使用 Apache Beam 和 Dataflow 从按日期分区的动态 GCS 存储桶中读取

Question

Right now I'm sending failed records from an API call to a GCS bucket thats partitioned by day

ex:

gs://path/to/file/2022/07/01

gs://path/to/file/2022/07/02

etc..

From there, I would like to schedule a batch job to retry these failed records the next day using Apache Beam and Dataflow. The issue is that the date at the end of the GCS path is only added when the initial template was uploaded to GCP and remains fixed at that date regardless of when you run the job similar to this and this . I'm using a ValueProvider but I cannot figure out a work around for this.

I've found that when I run the pipeline locally, everything works great, but in Dataflow, only the DoFns within the expand function actually get run.

Essentially what I'm doing here is:

1. getting the initial gcs path and appending current date minus 24 hours to the end in fileMatch

2. calling FileIO.readMatches() to convert each match() to a ReadableFile

3. calling MatchGCSFiles which contains the exact code below that uses the value providers to get the current date and append to the GCS path. I did this to override the existing path because this was the only way to get it to work as I thought I understood a DoFn cannot take in empty input (still learning Beam)

4. call FileIO.readMatches() again to convert the new match() to a new ReadableFile then call the API.

    String dateFormat = "yyyy/MM/dd"; ValueProvider<String> date = new ValueProvider<String>() { @Override public String get() { String currentDate = Instant.now().minus(86400000).toDateTime(DateTimeZone.UTC).toString(dateFormat); return currentDate; } @Override public boolean isAccessible() { return true; } }; ValueProvider<String> gcsPathWithDate = new ValueProvider<String>() { @Override public String get() { return String.format("%s/%s/*/*.json", gcsPathPrefix, date.get()); } @Override public boolean isAccessible() { return true; } }; fileMatch = FileIO.match().filepattern(gcsPathWithDate.get()); } PCollectionTuple mixedPColl = input .getPipeline() .apply("File match", fileMatch) .apply("applying read matches", FileIO.readMatches()) .apply("matching files", ParDo.of(new MatchGCSFiles())) .apply("applying read matches", FileIO.readMatches()). //problem here .apply("Read failed events from GCS", ParDo.of(new ReadFromGCS())) .apply(//call API)...

The problem is in the second FileIO.readMatches(), the return type does not match: reason: no instance(s) of type variable(s) exist so that PCollection conforms to PCollection

I've tried different work arounds for this but none seem to work.

Is there another/better way to dynamically add/replace the date in the GCS path? I'm still learning Beam so if I'm doing something wrong please let me know.

Thanks in advance.

Answer 1

It seems that there are a few improvements you can apply:

• Your pipeline always reads the file from yesterday. So you don't need a runtime parameter (and the usage of ValueProvider in your example is not right since it's not used as a runtime parameter provided by a pipeline option).

  • If you do need to set an arbitrary date, you'll have to follow the example to create a pipeline option.

   public interface MyOptions extends PipelineOptions { @Description("Date in yyyy-MM-dd") @Default.String("2022-01-01") ValueProvider<String> getDateValue(); void setDateValue(ValueProvider<String> value); }

  Then use it in your DoFns.

   ... @ProcessElement public void process(ProcessContext c) { MyOptions ops = c.getPipelineOptions().as(MyOptions.class); // Use it. ...(ops.getDateValue()) } ...

  When you start a job, you have to provide the pipeline option of that arbitrary date just like when you set any other PipelineOptions.

• You can always get yesterday's date by Get yesterday's date using Date .

   Instant now = Instant.now(); Instant yesterday = now.minus(1, ChronoUnit.DAYS);

  and then use it directly in your pipeline.