Hacktoberfest 2026：维护者为十月标记出来的 issue，仍然开放、适合新手。 浏览 Hacktoberfest issue

Perf: add in-process timing instrumentation to replace py-spy dependency

未关闭
  • 哪里不对？
#140 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
46/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
停滞
技术栈
python
领域
performance

调研方向

Start by reading read_xarray_table and _iteration_callback in the reader, then inspect perf_tests/profile.sh and perf_tests/query_large_synthetic.py to understand the current profiling workflow. Choose and implement a timing approach that reports per-partition costs across the relevant execution stages. Confirm the measurements distinguish I/O, conversion, and DataFusion work without requiring py-spy or root access.

由索引模型根据 Issue 内容生成。

描述

Problem

The current profiling workflow in perf_tests/ requires py-spy, which needs sudo on macOS:

sudo ./profile.sh groupby_air_full.py

This makes profiling unavailable in many CI environments and for users without root access. Additionally, cProfile only sees the Python side of query execution — the entire DataFusion execution appears as a single opaque to_arrow_table() call (0.66s for 320 MB of data in our benchmarks), making it impossible to distinguish:

  • Time spent in xarray I/O (ds.isel(block), .values)
  • Time spent in numpy → Arrow conversion (iter_record_batches)
  • Time spent in DataFusion (aggregation, hash-grouping, row filtering)
  • Time spent waiting for GIL between partitions

Proposed solution

Add a lightweight per-partition timing mode to the _iteration_callback mechanism already in read_xarray_table:

# reader.py — extend make_stream to record timing
def make_stream(projected_names=None) -> pa.RecordBatchReader:
    if _timing_callback is not None:
        t0 = time.perf_counter()
    result = pa.RecordBatchReader.from_batches(...)
    if _timing_callback is not None:
        _timing_callback({"elapsed_s": time.perf_counter() - t0, "block": block})
    return result

Or alternatively: add a verbose=True flag to read_xarray_table that prints per-partition timings to stderr, giving insight into whether execution is I/O-bound or compute-bound without requiring external tools.

A lower-overhead option is to expose a PartitionTimingHook via LazyArrowStreamTable that fires before/after each factory call.

Current profiling workaround

For the query_large_synthetic.py benchmark (320 MB, 8 partitions):

Full scan AVG:    0.662s  (341 MB peak)   ← all in Rust's to_arrow_table()
Filter pushdown:  0.005s  (0.7 MB peak)   ← 7/8 partitions pruned
GROUP BY lat/lon: 0.839s  (341 MB peak)   ← +27% vs. AVG for hash aggregation

cProfile shows 246 Python calls for the full scan, but 0.660s of those are attributed to {method 'to_arrow_table' of 'datafusion.DataFrame' objects} — a Rust builtin that cProfile cannot introspect.

Related

  • _iteration_callback in XarrayRecordBatchReader and read_xarray_table — already used for testing
  • perf_tests/profile.sh — current py-spy based approach
  • perf_tests/query_large_synthetic.py — new benchmark script added alongside this issue
主要语言
Python
星标
142
派生
19
平均合并
1 天 9 小时
30 天内合并 PR
9

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

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。

xqlsystems/xarray-sql 的其他 Issue

查看 xqlsystems/xarray-sql 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。