tasks/offline-compute/PrestoSQL/prestosql_012_en/gt/ground_truth.sql

INSERT INTO internal_platform_db.t_quality_check_gpu_instance_detail_prestosql_012
WITH base_trace AS (
    -- Step 1: Find trace_ids with check.aborted but NOT check.completed (HAVING inverted condition)
    SELECT trace_id
    FROM internal_platform_db.t_quality_check_span_prestosql_012
    WHERE databus_imp_date >= '2026060600'
      AND databus_imp_date <= '2026060800'
      AND span_name IN ('check.aborted', 'check.completed')
    GROUP BY trace_id
    HAVING COUNT(CASE WHEN span_name = 'check.aborted' THEN 1 END) > 0
       AND COUNT(CASE WHEN span_name = 'check.completed' THEN 1 END) = 0
),
combined_spans AS (
    -- Step 2: All original spans for ALL traces + synthetic check.completed for aborted-only traces
    SELECT
        trace_id,
        span_name,
        start_time,
        end_time,
        status_code,
        project_id,
        task_id,
        check_type
    FROM internal_platform_db.t_quality_check_span_prestosql_012
    WHERE databus_imp_date >= '2026060600'
      AND databus_imp_date <= '2026060800'
      AND span_name IN ('check.aborted', 'check.start', 'check.completed', 'data.scan', 'rule.evaluate', 'resource.allocate')

    UNION ALL

    SELECT
        trace_id,
        'check.completed' AS span_name,
        MAX(CASE WHEN span_name = 'check.start' THEN start_time END) AS start_time,
        MAX(CASE WHEN span_name = 'check.aborted' THEN end_time END) AS end_time,
        2 AS status_code,
        MAX(CASE WHEN span_name = 'check.aborted' THEN project_id END) AS project_id,
        MAX(CASE WHEN span_name = 'check.aborted' THEN task_id END) AS task_id,
        MAX(CASE WHEN span_name = 'check.aborted' THEN check_type END) AS check_type
    FROM internal_platform_db.t_quality_check_span_prestosql_012
    WHERE databus_imp_date >= '2026060600'
      AND databus_imp_date <= '2026060800'
      AND trace_id IN (SELECT trace_id FROM base_trace)
      AND span_name IN ('check.start', 'check.aborted')
    GROUP BY trace_id
),
base_data_time_fixed AS (
    -- Step 3: Fix check.aborted start_time -> earliest of code-related spans
    SELECT
        trace_id,
        span_name,
        CASE
            WHEN span_name = 'check.aborted'
            THEN MIN(CASE WHEN span_name IN ('data.scan', 'rule.evaluate', 'check.aborted') THEN start_time END)
                    OVER(PARTITION BY trace_id)
            ELSE start_time
        END AS start_time,
        end_time,
        status_code,
        project_id,
        task_id,
        check_type
    FROM combined_spans
),
base_data AS (
    -- Step 4: Convert timestamps, compute diff_days for cross-day splitting
    SELECT
        trace_id,
        span_name,
        start_time,
        end_time,
        status_code,
        project_id,
        task_id,
        check_type,
        from_unixtime(CAST(start_time AS BIGINT) / 1000) AS start_date,
        from_unixtime(CAST(end_time AS BIGINT) / 1000) AS end_date,
        date_diff('day', from_unixtime(CAST(start_time AS BIGINT) / 1000),
                 from_unixtime(CAST(end_time AS BIGINT) / 1000)) AS diff_days,
        CAST(CAST(start_time AS BIGINT) / 86400000 AS BIGINT) * 86400000 AS start_day_midnight_ms
    FROM base_data_time_fixed
),
pos_series AS (
    -- Step 5: Position series for cross-day expansion
    SELECT 0 AS pos UNION ALL SELECT 1 UNION ALL SELECT 2 UNION ALL SELECT 3
),
daily_split_spans AS (
    -- Step 6: INNER JOIN with pos_series, compute boundary timestamps per day
    SELECT
        trace_id,
        span_name,
        project_id,
        task_id,
        check_type,
        status_code,
        start_date,
        end_date,
        diff_days,
        date_add('day', s.pos, start_date) AS calc_date,
        CASE WHEN s.pos = 0 THEN start_time
             ELSE CAST(start_day_midnight_ms + s.pos * 86400000 AS VARCHAR)
        END AS split_start_time,
        CASE WHEN s.pos = diff_days THEN end_time
             ELSE CAST(start_day_midnight_ms + (s.pos + 1) * 86400000 - 1 AS VARCHAR)
        END AS split_end_time,
        start_time AS span_start_time,
        end_time AS span_end_time
    FROM base_data b
    INNER JOIN pos_series s ON s.pos <= b.diff_days
),
span_metrics AS (
    -- Step 7: GROUP BY trace_id, calc_date -> compute instance_run_time, code_run_time, status_code
    SELECT
        trace_id,
        calc_date,
        MAX(project_id) AS project_id,
        MAX(task_id) AS task_id,
        MAX(check_type) AS check_type,
        MAX(MAX(CASE WHEN span_name = 'check.completed' THEN status_code END))
            OVER(PARTITION BY trace_id) AS status_code,
        CAST(ROUND(
            (MAX(CASE WHEN span_name = 'check.completed' THEN CAST(split_end_time AS BIGINT) END) -
             MIN(CASE WHEN span_name = 'check.completed' THEN CAST(split_start_time AS BIGINT) END)) / 1000.0
        ) AS INT) AS instance_run_time,
        CAST(ROUND(
            (MAX(CASE WHEN span_name IN ('data.scan', 'rule.evaluate', 'check.aborted') THEN CAST(split_end_time AS BIGINT) END) -
             MIN(CASE WHEN span_name IN ('data.scan', 'rule.evaluate', 'check.aborted') THEN CAST(split_start_time AS BIGINT) END)) / 1000.0
        ) AS INT) AS code_run_time
    FROM daily_split_spans
    GROUP BY trace_id, calc_date
),
gpu_stats AS (
    -- Step 8: FLOOR 5-min window bucketing + pod mapping + ROW_NUMBER config dedup + gpu_hour
    SELECT
        ig.instance_uuid,
        lc.host_gpu_num * ig.total_run_time_m / 60.0 AS gpu_hour,
        ig.gpu_name,
        lc.resource_id,
        CASE WHEN lc.is_dedicated = 'true' THEN true ELSE false END AS is_dedicated
    FROM (
        -- Aggregate GPU run time per instance
        SELECT
            m.instance_uuid,
            MAX(p.gpu_name) AS gpu_name,
            SUM(p.run_time_m) AS total_run_time_m
        FROM (
            -- pod_run_time: count distinct minutes per (pod_name, time_5min)
            SELECT
                pod_name,
                MAX(gpu_name) AS gpu_name,
                time_5min,
                CAST(COUNT(DISTINCT minute_timestamp) AS DOUBLE) AS run_time_m
            FROM (
                -- pod_run_minutes: FLOOR bucketing dedup
                SELECT
                    pod_name,
                    gpu_name,
                    FLOOR(CAST(pkg_time AS BIGINT) / 60) * 60 AS minute_timestamp,
                    FLOOR(CAST(pkg_time AS BIGINT) / 300) * 300 AS time_5min
                FROM internal_platform_db.t_quality_gpu_monitor_prestosql_012
                WHERE metric IN ('k8s_container_vgpu_gpu_mem_usage', 'k8s_dcgm_fi_dev_fb_util')
                GROUP BY pod_name, gpu_name,
                         FLOOR(CAST(pkg_time AS BIGINT) / 60) * 60,
                         FLOOR(CAST(pkg_time AS BIGINT) / 300) * 300
            ) pod_run_minutes
            GROUP BY pod_name, time_5min
        ) p
        INNER JOIN internal_platform_db.dwd_quality_podname_prestosql_012 m
            ON p.pod_name = m.pod_name
        GROUP BY m.instance_uuid
    ) ig
    LEFT JOIN (
        -- latest_task_config: ROW_NUMBER dedup from UNION ALL of two config tables
        SELECT
            instance_uuid,
            host_gpu_num,
            resource_id,
            is_dedicated
        FROM (
            SELECT
                instance_uuid,
                host_gpu_num,
                resource_id,
                is_dedicated,
                ROW_NUMBER() OVER (PARTITION BY instance_uuid ORDER BY last_modify DESC) AS rn
            FROM (
                SELECT instance_uuid, host_gpu_num, last_modify, resource_id, is_dedicated
                FROM internal_platform_db.dwd_quality_task_config_prestosql_012
                WHERE databus_imp_date = (
                    SELECT MAX(databus_imp_date)
                    FROM internal_platform_db.dwd_quality_task_config_prestosql_012
                )
                UNION ALL
                SELECT instance_uuid, host_gpu_num, last_modify, resource_id, is_dedicated
                FROM internal_platform_db.quality_task_config_prestosql_012
                WHERE databus_imp_date = (
                    SELECT MAX(databus_imp_date)
                    FROM internal_platform_db.quality_task_config_prestosql_012
                )
            ) combined
        ) ranked
        WHERE rn = 1
    ) lc ON ig.instance_uuid = lc.instance_uuid
)
-- Final: JOIN span_metrics with gpu_stats
SELECT
    date_format(sm.calc_date, '%Y-%m-%d') AS p_date,
    sm.trace_id,
    sm.project_id,
    sm.task_id,
    sm.check_type,
    sm.status_code,
    sm.instance_run_time,
    sm.code_run_time,
    gs.gpu_hour,
    gs.resource_id,
    gs.is_dedicated
FROM span_metrics sm
LEFT JOIN gpu_stats gs ON sm.trace_id = gs.instance_uuid
;