You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@flink.apache.org by "Purushottam Sinha (Jira)" <ji...@apache.org> on 2026/05/29 08:15:00 UTC

[jira] [Commented] (FLINK-39745) Web dashboard: surface watermark lag as a per-job timeline

    [ https://issues.apache.org/jira/browse/FLINK-39745?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18084284#comment-18084284 ] 

Purushottam Sinha commented on FLINK-39745:
-------------------------------------------

I will be picking it up when time allow. Dumping the idea for now.

> Web dashboard: surface watermark lag as a per-job timeline
> ----------------------------------------------------------
>
>                 Key: FLINK-39745
>                 URL: https://issues.apache.org/jira/browse/FLINK-39745
>             Project: Flink
>          Issue Type: New Feature
>          Components: Runtime / Web Frontend
>            Reporter: Purushottam Sinha
>            Priority: Major
>
> Problem
> The current per-operator Watermarks drawer (flink-runtime-web/web-dashboard/src/app/pages/job/overview/watermarks/job-overview-drawer-watermarks.component.ts) shows only the latest raw epoch-millis watermark per subtask. Operators must mentally subtract from wall clock and re-poll to see whether event-time is keeping up — the trend, which is the actual diagnostic signal, is invisible.
> Evidence
>   - Drawer table renders one row per subtask with the current currentInputWatermark only; previous values are discarded on each refresh.
>   - Peer systems (Dataflow "system lag", Materialize lag metric, Confluent Cloud Flink) ship a per-stage lag-over-time chart for this data.
>   - All required data is already exposed via GET /jobs/:id/vertices/:vid/watermarks — no backend change needed.
> Proposed fix
>   - New job-level "Watermarks" tab: multi-operator lag-over-time chart with a configurable alert band, plus per-vertex status cards (current lag, trend, healthy/tracked/elevated/idle) sorted worst-first.
>   - Inline lag sparkline next to each operator name on the overview operator list.
>   - Opt-in localStorage persistence of the rolling 30-min buffer so the chart survives a page refresh for the same job ID.
> Acceptance
>   - New tab plots one line per vertex with lag in seconds; idle vertices render as "Idle" rather than infinite lag.
>   - Toggle "Remember across refresh" survives Cmd+R; clears on job-ID change.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)