Is there an existing issue for the same bug?
Branch Name
Commit ID
Other Environment Information
- Hardware parameters: N/A（逻辑 bug，与机器无关）
- OS type: Linux / Darwin
- Others: 单机或 multi-CN 均可复现；触发条件是 Window 输入行数 > AggBatchSize(8192)
Actual Behavior
执行下面这条带 ROW_NUMBER + SUM OVER 的查询时报错：
Error: -16 ... [ERROR] [execute_sql.py:51, _executesql_impl]
(20101, 'internal error: the Window operator currently does not support sending split result of window function.')
Failing SQL（线上/回归实际触发）
WITH w AS (
WITH order_dim AS (
SELECT vbeln, MAX(zsal_rep) AS zsal_rep, MAX(vkgrp) AS vkgrp, MAX(vkbur) AS vkbur, MAX(zsal_det) AS zsal_det
FROM dwd_dcp.dwd_s4_ztbw_qcbb GROUP BY vbeln
),
filtered_acdoca AS (
SELECT a.rbukrs, a.belnr, a.gjahr, a.zuonr, a.kunnr, a.wsl, a.prctr, a.vkbur_pa
FROM dwd_dcp.dwd_s4_acdoca a
WHERE a.drcrk = 'H' AND (a.racct LIKE '1122%' OR a.racct LIKE '2205%')
AND a.blart IN ('DZ','DW','DM') AND (a.xtruerev != 'X' OR a.xtruerev IS NULL)
AND a.gjahr IN ('2022', '2023', '2024', '2025')
AND a.rbukrs IN ('1000', '1100', '1200', '1300', '2000', '2100', '2200', '3000')
)
SELECT
a.rbukrs, a.belnr, a.gjahr, a.zuonr, a.kunnr, a.wsl,
COALESCE(c.vkgrp, od.vkgrp) AS vkgrp,
COALESCE(tgrp.bezei, od.zsal_rep) AS sales_group_desc,
e.cbukrs, e.cbuktx,
COALESCE(f1.cprctr, f2.cprctr) AS cprctr,
g.cpatnr_txtlg, NULLIF(h.comp_head, '/') AS comp_head,
m_code.dept_id, m_desc.dept_name,
ROW_NUMBER() OVER (PARTITION BY a.rbukrs, a.kunnr ORDER BY a.wsl DESC) AS rn,
SUM(a.wsl) OVER (PARTITION BY a.rbukrs) AS bukrs_wsl_total
FROM filtered_acdoca a
LEFT JOIN order_dim od ON a.zuonr = od.vbeln
LEFT JOIN dwd_dcp.dwd_s4_vbak c ON a.zuonr = c.vbeln
LEFT JOIN dwd_dcp.dwd_s4_tvgrt tgrp ON c.vkgrp = tgrp.vkgrp
LEFT JOIN dwd_dcp.dwd_s4_tvkbt tbur ON c.vkbur = tbur.vkbur
LEFT JOIN dwd_dcp.dwd_bw_ztbpc002_com e ON a.rbukrs = e.sbukrs
LEFT JOIN dwd_dcp.dwd_bw_ztbpc002_prc f1
ON f1.sysid = 'CN1' AND f1.bukrs = a.rbukrs AND f1.sprctr = a.prctr AND f1.bukrs != ''
LEFT JOIN dwd_dcp.dwd_bw_ztbpc002_prc f2
ON f2.sysid = 'CN1' AND (f2.bukrs = '' OR f2.bukrs IS NULL) AND f2.sprctr = a.prctr
LEFT JOIN dws_dcp.dws_bpqx g ON g.sysid = 'CN1' AND g.customer = a.kunnr
LEFT JOIN dwd_dcp.dwd_s4_bp001 h ON a.kunnr = h.partner
LEFT JOIN staging_db.sales_office_mapping m_code
ON c.vkbur = m_code.sales_office_code
LEFT JOIN staging_db.sales_office_mapping m_desc
ON tbur.bezei = m_desc.sales_office_desc
)
SELECT * FROM w
WHERE 1 = 1
LIMIT 500;触发点主要是 JOIN 后的两个窗口函数（尤其 SUM(a.wsl) OVER (...)），当进入 Window 的行数 > AggBatchSize(8192) 时 Flush() 返回多段 vector，Window 直接报错。
Expected Behavior
Window 算子应能正确处理 Flush() 返回的多段 result vector（与 Group 算子已支持的 split result 行为一致），查询成功返回结果。
Steps to Reproduce
Minimal repro（不依赖业务表）
DROP TABLE IF EXISTS t_win_split;
CREATE TABLE t_win_split (g INT, v DECIMAL(20,2));
-- 插入 >8192 行，迫使 SUM OVER 的 agg state 分块
INSERT INTO t_win_split
SELECT
(result % 8) AS g,
CAST(result AS DECIMAL(20,2)) AS v
FROM generate_series(1, 10000) g(result);
-- 触发：Window 对整表/大 batch 做 SUM OVER，Flush 返回 >1 个 vector
SELECT
g, v,
SUM(v) OVER (PARTITION BY g) AS part_sum,
ROW_NUMBER() OVER (PARTITION BY g ORDER BY v DESC) AS rn
FROM t_win_split
LIMIT 100;
若环境没有 generate_series，可用任意方式插入 ≥9000 行后再跑同一条 SELECT。
（上面 Failing SQL 即业务侧完整复现；若无金盘表，用本节 Minimal repro 即可。）
Additional information
Root cause（代码位置）
pkg/sql/colexec/window/window.go 在 Flush() 后只接受单个 vector：
vecs, err := ctr.batAggs[idx].Flush()
...
if len(vecs) > 1 {
return moerr.NewInternalErrorNoCtx(
"the Window operator currently does not support sending split result of window function.")
}
-
聚合执行器按 AggBatchSize = 8192 分块（pkg/sql/colexec/aggexec/aggState.go）。
SUM / decimal SUM fast path（sum_decimal_fast.go / sumavg2.go）的 Flush() 返回 len(exec.state) 个 vector。当 group/行数 >8192 时 len(vecs) > 1。
-
对比：Group 已能消费多段 Flush 结果（把各段挂到对应 output batch），Window 尚未对齐。
Suggested fix
- Window：对
Flush() 多段结果做 UnionBatch 合并，或按 chunk 拆成多个 output batch（与 Group 对齐）。
- 补充回归：
SUM/AVG/... OVER + 输入行数 >8192（含 PARTITION BY / 无 PARTITION BY 的 receiveAll 路径）。
Workaround
将窗口改写为 JOIN + GROUP BY（例如 SUM OVER (PARTITION BY g) → 先 GROUP BY g 再 join 回明细），可绕过 Window 算子。
Is there an existing issue for the same bug?
Branch Name
Commit ID
Other Environment Information
Actual Behavior
执行下面这条带
ROW_NUMBER+SUM OVER的查询时报错：Failing SQL（线上/回归实际触发）
触发点主要是 JOIN 后的两个窗口函数（尤其
SUM(a.wsl) OVER (...)），当进入 Window 的行数 >AggBatchSize(8192) 时Flush()返回多段 vector，Window 直接报错。Expected Behavior
Window 算子应能正确处理
Flush()返回的多段 result vector（与Group算子已支持的 split result 行为一致），查询成功返回结果。Steps to Reproduce
Minimal repro（不依赖业务表）
若环境没有
generate_series，可用任意方式插入 ≥9000 行后再跑同一条SELECT。（上面 Failing SQL 即业务侧完整复现；若无金盘表，用本节 Minimal repro 即可。）
Additional information
Root cause（代码位置）
pkg/sql/colexec/window/window.go在Flush()后只接受单个 vector：聚合执行器按
AggBatchSize = 8192分块（pkg/sql/colexec/aggexec/aggState.go）。SUM/ decimal SUM fast path（sum_decimal_fast.go/sumavg2.go）的Flush()返回len(exec.state)个 vector。当 group/行数 >8192 时len(vecs) > 1。对比：
Group已能消费多段 Flush 结果（把各段挂到对应 output batch），Window 尚未对齐。Suggested fix
Flush()多段结果做UnionBatch合并，或按 chunk 拆成多个 output batch（与 Group 对齐）。SUM/AVG/... OVER+ 输入行数>8192（含PARTITION BY/ 无PARTITION BY的receiveAll路径）。Workaround
将窗口改写为
JOIN+GROUP BY（例如SUM OVER (PARTITION BY g)→ 先GROUP BY g再 join 回明细），可绕过 Window 算子。