Spark Structured Streaming
Iceberg uses Apache Spark’s DataSourceV2 API for data source and catalog implementations. Spark DSv2 is an evolving API with different levels of support in Spark versions.
As of Spark 3.0, DataFrame reads and writes are supported.
|Feature support||Spark 3.0||Spark 2.4||Notes|
Iceberg supports processing incremental data in spark structured streaming jobs which starts from a historical timestamp:
val df = spark.readStream .format("iceberg") .option("stream-from-timestamp", Long.toString(streamStartTimestamp)) .load("database.table_name")
streaming-skip-overwrite-snapshots=true. Similarly, delete snapshots will cause an exception by default, and deletes may be ignored by setting
To write values from streaming query to Iceberg table, use
val tableIdentifier: String = ... data.writeStream .format("iceberg") .outputMode("append") .trigger(Trigger.ProcessingTime(1, TimeUnit.MINUTES)) .option("path", tableIdentifier) .option("checkpointLocation", checkpointPath) .start()
tableIdentifier can be:
- The fully-qualified path to a HDFS table, like
- A table name if the table is tracked by a catalog, like
Iceberg doesn’t support “continuous processing”, as it doesn’t provide the interface to “commit” the output.
complete output modes:
append: appends the rows of every micro-batch to the table
complete: replaces the table contents every micro-batch
The table should be created in prior to start the streaming query. Refer SQL create table on Spark page to see how to create the Iceberg table.
Writing against partitioned table
Iceberg requires the data to be sorted according to the partition spec per task (Spark partition) in prior to write against partitioned table. For batch queries you’re encouraged to do explicit sort to fulfill the requirement (see here), but the approach would bring additional latency as repartition and sort are considered as heavy operations for streaming workload. To avoid additional latency, you can enable fanout writer to eliminate the requirement.
val tableIdentifier: String = ... data.writeStream .format("iceberg") .outputMode("append") .trigger(Trigger.ProcessingTime(1, TimeUnit.MINUTES)) .option("path", tableIdentifier) .option("fanout-enabled", "true") .option("checkpointLocation", checkpointPath) .start()
Fanout writer opens the files per partition value and doesn’t close these files till write task is finished. This functionality is discouraged for batch query, as explicit sort against output rows isn’t expensive for batch workload.
Maintenance for streaming tables
Streaming queries can create new table versions quickly, which creates lots of table metadata to track those versions. Maintaining metadata by tuning the rate of commits, expiring old snapshots, and automatically cleaning up metadata files is highly recommended.
Tune the rate of commits
Having high rate of commits would produce lots of data files, manifests, and snapshots which leads the table hard to maintain. We encourage having trigger interval 1 minute at minimum, and increase the interval if needed.
The triggers section in Structured Streaming Programming Guide documents how to configure the interval.
Expire old snapshots
Each micro-batch written to a table produces a new snapshot, which are tracked in table metadata until they are expired to remove the metadata and any data files that are no longer needed. Snapshots accumulate quickly with frequent commits, so it is highly recommended that tables written by streaming queries are regularly maintained.
Compacting data files
The amount of data written in a micro batch is typically small, which can cause the table metadata to track lots of small files. Compacting small files into larger files reduces the metadata needed by the table, and increases query efficiency.
To optimize write latency on streaming workload, Iceberg may write the new snapshot with a “fast” append that does not automatically compact manifests. This could lead lots of small manifest files. Manifests can be rewritten to optimize queries and to compact.