Partial aggregation stops skipping after early emits under memory pressure

未关闭
  • 哪里不对？
#27,350 0 条评论 0 个 reaction 已指派 1 人 在 GitHub 查看

维护者通常 1 天内回复

@comphead 已经在做这个了。

开始于 2026年11月12日。

  • #27351 来自 @comphead —— 未关闭

评估

这个 Issue 还没有评估数据。

描述

bug
Describe the bug

PartialHashAggregateStream gives the skip probe all rows seen so far, but only the groups currently in its table (hash_stream.rs#L401):

self.update_skip_aggregation_probe(input_rows, hash_table.building_group_count());

Under memory pressure the stream emits all groups early and continues with an empty table (hash_stream.rs#L306-L313). After that, num_groups counts only the groups since the last emit, while input_rows keeps growing (skip_partial.rs#L78-L91). The ratio stays below skip_partial_aggregation_probe_ratio_threshold, so the probe fires late or never, even for unique keys. The legacy GroupedHashAggregateStream has the same pattern (grouped_hash_stream.rs#L1399).

To Reproduce
SET datafusion.execution.target_partitions = 2;
SET datafusion.runtime.memory_limit = '4M';
EXPLAIN ANALYZE
SELECT (value * 1103515245 + 12345) % 2147483648 AS k, SUM(value)
FROM generate_series(1, 2000000)
GROUP BY k;

All 2M keys are unique. The affine map only makes the input unordered, so the hash path is used instead of the clustered one.

AggregateExec: mode=Partial on 51960dbad:

Memory limit early_emit_count skipped_aggregation_rows elapsed_compute
None 0 1.79 M 6–9 ms
8M 15–24 894 K 20 ms
4M 94–101 0 33 ms

The limited runs send the same 2M rows downstream, at about 4x the partial CPU. The fair pool (datafusion-cli -m 16M --mem-pool-type fair) behaves the same: 48 emits, 0 skipped.

Expected behavior

The probe should still skip after skip_partial_aggregation_probe_rows_threshold rows. Counting the groups already emitted would fix it:

self.update_skip_aggregation_probe(
    input_rows,
    self.emitted_groups + hash_table.building_group_count(),
);

This counts output rows, the same total reduction_factor reports, which is what the skip decision needs. Resetting input_rows on each emit would not work, because frequent emits would keep the probe below its rows threshold.

Additional context

Repeated keys are affected too. 20M rows over about 200K keys, 8 partitions and a 32 MB fair pool gave 590 emits, 0 skipped rows, and 360 ms instead of 70 ms of partial CPU.

Related: #26124 (proactive flushing would hit this on every flush), #22405, #24996.

Willingness to contribute

I plan to contribute a fix for this bug independently

主要语言
Rust
星标
9.6k
派生
2.6k
平均合并
3 天 4 小时
30 天内合并 PR
345

环境准备

在 Codespaces 中打开

在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

apache/datafusion 的其他 Issue

查看 apache/datafusion 的全部 Issue

相似的 Issue

更多 Rust Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。