Streaming read and writing with aggregation

Guillermo-HR
New Contributor

‎04-29-2026 10:41 PM

Hi,

I have the following problem: on a medallion architecture on a bronze volume I get files every month containing the data for each sensor reading during the period 1 of month 00:00 to last day 23:00. I have a manual job that calls the python files to load and transform from landing -> bronze-> silver. Now I want to add the aggregated data to a gold table. I have the following code:

df_silver_swell_metrics = (

    spark.readStream

        .format("delta")

        .table(f"cor_project.silver.swell_metrics")

        .withWatermark("datetime", "0 second")

)

df_silver_swell_metrics_transformed = (

    df_silver_swell_metrics

    .groupBy(

        F.window("datetime", "1 day").alias("window"),

        "coast_name"

    ).agg(

...

            )

)

df_gold_wave_daily_summary = (

    df_silver_swell_metrics_transformed

    .select(

        ...

    )

)

(

    df_gold_wave_daily_summary.writeStream

        .format("delta")

        .outputMode("append")

        .trigger(availableNow=True)

        .option("checkpointLocation", f"/Volumes/cor_{ambiente}/gold/data/checkpoints/wave_daily_summary")

        .toTable(f"cor_project.gold.wave_daily_summary")

)

This works it doesn't repeat data that has already been processed  but it always misses the last day of the month. I have asked in other groups and they told me to use batch processing. Any thoughts on posible solutions?

Thanks for any comment 

0 Kudos

Saritha_S
Databricks Employee
Databricks Employee

‎05-12-2026 09:36 PM

Hi @Guillermo-HR 

Yes — batch is usually the right fix here.

What’s happening is that your query is using event-time window aggregation in Structured Streaming with append output mode. In that mode, Spark only emits a window after it is sure the window is closed according to the watermark. With monthly files and availableNow=True, the final day’s 1-day window often never gets finalized during that run, so it stays buffered and appears “missing”.

A few key points:

  • withWatermark("datetime", "0 second") does not mean “emit immediately”; it still depends on seeing later event-time data to advance the watermark past the end of the last window.
  • If your last record is on 2026-04-30 23:00, there may be no later event to push the watermark beyond the 2026-04-30 daily window.
  • append + windowed aggregation is therefore a bad fit for bounded monthly backfills/files like this.

Recommendation:

  • Use batch instead of streaming
  • If you want to use streaming, Use outputMode("complete")
0 Kudos