Working AtlasStreaming & event time

Watermark Tide

See what happens when an event arrives after its reporting window has closed.

TryLower Watermark lag and watch Late events. With reduced motion, advance with Step model.

Watch forThe watermark is a progress marker; when it reaches a window’s end, that window closes.

Jump to the model
Separate teaching model

The linked story describes related work; the scenarios and results here are synthetic.

Interactive model

Preparing the interactive experiment…

Skill practicedChoose how long to wait for late events before closing a window

Connection to my workClickTrail covers campaign-event ingestion, but not watermark rules; this model shows how a late event can miss a closed window.

Related project:ClickTrail project story →

Understand the result

A watermark is a progress marker: the newest event time seen minus a delay you choose. When it reaches or passes the end boundary of a time window, this model closes that window. A later event for a closed window goes to Late events and does not change its count. This is one teaching model, not a universal stream-processing rule.

Controls

Window type
Picks the windowing strategy: fixed non-overlapping buckets, fixed overlapping buckets, or gap-based sessions that merge when events land close together in time.
Watermark lag
Sets how far the progress marker trails the newest event time. A shorter delay closes windows sooner; production systems may separately retain late data or allow a grace period.
Window size
In Tumbling and Sliding modes, sets the width of each fixed window.
Gap timeout
In Session mode, sets the inactivity gap that closes a session. This replaces Window size when you select Session.
Reseed
Generates a new deterministic event sequence so you can inspect a different arrival order under the same settings.
Step model
When reduced motion is enabled, advances the simulation by about 2.25 seconds of event-arrival time while automatic animation stays paused.

What to watch for

  • With normal motion, watch the watermark tide line sweep across and seal window panes. With reduced motion, each Step model action advances about 2.25 seconds of event-arrival time; panes still seal when the line reaches their end boundary.
  • Lower Watermark lag and compare the synthetic stream: a shorter lag can seal windows earlier, sending more generated late events to this model’s Late-event side output.
  • A late event landing after its window fired never changes that window's already-emitted count — this model routes it to the Late-event side output.

Why it happens

Here the watermark is maximum event time observed minus configured lag, and a window fires when it reaches or passes the end. This slider models watermark lag; production systems may separately configure retained state or grace periods for late data. This model routes arrivals for sealed windows to its Late-event side output.

How this idea relates to the linked work

ClickTrail documents event ingestion for campaign analytics, but its project story does not describe watermark windows or a late-event policy. This separate teaching model helps me reason about those adjacent streaming concepts; it does not present them as ClickTrail capabilities or reproduce its ingestion path.

See the broader engineering brief