你们tdengine的python-sdk包taos-ws-py库在持有查询不会释放GIL还不修复么？我都查出这个问题了。并且提交PR你们不处理的么？

13450519157 · 2026 年11 月 1 日 10:04

【TDengine 使用环境】
生产环境 /测试/ Poc/预生产环境

【TDengine 版本】

【操作系统以及版本】

【部署方式】容器/非容器部署

【集群节点数】

【集群副本数】

【描述业务影响】

【问题复现路径/shan】做过哪些操作出现的问题

【遇到的问题：问题现象及影响】

【资源配置】

【报错完整截图】（不要大段的粘贴报错代码，论坛直接看报错代码不直观）

github.com/taosdata/taos-connector-python

fix: release GIL during connection, query and fetch operations (#396)

main ← TDxhfeng:fix/release-gil-during-query-fetch

已打开 10:42AM - 27 Oct 26 UTC

TDxhfeng

+243 -49

Synchronous WebSocket connection, query and raw-block fetch operations currently… retain the CPython GIL while waiting in Rust. A slow query or a large result can therefore stop unrelated Python threads, even when the application uses separate pooled connections. This releases the GIL around the Rust I/O operations and bounds uninterrupted Python row conversion, while preserving the Python method arguments, result types and row order. ## Changes - Release the GIL during connection/cursor creation and connection/DB-API query and execute operations, including request-ID variants. - Release the GIL while fetching raw result blocks through DB-API cursors and direct result iterators. - Keep Python object creation under the GIL, with a scheduling opportunity between batches of 1,024 converted rows. - Add a delayed loopback handshake regression using a separate server process, so the test server does not share the client's GIL. - Add an opt-in local database test using a delayed TCP proxy for query/fetch responsiveness, cross-block fetching, fetchone/fetchmany, dictionary results and direct iteration. - Run the handshake regression in the existing Windows job and enable the local integration test in the existing Linux test job. ## Validation The Rust changes are identical to the release-based patch already built and tested on Windows x64 and Linux x86_64: https://github.com/TDxhfeng/taos-connector-python/actions/runs/40486167834 - Both wheel builds and delayed-handshake regressions passed. - Linux disposable TDengine integration: 2 tests passed, including the delayed query/fetch test with 12,000 rows. - Read-only comparisons against the official 0.7.0 wheels used identical historical queries for 1,000, 20,000 and 100,000 rows, twice per size and platform. All six result checksums, tuple types and column types matched on each platform. - Windows pooled-connection smoke test: 8 queries across two connections passed. - Local tests for this contribution, using that previously built patched extension: `python -m pytest taos-ws-py/tests/test_gil_release.py taos-ws-py/tests/test_utils.py -q` -> 7 passed, 1 skipped. The skipped case requires a disposable local TDengine instance. For the 100,000-row fetch comparisons, the largest continuous pause of another Python thread across the two runs was: | Platform | Official 0.7.0 | Patched | | --- | ---: | ---: | | Windows, CPython 3.10 | 3,014 ms | 17.7 ms | | Linux, CPython 3.12 | 2,631 ms | 6.8 ms | These are local thread-responsiveness measurements, not production latency percentiles. Network/server conditions were not controlled for a query-speed benchmark. An earlier Windows run also had a 179 ms outlier; the change does not guarantee a maximum scheduling latency. The full upstream suite, including cloud/enterprise-dependent cases and the build against current main's dependency lockfile, remains for the upstream CI. This PR is based on current main (`ecfa0b2`) and preserves its package version and dependency lockfile. Related to #371 and #372. This also addresses the DB-API cursor query/fetch paths used by connection pools, and adds regression coverage for them.

这是我提的pr

TDuser_OEzS_1621 · 2026 年11 月 2 日 01:38

收到，我反馈给我们研发处理。多谢！

qevolg · 2026 年11 月 2 日 04:48

已 Merge，感谢提交 PR！

13450519157 · 2026 年11 月 4 日 13:16

那这个版本什么时候发布呢我看pypi还是0.7.0