Skip to content

nathan-gs/eventhubs-reingest

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

15 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

EventHubs Reingestion (Capture or other sources)

Inputs

We require a Dataset with, following columns:

  • a ts as java.sql.Timestamp column
  • a body as Array[Byte]
  • a eh_partion as int
    • Either randomly using random_partition()
    • Or based on a field column_to_partition()

EventHubs Capture data

If you want to ingest everything inside a EventHubs Capture target, you can leave the spark.eventhubsreingest.query empty. It will look for all inputs with as format: avro or com.databricks.spark.avro.

As path you can specify the path to Azure Blob or Azure Datalake, Spark accepts wildcards, although they are unnecessary if you use {Namespace}/{EventHub}/year={Year}/month={Month}/day={Day}/hour={Hour}/minute={Minute}/capture-{Second}-{PartitionId} as the capture file name format (it's seen as a partitioned table).

The query being used for this is:

SELECT
  ehcapture_timestamp(`EnqueuedTimeUtc`) as `ts`,
  `Body` as `body`,
  random_partition() as `eh_partition`
FROM internal_capture_alias
ORDER BY
  year(`ts`), dayofyear(`ts`), hour(`ts`), `eh_partition`

Everything else is done automatically.

Query based inputs

By specifying spark.eventhubsreingest.query you can customize the query used to read the data. You can read from Hive & Spark SQL tables, if you have a centralized metastore. Additionaly, you can add temporary views by specifying spark.eventhubsreingest.inputs., each alias is queryable.

For example we want to read EventHubs capture data, but only data for a specific day, we could use following query:

SELECT 
    ehcapture_timestamp(EnqueuedTimeUtc) as `ts`, 
    Body as body,
    random_partition() as `eh_partition` 
    FROM capture WHERE year='2018' AND month='01' AND day='16'
    ORDER BY year(`ts`), dayofyear(`ts`), hour(`ts`), `eh_partition`

Do note the ehcapture_timestamp function, which translates EventHubs capture Date Time strings into the right format. The data itself is using the capture file name format from above.

Specifying the query and the rest of the inputs config now looks like this:

spark.eventhubsreingest.inputs.capture.path="wasbs://..."
spark.eventhubsreingest.inputs.capture.format=com.databricks.spark.avro

spark.eventhubsreingest.query="SELECT ehcapture_timestamp(EnqueuedTimeUtc) as ts, Body as body, random_partition() as `eh_partition` FROM capture WHERE year='2018' AND month='01' AND day='16' ORDER BY `year`, `month`, `hour`, `eh_partition`"

You have access to the full Spark SQL functionality, including joins, etc.

Partitioning based on a field within the message.

It's possible to use your own partitioning strategy, by looking at the contents of the message.

Suppose we have json messages, that look like this:

{"id":"arandomid1", "k":"d"}
{"id":"arandomid2", "k":"d"}
{"id":"arandomid1", "k":"d"}

We could use following query:

SELECT
  `ts`,
  `data` as `body`,
  column_to_partition(get_json_object(`data`, '$.id')) as `eh_partition`
FROM `testds_hash`
ORDER BY
  year(`ts`), dayofyear(`ts`), hour(`ts`), `eh_partition`

The column_to_partition udf will partition the messages based on the String.hashCode % NUMBER_OF_PARTITIONS.

Caching

By default the input dataset is cached using .cache, if you want to disable this, set:

spark.eventhubsreingest.cache=false

Output

The output is always a Azure EventHub, you can specify the credentials:

spark.eventhubsreingest.output.eh.ns=""
spark.eventhubsreingest.output.eh.name=""
spark.eventhubsreingest.output.eh.keyName=""
spark.eventhubsreingest.output.eh.keyValue=""

# set by default to 260 ms (because 1 throughput unit is 1mb/s and batch size is 256kb, with some margins).
spark.eventhubsreingest.output.eh.ms_to_wait_after_batch="260"

The application connects to EventHub to discover the partition layout, and divides the dataset in the exact same amount of partitions (random assignment). It sorts the data based on year, dayofmonth, hour, random partition.

The application will try to batch as many messages as possible, taking into account the limits of EventHub. Most importantly:

  • Max Event/Batch size of 256kb
  • Up to 1 MB per second of ingress events, see EventHubs FaQ, we accomplish this by assuming there are as many partitions as throughput units and waiting for 300ms after we've send each batch. Multiple partitions can be send in parallel though.

Executing on HDInsight using livy

The livy.sh script will upload and submit the jar file. The livy.sh script assumes you have build the source code from scratch (sbt assembly).

export STORAGE_ACCOUNT_NAME="YOUR STORAGE ACC"
export STORAGE_CONTAINER_NAME="YOUR STORAGE CONTAINER"

export CLUSTER_URL="https://CLUSTERNAME.azurehdinsight.net"
export CLUSTER_USER="admin"
export CLUSTER_PASSWORD='A PASSWORD'

APP_CONF="\"spark.eventhubsreingest.inputs.capture.path\":\"wasbs://CAPTURE_CONTAINER@CAPTURE_STORAGE_ACC.blob.core.windows.net/PATH_TO_FILES/\""
APP_CONF="$APP_CONF,\"spark.eventhubsreingest.inputs.capture.format\":\"avro\""

APP_CONF="$APP_CONF,\"spark.eventhubsreingest.output.eh.ns\":\"EVENTHUB NAMESPACE\""
APP_CONF="$APP_CONF,\"spark.eventhubsreingest.output.eh.name\":\"EVENTHUB NAME\""
APP_CONF="$APP_CONF,\"spark.eventhubsreingest.output.eh.keyName\":\"RootManageSharedAccessKey\""
APP_CONF="$APP_CONF,\"spark.eventhubsreingest.output.eh.keyValue\":\"EVENTHUB Key"

export APP_CONF="$APP_CONF"

./livy.sh

Checking the status can be done through the Yarn UI, or using curl -k --user "$CLUSTER_USER:$CLUSTER_PASSWORD" "$CLUSTER_URL/livy/batches"

Performance

On a:

  • 3 worker node HDI cluster
  • a target EventHub
    • 12 partitions
    • 12 throughput units
  • 5.6gb of capture files, with some small and some large events:
    • 1,592 blobs
    • 5,616,207,929 bytes

We manage to process the data in 15 minutes.

Throughput in mb

Number of incoming megabytes Number of incoming megabytes per min

Throughput in msg/s

Number of Incoming Messages Number of Incoming Messages per min

Distribution of Spark taks

Spark Task distribution

Do notice that the task time is highly correlated with the input size. Spark Task distribution