One post tagged with "performance"

In this article, we're taking a look on how we use SDLB's housekeeping features to keep our pipelines running efficiently.

Some DataObject contain housekeeping features of their own. Make sure you use them! For example, Delta Tables support commands like optimize and vacuum to optimize storage and delete no longer needed files.

But usually, those commands do not re-organize your partitions. This is where SDLBs housekeeping mode comes in.

The example is taken from a real world project we've implemented.

Context​

In this particular project we are collecting data from various reporting units and process it in batches. The reporting units use an Azure Function to upload JSON files to an Azure Data Lake Storage. From there, we pick them up for validation and processing. Reporting units can upload data anytime, but it is only processed a few times a day.

Once validated, we use Delta Lake tables in Databricks to process data through the layers of the Lakehouse.

Partitioning​

The Azure Function puts uploaded JSON files in a subfolder for each reporting unit. As such, JSON files are already neatly partitioned by reporting_unit:

uploadFolder/
  reporting_unit=rp01
    file1.json
    file2.json
    file3.json
  reporting_unit=rp02
    file1.json
  reporting_unit=rp03
    fileX.json

To read these JSON files, we can therefore use the following DataObject definition:

import_json {
    type = JsonFileDataObject
    path = uploadFolder/
    partitions = [reporting_unit]    
}

These files are then processed with a FileTransferAction into an output DataObject stage_json:

stage_json {
    type = FileTransferAction
    inputId = import_json
    outputId = stage_json
    executionMode = { type = FileIncrementalMoveMode }
    metadata.feed = stage_json
}

Each time we start to process uploaded data, we use the run_id to keep track of all batch jobs and version of files delivered. If you use a state path (see commandLine), your runs automatically generate a run_id to identify the run, and you can use it by extending your DataObject:

stage_json {
    type = JsonFileDataObject
    path = processedFolder
    partitions = [run_id,reporting_unit]
    schema = """reporting_unit string, run_id string, ...."""
}

Note how we just use run_id as part of the schema without any further declaration. Since we use the state path, SDLB uses a run_id internally, and if it's referenced as partition column in a DataObject, processed data get automatically assigned the id of the current run.

Drawback​

Let's take a look at the resulting partition layout of stage_json:

processedFolder/
  run_id=1/
    reporting_unit=rp01/
      file1.json
      file2.json
      file3.json
    reporting_unit=rp02/
      file1.json
    reporting_unit=rp03/
      fileX.json

This partition layout has many advantages in our case as we know exactly during which run a particular file was processed and which reporting unit uploaded it. In further stages we can clearly work with files that were processed in the current run and not touch any old run_ids.

For this use case, a few things are important to note:

• Some reporting units don't upload data for days. You end up with only a few reporting_unit partitions per run_id.
• File sizes are rather small (< 1 MiB), partition sizes end up very small too.
• If you use hourly runs and run 24/7, you end up with 168 partitions per week, plus sub-partitions for reporting units.
• Once files are correctly processed, we don't read the uploaded files anymore. We still keep them as raw files should we ever need to re-process them.

The drawback becomes apparent when you have actions working with all partitions, they will become very slow. Spark doesn't like a lot of small partitions.

To mitigate that, we use SDLB's Housekeeping Feature.

HousekeepingMode​

If you take a look at DataObject's parameters, you will see a housekeepingMode. There are two modes available:

• PartitionArchiveCompactionMode: to compact / archive partitions
• PartitionRetentionMode: to delete certain partitions completely

PartitionArchiveCompactionMode​

In this mode, you solve two tasks at once:

• You define how many smaller partitions are aggregated into one larger partition (archive)
• Rewrite all files in a partition to combine many small files into larger files (compact)

Archive​

In our example above, we stated that we don't want to alter any input files, so we won't use compaction. We want to keep them as is (raw data). But we do want to get rid of all the small partitions after a certain amount of time. For that, we extend stage_json to include the housekeepingMode with a archivePartitionExpression:

stage_json {
    type = JsonFileDataObject
    path = processedFolder
    partitions = [run_id,reporting_unit]
    schema = """reporting_unit string, run_id string, ...."""
    housekeepingMode = {
      type = PartitionArchiveCompactionMode
      archivePartitionExpression = "if( elements.run_id < (runId - 500), map('run_id', (cast(elements.run_id as integer) div 500) * 500, 'reporting_unit', elements.reporting_unit), elements)"
    }
}

This expression probably needs some explanation:
The Spark SQL expression works with attributes of PartitionExpressionData. In this case we use runId (the current runId) and elements (all partition values as map(string,string)). It needs to return a map(string,string) to define new partition values. In our case, it needs to define run_id and reporting_unit because these are the partitions defined in stage_json.

Let's take the expression apart:
if(elements.run_id < (runId - 500), ...
Only archive the partition if it's runId is older than 500 run_ids ago.

map('run_id', (cast(elements.run_id as integer) div 500) * 500, 'reporting_unit', elements.reporting_unit)
Creates the map with the new values for the partitions. The run_id is floored to the next 500 value, so as example, the new value of run_id 1984 will be 1500 (because integer 1984/500=3, 3*500=1500).
Remember that we need to return all partition values in the map, also the ones we don't want to alter. For reporting_unit we simply return the existing value elements.reporting_unit.

..., elements)
This is the else condition and simply returns the existing partition values if there is nothing to archive.

Compaction​

We don't want to compact files in our case. But from the documentation you can see that compaction works very similarly:
You also work with attributes from PartitionExpressionData but instead of new partition values, you return a boolean to indicate for each partition if it should be compacted or not.

PartitionRetentionMode​

Again, not used in our example as we never delete old files. But if you need to, you define a Spark SQL expression returning a boolean indicating if a partition should be retained or deleted.

stage_json {
    type = JsonFileDataObject
    path = processedFolder
    partitions = [run_id,reporting_unit]
    schema = """reporting_unit string, run_id string, ...."""
    housekeepingMode = {
      type = PartitionRetentionMode
      retentionCondition = "elements.run_id > (runId - 500)"
    }
}

Result​

In our example, we had performance gradually decreasing because Spark had to read more than 10'000 partitions and subpartitions. Just listing all available partitions, even if you only worked with the most recent one, took a few minutes and these operations added up.

With the housekeeping mode enabled, older partitions continuously get merged into larger partitions containing up to 500 runs. This brought the duration of list operations back to a few seconds.

The operations are fully automated, no manual intervention is required.