[BUG] Cancelling a PPL query sent by another plugin doesn't stop it, and its searches have no parent task

Đang mở
  • Có gì không ổn?
#5,911 0 bình luận 0 reaction 0 người được giao Xem trên GitHub

Maintainer thường phản hồi trong vòng 1 ngày

Chưa có ai nhận issue này.

Đánh giá

Độ khó
3/5
Thời gian dự kiến
1-2 ngày
Mức phù hợp với người mới
64/100
Loại issue
Lỗi
Độ rõ ràng
Đặc tả rõ ràng
Mức độ hoạt động
Sôi nổi
Công nghệ
java
Lĩnh vực
backend, distributed-systems

Hướng nghiên cứu

Start in plugin/src/main/java/org/opensearch/sql/plugin/transport/TransportPPLQueryAction.java at the task checks in doExecute and submitAsync; compare them with OpenSearchQueryManager.setCancellableTask and OpenSearchNodeClient.applyParentTask. The issue proposes using the shared CancellableTask type for both paths. Done means PPL requests from another plugin link searches to their task, cancellation stops the query, and the async path accepts the task.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.

Mô tả

untriaged

What is the bug?
A PPL query that another plugin sends through the cluster:admin/opensearch/ppl transport action, using its own bundled copy of the SQL classes, is not linked to its task. Examples are Alerting PPL monitors (PPLUtils → PPLPluginInterface.executeQuery) and Skills PPL tools (PPLTool, LogPatternTool, and tools that use PPLExecuteHelper). As a result:

  • the searches the query runs have no parent task, and
  • POST _tasks/<ppl task id>/_cancel marks the task cancelled, but the query keeps running.

TransportPPLQueryAction.doExecute only sets the query's cancellable task when the task is this plugin's own PPLQueryTask class (L212-L214):

if (task instanceof PPLQueryTask pplQueryTask) {
  OpenSearchQueryManager.setCancellableTask(pplQueryTask);
}

Alerting and Skills bundle their own copy of the SQL classes (the opensearch-sql-thin jar) and don't extend opensearch-sql, so those classes load in the caller's classloader. Core creates the task through request.createTask(...) on the caller's TransportPPLQueryRequest, which returns the caller's copy of PPLQueryTask. That copy is a different class from this plugin's PPLQueryTask, so the instanceof check fails and the ThreadLocal is never set. OpenSearchNodeClient.applyParentTask (L254-L258) then sets no parent, and the engine's cancellation checks find no task. submitAsync makes the same check (L314-L316), so an async PPL request from another plugin fails with Async PPL queries require a cancellable PPLQueryTask. Async PPL is only on main so far, and neither Alerting nor Skills sends async requests today, so this part is latent.

How can one reproduce the bug?
Reproduced on a local single-node cluster with opensearch-sql and opensearch-alerting, on 3.10.0-SNAPSHOT (SQL at 24bfaaf) and on 3.9.0:

  1. Index 2M small docs into big. To make one query run about 20 s, set index.max_result_window: 10 on big, plus plugins.query.size_limit: 100000000 and plugins.calcite.pushdown.enabled: false. Raise plugins.alerting.ppl_query_max_execution_duration so the monitor waits for the whole query.
  2. Run source=big | where n < 0 | head 10 from an Alerting PPL monitor (POST _plugins/_alerting/monitors/_execute?dryrun=true with a ppl_monitor), and the same query through POST _plugins/_ppl.
  3. GET _tasks?detailed=true&actions=cluster:admin/opensearch/ppl,indices:data/read/search*. On main, the _plugins/_ppl query's searches have the PPL task as parent_task_id (300 of 300 sampled); the Alerting query's searches have none (0 of 279).
  4. POST _tasks/<ppl task id>/_cancel about 5 s in. The _plugins/_ppl query stops right away. The Alerting query's task reports "cancelled": true, but the query runs all ~200k searches to completion, unless plugins.ppl.query.timeout fires first. Same result on 3.9.0.

What is the expected behavior?
A PPL query should be linked to its task whichever plugin sends it. Its searches should be children of the PPL task, and cancelling the task should stop the query.

Suggested fix: check for the core type instead of this plugin's class. CancellableTask is loaded by the server classloader, so it is the same class in every plugin.

if (task instanceof CancellableTask cancellableTask) {
  OpenSearchQueryManager.setCancellableTask(cancellableTask);
}

Apply the same change in submitAsync; CancellableTask#cancel(String) is public. OpenSearchQueryManager.setCancellableTask already takes a CancellableTask. Features that need PPLQueryTask-specific state can stay limited to PPLQueryTask. For example, #5760 adds Query Insights fields and edits this same check.

What is your host/environment?

  • Version: main (24bfaaf). The cancellation problem also affects releases since #5254: 3.6.0+ with Skills and 3.7.0+ with Alerting (reproduced on 3.9.0). In releases, searches issued by the background scanner have no parent task whoever sends the query (#5647, fixed on main by #5888), so the missing-parent part is specific to main. The async failure is main-only too.
  • Plugins: opensearch-sql together with alerting and/or skills (all in the default distribution)

Do you have any additional context?
Found while reviewing opensearch-project/query-insights#699, which shows a running PPL query in Live Queries as one row by nesting its searches under the PPL task. PPL queries from other plugins don't get that grouping.

Ngôn ngữ chính
Java
Star
175
Fork
233
Merge trung bình
2 ngày 7 giờ
Pull request đã merge (30 ngày)
36

Chuẩn bị môi trường

Bắt đầu từ đâu

  1. Đọc hết issue, rồi đọc hướng dẫn đóng góp của dự án.
  2. Bình luận trên issue rằng bạn sẽ nhận — tránh hai người làm cùng một việc.
  3. Fork repository và làm thay đổi trên một nhánh.
  4. Mở pull request có tham chiếu số hiệu của issue.

Issue khác của opensearch-project/sql

Tất cả issue của opensearch-project/sql

Issue tương tự

Thêm issue về Java

Nhận issue mới trong hộp thư của bạn

Bản tóm tắt ngắn những issue GitHub phù hợp với người mới.