOUR:
+                field = "hour"
+            elif time_bin_config.scale is TimeBinScale.DAY:
+                field = "day"
+            elif time_bin_config.scale is TimeBinScale.WEEK:
+                field = "week"
+            elif time_bin_config.scale is TimeBinScale.MONTH:
+                field = "month"
+            elif time_bin_config.scale is TimeBinScale.YEAR:
+                field = "year"
+        bucket = date_trunc(dialect, field, models.Trace.start_time, utc_offset_minutes)
+        trace_error_status_counts = (
+            select(
+                models.Span.trace_rowid,
+            )
+            .where(models.Span.parent_id.is_(None))
+            .group_by(models.Span.trace_rowid)
+            .having(func.max(models.Span.cumulative_error_count) > 0)
+        ).subquery()
+        stmt = (
+            select(
+                bucket,
+                func.count(models.Trace.id).label("total_count"),
+                func.coalesce(func.count(trace_error_status_counts.c.trace_rowid), 0).label(
+                    "error_count"
+                ),
+            )
+            .join_from(
+                models.Trace,
+                trace_error_status_counts,
+                onclause=trace_error_status_counts.c.trace_rowid == models.Trace.id,
+                isouter=True,
+            )
+            .where(models.Trace.project_rowid == self.project_rowid)
+            .group_by(bucket)
+            .order_by(bucket)
+        )
+        if time_range:
+            if time_range.start:
+                stmt = stmt.where(time_range.start <= models.Trace.start_time)
+            if time_range.end:
+                stmt = stmt.where(models.Trace.start_time < time_range.end)
+        data: dict[datetime, TraceCountByStatusTimeSeriesDataPoint] = {}
+        async with info.context.db() as session:
+            async for t, total_count, error_count in await session.stream(stmt):
+                timestamp = _as_datetime(t)
+                data[timestamp] = TraceCountByStatusTimeSeriesDataPoint(
+                    timestamp=timestamp,
+                    ok_count=total_count - error_count,
+                    error_count=error_count,
+                    total_count=total_count,
+                )
+
+        data_timestamps: list[datetime] = [data_point.timestamp for data_point in data.values()]
+        min_time = min([*data_timestamps, time_range.start])
+        max_time = max(
+            [
+                *data_timestamps,
+                *([time_range.end] if time_range.end else []),
+            ],
+            default=datetime.now(timezone.utc),
+        )
+        for timestamp in get_timestamp_range(
+            start_time=min_time,
+            end_time=max_time,
+            stride=field,
+            utc_offset_minutes=utc_offset_minutes,
+        ):
+            if timestamp not in data:
+                data[timestamp] = TraceCountByStatusTimeSeriesDataPoint(
+                    timestamp=timestamp,
+                    ok_count=0,
+                    error_count=0,
+                    total_count=0,
+                )
+        return TraceCountByStatusTimeSeries(data=sorted(data.values(), key=lambda x: x.timestamp))
+
+    @strawberry.field
+    async def trace_latency_ms_percentile_time_series(
+        self,
+        info: Info[Context, None],
+        time_range: TimeRange,
+        time_bin_config: Optional[TimeBinConfig] = UNSET,
+    ) -> TraceLatencyPercentileTimeSeries:
+        if time_range.start is None:
+            raise BadRequest("Start time is required")
+
+        dialect = info.context.db.dialect
+        utc_offset_minutes = 0
+        field: Literal["minute", "hour", "day", "week", "month", "year"] = "hour"
+        if time_bin_config:
+            utc_offset_minutes = time_bin_config.utc_offset_minutes
+            if time_bin_config.scale is TimeBinScale.MINUTE:
+                field = "minute"
+            elif time_bin_config.scale is TimeBinScale.HOUR:
+                field = "hour"
+            elif time_bin_config.scale is TimeBinScale.DAY:
+                field = "day"
+            elif time_bin_config.scale is TimeBinScale.WEEK:
+                field = "week"
+            elif time_bin_config.scale is TimeBinScale.MONTH:
+                field = "month"
+            elif time_bin_config.scale is TimeBinScale.YEAR:
+                field = "year"
+        bucket = date_trunc(dialect, field, models.Trace.start_time, utc_offset_minutes)
+
+        stmt = select(bucket).where(models.Trace.project_rowid == self.project_rowid)
+        if time_range.start:
+            stmt = stmt.where(time_range.start <= models.Trace.start_time)
+        if time_range.end:
+            stmt = stmt.where(models.Trace.start_time < time_range.end)
+
+        if dialect is SupportedSQLDialect.POSTGRESQL:
+            stmt = stmt.add_columns(
+                percentile_cont(0.50).within_group(models.Trace.latency_ms.asc()).label("p50"),
+                percentile_cont(0.75).within_group(models.Trace.latency_ms.asc()).label("p75"),
+                percentile_cont(0.90).within_group(models.Trace.latency_ms.asc()).label("p90"),
+                percentile_cont(0.95).within_group(models.Trace.latency_ms.asc()).label("p95"),
+                percentile_cont(0.99).within_group(models.Trace.latency_ms.asc()).label("p99"),
+                percentile_cont(0.999).within_group(models.Trace.latency_ms.asc()).label("p999"),
+                func.max(models.Trace.latency_ms).label("max"),
+            )
+        elif dialect is SupportedSQLDialect.SQLITE:
+            stmt = stmt.add_columns(
+                func.percentile(models.Trace.latency_ms, 50).label("p50"),
+                func.percentile(models.Trace.latency_ms, 75).label("p75"),
+                func.percentile(models.Trace.latency_ms, 90).label("p90"),
+                func.percentile(models.Trace.latency_ms, 95).label("p95"),
+                func.percentile(models.Trace.latency_ms, 99).label("p99"),
+                func.percentile(models.Trace.latency_ms, 99.9).label("p999"),
+                func.max(models.Trace.latency_ms).label("max"),
+            )
         else:
-            assert_never(info.context.db.dialect)
+            assert_never(dialect)
 
-        # Build the base query to count spans grouped by hour
+        stmt = stmt.group_by(bucket).order_by(bucket)
+
+        data: dict[datetime, TraceLatencyMsPercentileTimeSeriesDataPoint] = {}
+        async with info.context.db() as session:
+            async for (
+                bucket_time,
+                p50,
+                p75,
+                p90,
+                p95,
+                p99,
+                p999,
+                max_latency,
+            ) in await session.stream(stmt):
+                timestamp = _as_datetime(bucket_time)
+                data[timestamp] = TraceLatencyMsPercentileTimeSeriesDataPoint(
+                    timestamp=timestamp,
+                    p50=p50,
+                    p75=p75,
+                    p90=p90,
+                    p95=p95,
+                    p99=p99,
+                    p999=p999,
+                    max=max_latency,
+                )
+
+        data_timestamps: list[datetime] = [data_point.timestamp for data_point in data.values()]
+        min_time = min([*data_timestamps, time_range.start])
+        max_time = max(
+            [
+                *data_timestamps,
+                *([time_range.end] if time_range.end else []),
+            ],
+            default=datetime.now(timezone.utc),
+        )
+        for timestamp in get_timestamp_range(
+            start_time=min_time,
+            end_time=max_time,
+            stride=field,
+            utc_offset_minutes=utc_offset_minutes,
+        ):
+            if timestamp not in data:
+                data[timestamp] = TraceLatencyMsPercentileTimeSeriesDataPoint(timestamp=timestamp)
+        return TraceLatencyPercentileTimeSeries(
+            data=sorted(data.values(), key=lambda x: x.timestamp)
+        )
+
+    @strawberry.field
+    async def trace_token_count_time_series(
+        self,
+        info: Info[Context, None],
+        time_range: TimeRange,
+        time_bin_config: Optional[TimeBinConfig] = UNSET,
+    ) -> TraceTokenCountTimeSeries:
+        if time_range.start is None:
+            raise BadRequest("Start time is required")
+
+        dialect = info.context.db.dialect
+        utc_offset_minutes = 0
+        field: Literal["minute", "hour", "day", "week", "month", "year"] = "hour"
+        if time_bin_config:
+            utc_offset_minutes = time_bin_config.utc_offset_minutes
+            if time_bin_config.scale is TimeBinScale.MINUTE:
+                field = "minute"
+            elif time_bin_config.scale is TimeBinScale.HOUR:
+                field = "hour"
+            elif time_bin_config.scale is TimeBinScale.DAY:
+                field = "day"
+            elif time_bin_config.scale is TimeBinScale.WEEK:
+                field = "week"
+            elif time_bin_config.scale is TimeBinScale.MONTH:
+                field = "month"
+            elif time_bin_config.scale is TimeBinScale.YEAR:
+                field = "year"
+        bucket = date_trunc(dialect, field, models.Trace.start_time, utc_offset_minutes)
         stmt = (
-            select(hour, func.count())
-            .join(models.Trace)
+            select(
+                bucket,
+                func.sum(models.SpanCost.total_tokens),
+                func.sum(models.SpanCost.prompt_tokens),
+                func.sum(models.SpanCost.completion_tokens),
+            )
+            .join_from(
+                models.Trace,
+                models.SpanCost,
+                onclause=models.SpanCost.trace_rowid == models.Trace.id,
+            )
             .where(models.Trace.project_rowid == self.project_rowid)
-            .group_by(hour)
-            .order_by(hour)
+            .group_by(bucket)
+            .order_by(bucket)
         )
+        if time_range:
+            if time_range.start:
+                stmt = stmt.where(time_range.start <= models.Trace.start_time)
+            if time_range.end:
+                stmt = stmt.where(models.Trace.start_time < time_range.end)
+        data: dict[datetime, TraceTokenCountTimeSeriesDataPoint] = {}
+        async with info.context.db() as session:
+            async for (
+                t,
+                total_tokens,
+                prompt_tokens,
+                completion_tokens,
+            ) in await session.stream(stmt):
+                timestamp = _as_datetime(t)
+                data[timestamp] = TraceTokenCountTimeSeriesDataPoint(
+                    timestamp=timestamp,
+                    prompt_token_count=prompt_tokens,
+                    completion_token_count=completion_tokens,
+                    total_token_count=total_tokens,
+                )
+
+        data_timestamps: list[datetime] = [data_point.timestamp for data_point in data.values()]
+        min_time = min([*data_timestamps, time_range.start])
+        max_time = max(
+            [
+                *data_timestamps,
+                *([time_range.end] if time_range.end else []),
+            ],
+            default=datetime.now(timezone.utc),
+        )
+        for timestamp in get_timestamp_range(
+            start_time=min_time,
+            end_time=max_time,
+            stride=field,
+            utc_offset_minutes=utc_offset_minutes,
+        ):
+            if timestamp not in data:
+                data[timestamp] = TraceTokenCountTimeSeriesDataPoint(timestamp=timestamp)
+        return TraceTokenCountTimeSeries(data=sorted(data.values(), key=lambda x: x.timestamp))
+
+    @strawberry.field
+    async def trace_token_cost_time_series(
+        self,
+        info: Info[Context, None],
+        time_range: TimeRange,
+        time_bin_config: Optional[TimeBinConfig] = UNSET,
+    ) -> TraceTokenCostTimeSeries:
+        if time_range.start is None:
+            raise BadRequest("Start time is required")
 
-        # Apply time range filtering if provided
+        dialect = info.context.db.dialect
+        utc_offset_minutes = 0
+        field: Literal["minute", "hour", "day", "week", "month", "year"] = "hour"
+        if time_bin_config:
+            utc_offset_minutes = time_bin_config.utc_offset_minutes
+            if time_bin_config.scale is TimeBinScale.MINUTE:
+                field = "minute"
+            elif time_bin_config.scale is TimeBinScale.HOUR:
+                field = "hour"
+            elif time_bin_config.scale is TimeBinScale.DAY:
+                field = "day"
+            elif time_bin_config.scale is TimeBinScale.WEEK:
+                field = "week"
+            elif time_bin_config.scale is TimeBinScale.MONTH:
+                field = "month"
+            elif time_bin_config.scale is TimeBinScale.YEAR:
+                field = "year"
+        bucket = date_trunc(dialect, field, models.Trace.start_time, utc_offset_minutes)
+        stmt = (
+            select(
+                bucket,
+                func.sum(models.SpanCost.total_cost),
+                func.sum(models.SpanCost.prompt_cost),
+                func.sum(models.SpanCost.completion_cost),
+            )
+            .join_from(
+                models.Trace,
+                models.SpanCost,
+                onclause=models.SpanCost.trace_rowid == models.Trace.id,
+            )
+            .where(models.Trace.project_rowid == self.project_rowid)
+            .group_by(bucket)
+            .order_by(bucket)
+        )
         if time_range:
-            if t := time_range.start:
-                # Round down to nearest hour for the start time
-                start = t.replace(minute=0, second=0, microsecond=0)
-                stmt = stmt.where(start <= models.Span.start_time)
-            if t := time_range.end:
-                # Round up to nearest hour for the end time
-                # If the time is already at the start of an hour, use it as is
-                if t.minute == 0 and t.second == 0 and t.microsecond == 0:
-                    end = t
-                else:
-                    # Otherwise, round up to the next hour
-                    end = t.replace(minute=0, second=0, microsecond=0) + timedelta(hours=1)
-                stmt = stmt.where(models.Span.start_time < end)
+            if time_range.start:
+                stmt = stmt.where(time_range.start <= models.Trace.start_time)
+            if time_range.end:
+                stmt = stmt.where(models.Trace.start_time < time_range.end)
+        data: dict[datetime, TraceTokenCostTimeSeriesDataPoint] = {}
+        async with info.context.db() as session:
+            async for (
+                t,
+                total_cost,
+                prompt_cost,
+                completion_cost,
+            ) in await session.stream(stmt):
+                timestamp = _as_datetime(t)
+                data[timestamp] = TraceTokenCostTimeSeriesDataPoint(
+                    timestamp=timestamp,
+                    prompt_cost=prompt_cost,
+                    completion_cost=completion_cost,
+                    total_cost=total_cost,
+                )
 
-        # Execute the query and convert the results to a time series
+        data_timestamps: list[datetime] = [data_point.timestamp for data_point in data.values()]
+        min_time = min([*data_timestamps, time_range.start])
+        max_time = max(
+            [
+                *data_timestamps,
+                *([time_range.end] if time_range.end else []),
+            ],
+            default=datetime.now(timezone.utc),
+        )
+        for timestamp in get_timestamp_range(
+            start_time=min_time,
+            end_time=max_time,
+            stride=field,
+            utc_offset_minutes=utc_offset_minutes,
+        ):
+            if timestamp not in data:
+                data[timestamp] = TraceTokenCostTimeSeriesDataPoint(timestamp=timestamp)
+        return TraceTokenCostTimeSeries(data=sorted(data.values(), key=lambda x: x.timestamp))
+
+    @strawberry.field
+    async def span_annotation_score_time_series(
+        self,
+        info: Info[Context, None],
+        time_range: TimeRange,
+        time_bin_config: Optional[TimeBinConfig] = UNSET,
+    ) -> SpanAnnotationScoreTimeSeries:
+        if time_range.start is None:
+            raise BadRequest("Start time is required")
+
+        dialect = info.context.db.dialect
+        utc_offset_minutes = 0
+        field: Literal["minute", "hour", "day", "week", "month", "year"] = "hour"
+        if time_bin_config:
+            utc_offset_minutes = time_bin_config.utc_offset_minutes
+            if time_bin_config.scale is TimeBinScale.MINUTE:
+                field = "minute"
+            elif time_bin_config.scale is TimeBinScale.HOUR:
+                field = "hour"
+            elif time_bin_config.scale is TimeBinScale.DAY:
+                field = "day"
+            elif time_bin_config.scale is TimeBinScale.WEEK:
+                field = "week"
+            elif time_bin_config.scale is TimeBinScale.MONTH:
+                field = "month"
+            elif time_bin_config.scale is TimeBinScale.YEAR:
+                field = "year"
+        bucket = date_trunc(dialect, field, models.Trace.start_time, utc_offset_minutes)
+        stmt = (
+            select(
+                bucket,
+                models.SpanAnnotation.name,
+                func.avg(models.SpanAnnotation.score).label("average_score"),
+            )
+            .join_from(
+                models.SpanAnnotation,
+                models.Span,
+                onclause=models.SpanAnnotation.span_rowid == models.Span.id,
+            )
+            .join_from(
+                models.Span,
+                models.Trace,
+                onclause=models.Span.trace_rowid == models.Trace.id,
+            )
+            .where(models.Trace.project_rowid == self.project_rowid)
+            .group_by(bucket, models.SpanAnnotation.name)
+            .order_by(bucket)
+        )
+        if time_range:
+            if time_range.start:
+                stmt = stmt.where(time_range.start <= models.Trace.start_time)
+            if time_range.end:
+                stmt = stmt.where(models.Trace.start_time < time_range.end)
+        scores: dict[datetime, dict[str, float]] = {}
+        unique_names: set[str] = set()
         async with info.context.db() as session:
-            data = await session.stream(stmt)
-            return SpanCountTimeSeries(
-                data=[
-                    TimeSeriesDataPoint(
-                        timestamp=_as_datetime(t),
-                        value=v,
-                    )
-                    async for t, v in data
-                ]
+            async for (
+                t,
+                name,
+                average_score,
+            ) in await session.stream(stmt):
+                timestamp = _as_datetime(t)
+                if timestamp not in scores:
+                    scores[timestamp] = {}
+                scores[timestamp][name] = average_score
+                unique_names.add(name)
+
+        score_timestamps: list[datetime] = [timestamp for timestamp in scores]
+        min_time = min([*score_timestamps, time_range.start])
+        max_time = max(
+            [
+                *score_timestamps,
+                *([time_range.end] if time_range.end else []),
+            ],
+            default=datetime.now(timezone.utc),
+        )
+        data: dict[datetime, SpanAnnotationScoreTimeSeriesDataPoint] = {
+            timestamp: SpanAnnotationScoreTimeSeriesDataPoint(
+                timestamp=timestamp,
+                scores_with_labels=[
+                    SpanAnnotationScoreWithLabel(label=label, score=scores[timestamp][label])
+                    for label in scores[timestamp]
+                ],
             )
+            for timestamp in score_timestamps
+        }
+        for timestamp in get_timestamp_range(
+            start_time=min_time,
+            end_time=max_time,
+            stride=field,
+            utc_offset_minutes=utc_offset_minutes,
+        ):
+            if timestamp not in data:
+                data[timestamp] = SpanAnnotationScoreTimeSeriesDataPoint(
+                    timestamp=timestamp,
+                    scores_with_labels=[],
+                )
+        return SpanAnnotationScoreTimeSeries(
+            data=sorted(data.values(), key=lambda x: x.timestamp),
+            names=sorted(list(unique_names)),
+        )
 
 
 @strawberry.type
@@ -782,6 +1317,85 @@ class SpanCountTimeSeries(TimeSeries):
     """A time series of span count"""
 
 
+@strawberry.type
+class TraceCountTimeSeries(TimeSeries):
+    """A time series of trace count"""
+
+
+@strawberry.type
+class TraceCountByStatusTimeSeriesDataPoint:
+    timestamp: datetime
+    ok_count: int
+    error_count: int
+    total_count: int
+
+
+@strawberry.type
+class TraceCountByStatusTimeSeries:
+    data: list[TraceCountByStatusTimeSeriesDataPoint]
+
+
+@strawberry.type
+class TraceLatencyMsPercentileTimeSeriesDataPoint:
+    timestamp: datetime
+    p50: Optional[float] = None
+    p75: Optional[float] = None
+    p90: Optional[float] = None
+    p95: Optional[float] = None
+    p99: Optional[float] = None
+    p999: Optional[float] = None
+    max: Optional[float] = None
+
+
+@strawberry.type
+class TraceLatencyPercentileTimeSeries:
+    data: list[TraceLatencyMsPercentileTimeSeriesDataPoint]
+
+
+@strawberry.type
+class TraceTokenCountTimeSeriesDataPoint:
+    timestamp: datetime
+    prompt_token_count: Optional[float] = None
+    completion_token_count: Optional[float] = None
+    total_token_count: Optional[float] = None
+
+
+@strawberry.type
+class TraceTokenCountTimeSeries:
+    data: list[TraceTokenCountTimeSeriesDataPoint]
+
+
+@strawberry.type
+class TraceTokenCostTimeSeriesDataPoint:
+    timestamp: datetime
+    prompt_cost: Optional[float] = None
+    completion_cost: Optional[float] = None
+    total_cost: Optional[float] = None
+
+
+@strawberry.type
+class TraceTokenCostTimeSeries:
+    data: list[TraceTokenCostTimeSeriesDataPoint]
+
+
+@strawberry.type
+class SpanAnnotationScoreWithLabel:
+    label: str
+    score: float
+
+
+@strawberry.type
+class SpanAnnotationScoreTimeSeriesDataPoint:
+    timestamp: datetime
+    scores_with_labels: list[SpanAnnotationScoreWithLabel]
+
+
+@strawberry.type
+class SpanAnnotationScoreTimeSeries:
+    data: list[SpanAnnotationScoreTimeSeriesDataPoint]
+    names: list[str]
+
+
 INPUT_VALUE = SpanAttributes.INPUT_VALUE.split(".")
 OUTPUT_VALUE = SpanAttributes.OUTPUT_VALUE.split(".")
 
@@ -790,5 +1404,213 @@ def _as_datetime(value: Any) -> datetime:
     if isinstance(value, datetime):
         return value
     if isinstance(value, str):
-        return datetime.fromisoformat(value)
+        return cast(datetime, normalize_datetime(datetime.fromisoformat(value), timezone.utc))
     raise ValueError(f"Cannot convert {value} to datetime")
+
+
+async def _paginate_span_by_trace_start_time(
+    db: DbSessionFactory,
+    project_rowid: int,
+    time_range: Optional[TimeRange] = None,
+    first: Optional[int] = DEFAULT_PAGE_SIZE,
+    after: Optional[CursorString] = None,
+    sort: SpanSort = SpanSort(col=SpanColumn.startTime, dir=SortDir.desc),
+    orphan_span_as_root_span: Optional[bool] = True,
+    retries: int = 3,
+) -> Connection[Span]:
+    """Return one representative root span per trace, ordered by trace start time.
+
+    **Note**: Despite the function name, cursors are based on trace rowids, not span rowids.
+    This is because we paginate by traces (one span per trace), not individual spans.
+
+    **Important**: The edges list can be empty while has_next_page=True. This happens
+    when traces exist but have no matching root spans. Pagination continues because there
+    may be more traces ahead with spans.
+
+    Args:
+        db: Database session factory.
+        project_rowid: Project ID to query spans from.
+        time_range: Optional time range filter on trace start times.
+        first: Maximum number of edges to return (default: DEFAULT_PAGE_SIZE).
+        after: Cursor for pagination (points to trace position, not span).
+        sort: Sort by trace start time (asc/desc only).
+        orphan_span_as_root_span: Whether to include orphan spans as root spans.
+            True: spans with parent_id=NULL OR pointing to non-existent spans.
+            False: only spans with parent_id=NULL.
+        retries: Maximum number of retry attempts when insufficient edges are found.
+            When traces exist but lack root spans, the function retries pagination
+            to find traces with spans. Set to 0 to disable retries.
+
+    Returns:
+        Connection[Span] with:
+        - edges: At most one Edge per trace (may be empty list).
+        - page_info: Pagination info based on trace positions.
+
+    Key Points:
+        - Traces without root spans produce NO edges
+        - Spans ordered by trace start time, not span start time
+        - Cursors track trace positions for efficient large-scale pagination
+    """
+    # Build base trace query ordered by start time
+    traces = select(
+        models.Trace.id,
+        models.Trace.start_time,
+    ).where(models.Trace.project_rowid == project_rowid)
+    if sort.dir is SortDir.desc:
+        traces = traces.order_by(
+            models.Trace.start_time.desc(),
+            models.Trace.id.desc(),
+        )
+    else:
+        traces = traces.order_by(
+            models.Trace.start_time.asc(),
+            models.Trace.id.asc(),
+        )
+
+    # Apply time range filters
+    if time_range:
+        if time_range.start:
+            traces = traces.where(time_range.start <= models.Trace.start_time)
+        if time_range.end:
+            traces = traces.where(models.Trace.start_time < time_range.end)
+
+    # Apply cursor pagination
+    if after:
+        cursor = Cursor.from_string(after)
+        assert cursor.sort_column
+        compare = operator.lt if sort.dir is SortDir.desc else operator.gt
+        traces = traces.where(
+            compare(
+                tuple_(models.Trace.start_time, models.Trace.id),
+                (cursor.sort_column.value, cursor.rowid),
+            )
+        )
+
+    # Limit for pagination
+    if first:
+        traces = traces.limit(
+            first + 1  # over-fetch by one to determine whether there's a next page
+        )
+    traces_cte = traces.cte()
+
+    # Define join condition for root spans
+    if orphan_span_as_root_span:
+        # Include both NULL parent_id and orphaned spans
+        parent_spans = select(models.Span.span_id).alias("parent_spans")
+        onclause = and_(
+            models.Span.trace_rowid == traces_cte.c.id,
+            or_(
+                models.Span.parent_id.is_(None),
+                ~exists().where(models.Span.parent_id == parent_spans.c.span_id),
+            ),
+        )
+    else:
+        # Only spans with no parent (parent_id is NULL, excludes orphaned spans)
+        onclause = and_(
+            models.Span.trace_rowid == traces_cte.c.id,
+            models.Span.parent_id.is_(None),
+        )
+
+    # Join traces with root spans (left join allows traces without spans)
+    stmt = select(
+        traces_cte.c.id,
+        traces_cte.c.start_time,
+        models.Span.id,
+    ).join_from(
+        traces_cte,
+        models.Span,
+        onclause=onclause,
+        isouter=True,
+    )
+
+    # Order by trace time, then pick earliest span per trace
+    if sort.dir is SortDir.desc:
+        stmt = stmt.order_by(
+            traces_cte.c.start_time.desc(),
+            traces_cte.c.id.desc(),
+            models.Span.start_time.asc(),  # earliest span
+            models.Span.id.desc(),
+        )
+    else:
+        stmt = stmt.order_by(
+            traces_cte.c.start_time.asc(),
+            traces_cte.c.id.asc(),
+            models.Span.start_time.asc(),  # earliest span
+            models.Span.id.desc(),
+        )
+
+    # Use DISTINCT for PostgreSQL, manual grouping for SQLite
+    if db.dialect is SupportedSQLDialect.POSTGRESQL:
+        stmt = stmt.distinct(traces_cte.c.start_time, traces_cte.c.id)
+    elif db.dialect is SupportedSQLDialect.SQLITE:
+        # too complicated for SQLite, so we rely on groupby() below
+        pass
+    else:
+        assert_never(db.dialect)
+
+    # Process results and build edges
+    edges: list[Edge[Span]] = []
+    start_cursor: Optional[str] = None
+    end_cursor: Optional[str] = None
+    async with db() as session:
+        records = groupby(await session.stream(stmt), key=lambda record: record[:2])
+        async for (trace_rowid, trace_start_time), group in islice(records, first):
+            cursor = Cursor(
+                rowid=trace_rowid,
+                sort_column=CursorSortColumn(
+                    type=CursorSortColumnDataType.DATETIME,
+                    value=trace_start_time,
+                ),
+            )
+            if start_cursor is None:
+                start_cursor = str(cursor)
+            end_cursor = str(cursor)
+            first_record = group[0]
+            # Only create edge if trace has a root span
+            if (span_rowid := first_record[2]) is not None:
+                edges.append(Edge(node=Span(span_rowid=span_rowid), cursor=str(cursor)))
+        has_next_page = True
+        try:
+            await records.__anext__()
+        except StopAsyncIteration:
+            has_next_page = False
+
+    # Retry if we need more edges and more traces exist
+    if first and len(edges) < first and has_next_page:
+        while retries and (num_needed := first - len(edges)) and has_next_page:
+            retries -= 1
+            batch_size = max(first, 1000)
+            more = await _paginate_span_by_trace_start_time(
+                db=db,
+                project_rowid=project_rowid,
+                time_range=time_range,
+                first=batch_size,
+                after=end_cursor,
+                sort=sort,
+                orphan_span_as_root_span=orphan_span_as_root_span,
+                retries=0,
+            )
+            edges.extend(more.edges[:num_needed])
+            start_cursor = start_cursor or more.page_info.start_cursor
+            end_cursor = more.page_info.end_cursor if len(edges) < first else edges[-1].cursor
+            has_next_page = len(more.edges) > num_needed or more.page_info.has_next_page
+
+    return Connection(
+        edges=edges,
+        page_info=PageInfo(
+            start_cursor=start_cursor,
+            end_cursor=end_cursor,
+            has_previous_page=False,
+            has_next_page=has_next_page,
+        ),
+    )
+
+
+def to_gql_project(project: models.Project) -> Project:
+    """
+    Converts an ORM project to a GraphQL project.
+    """
+    return Project(
+        project_rowid=project.id,
+        db_project=project,
+    )