Hancheng Deng created FLINK-40289:
-------------------------------------
Summary: Window Top-N examples are not reproducible in streaming
mode
Key: FLINK-40289
URL: https://issues.apache.org/jira/browse/FLINK-40289
Project: Flink
Issue Type: Improvement
Components: Documentation, Table SQL / API
Affects Versions: 2.3.0
Reporter: Hancheng Deng
h2. Description
The Window Top-N documentation shows six output rows for two 10-minute tumbling
windows while defining the following watermark:
{code:sql}
WATERMARK FOR bidtime AS bidtime - INTERVAL '1' SECOND
{code}
However, the maximum event time in the displayed input is {{2020-04-15
08:17:00.000}}. Therefore, in streaming mode, the watermark can advance only to
approximately {{2020-04-15 08:16:59.000}}.
This closes the {{[08:00, 08:10)}} window, but does not close the {{[08:10,
08:20)}} window. With an unbounded streaming source and only the displayed
records, only the first three output rows can be reproduced. The three rows
shown for the second window are not emitted until later records advance the
watermark past the window end.
The six displayed rows are valid as final results for bounded/batch input, or
when subsequent streaming records advance the watermark sufficiently. Neither
condition is explained in the example.
This issue is present in the Flink 1.20 LTS documentation, the Flink 2.3 stable
documentation, and the current documentation source.
h2. Affected documentation
* [Flink 1.20 Window
Top-N|https://nightlies.apache.org/flink/flink-docs-release-1.20/docs/dev/table/sql/queries/window-topn/]
* [Flink 2.3 Window
Top-N|https://nightlies.apache.org/flink/flink-docs-release-2.3/docs/sql/reference/queries/window-topn/]
* {{docs/content/docs/sql/reference/queries/window-topn.md}}
* {{docs/content.zh/docs/sql/reference/queries/window-topn.md}}
h2. Steps to reproduce
# Use an unbounded streaming source containing only the nine records displayed
in the documentation.
# Define {{WATERMARK FOR bidtime AS bidtime - INTERVAL '1' SECOND}}.
# Run the documented Window Top-N query.
# Observe that only the Top 3 rows for {{[08:00, 08:10)}} are emitted.
# No rows for {{[08:10, 08:20)}} are emitted because the watermark never
reaches the end of that window.
h2. Expected documentation improvement
Clarify that the displayed tables represent final results after both windows
have closed:
* In batch mode, bounded input completion closes the remaining windows.
* In streaming mode, subsequent events must advance the watermark past the
window end before Window Top-N emits the final result.
* With only the displayed records in an unbounded source, only the first window
result is emitted.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
