Flink Writes

Iceberg support batch and streaming writes With Apache Flink's DataStream API and Table API.

Writing with SQL

Iceberg support both INSERT INTO and INSERT OVERWRITE.

INSERT INTO

To append new data to a table with a Flink streaming job, use INSERT INTO:

INSERT INTO `hive_catalog`.`default`.`sample` VALUES (1, 'a');
INSERT INTO `hive_catalog`.`default`.`sample` SELECT id, data from other_kafka_table;

INSERT OVERWRITE

To replace data in the table with the result of a query, use INSERT OVERWRITE in batch job (flink streaming job does not support INSERT OVERWRITE). Overwrites are atomic operations for Iceberg tables.

Partitions that have rows produced by the SELECT query will be replaced, for example:

INSERT OVERWRITE sample VALUES (1, 'a');

Iceberg also support overwriting given partitions by the select values:

INSERT OVERWRITE `hive_catalog`.`default`.`sample` PARTITION(data='a') SELECT 6;

For a partitioned iceberg table, when all the partition columns are set a value in PARTITION clause, it is inserting into a static partition, otherwise if partial partition columns (prefix part of all partition columns) are set a value in PARTITION clause, it is writing the query result into a dynamic partition. For an unpartitioned iceberg table, its data will be completely overwritten by INSERT OVERWRITE.

UPSERT

Iceberg supports UPSERT based on the primary key when writing data into v2 table format. There are two ways to enable upsert.

1. Enable the UPSERT mode as table-level property write.upsert.enabled. Here is an example SQL statement to set the table property when creating a table. It would be applied for all write paths to this table (batch or streaming) unless overwritten by write options as described later.

   CREATE TABLE `hive_catalog`.`default`.`sample` (
       `id` INT COMMENT 'unique id',
       `data` STRING NOT NULL,
       PRIMARY KEY(`id`) NOT ENFORCED
   ) with ('format-version'='2', 'write.upsert.enabled'='true');
   

2. Enabling UPSERT mode using upsert-enabled in the write options provides more flexibility than a table level config. Note that you still need to use v2 table format and specify the primary key or identifier fields when creating the table.

   INSERT INTO tableName /*+ OPTIONS('upsert-enabled'='true') */
   ...
   

Info

OVERWRITE and UPSERT can't be set together. In UPSERT mode, if the table is partitioned, the partition fields should be included in equality fields.

Writing with DataStream

Iceberg support writing to iceberg table from different DataStream input.

Appending data

Flink supports writing DataStream<RowData> and DataStream<Row> to the sink iceberg table natively.

StreamExecutionEnvironment env = ...;

DataStream<RowData> input = ... ;
Configuration hadoopConf = new Configuration();
TableLoader tableLoader = TableLoader.fromHadoopTable("hdfs://nn:8020/warehouse/path", hadoopConf);

FlinkSink.forRowData(input)
    .tableLoader(tableLoader)
    .append();

env.execute("Test Iceberg DataStream");

Overwrite data

Set the overwrite flag in FlinkSink builder to overwrite the data in existing iceberg tables:

StreamExecutionEnvironment env = ...;

DataStream<RowData> input = ... ;
Configuration hadoopConf = new Configuration();
TableLoader tableLoader = TableLoader.fromHadoopTable("hdfs://nn:8020/warehouse/path", hadoopConf);

FlinkSink.forRowData(input)
    .tableLoader(tableLoader)
    .overwrite(true)
    .append();

env.execute("Test Iceberg DataStream");

Upsert data

Set the upsert flag in FlinkSink builder to upsert the data in existing iceberg table. The table must use v2 table format and have a primary key.

StreamExecutionEnvironment env = ...;

DataStream<RowData> input = ... ;
Configuration hadoopConf = new Configuration();
TableLoader tableLoader = TableLoader.fromHadoopTable("hdfs://nn:8020/warehouse/path", hadoopConf);

FlinkSink.forRowData(input)
    .tableLoader(tableLoader)
    .upsert(true)
    .append();

env.execute("Test Iceberg DataStream");

Info

OVERWRITE and UPSERT can't be set together. In UPSERT mode, if the table is partitioned, the partition fields should be included in equality fields.

Write with Avro GenericRecord

Flink Iceberg sink provides AvroGenericRecordToRowDataMapper that converts Avro GenericRecord to Flink RowData. You can use the mapper to write Avro GenericRecord DataStream to Iceberg.

Please make sure flink-avro jar is included in the classpath. Also iceberg-flink-runtime shaded bundle jar can't be used because the runtime jar shades the avro package. Please use non-shaded iceberg-flink jar instead.

DataStream<org.apache.avro.generic.GenericRecord> dataStream = ...;

Schema icebergSchema = table.schema();


// The Avro schema converted from Iceberg schema can't be used
// due to precision difference between how Iceberg schema (micro)
// and Flink AvroToRowDataConverters (milli) deal with time type.
// Instead, use the Avro schema defined directly.
// See AvroGenericRecordToRowDataMapper Javadoc for more details.
org.apache.avro.Schema avroSchema = AvroSchemaUtil.convert(icebergSchema, table.name());

GenericRecordAvroTypeInfo avroTypeInfo = new GenericRecordAvroTypeInfo(avroSchema);
RowType rowType = FlinkSchemaUtil.convert(icebergSchema);

FlinkSink.builderFor(
    dataStream,
    AvroGenericRecordToRowDataMapper.forAvroSchema(avroSchema),
    FlinkCompatibilityUtil.toTypeInfo(rowType))
  .table(table)
  .tableLoader(tableLoader)
  .append();

Branch Writes

Writing to branches in Iceberg tables is also supported via the toBranch API in FlinkSink For more information on branches please refer to branches.

FlinkSink.forRowData(input)
    .tableLoader(tableLoader)
    .toBranch("audit-branch")
    .append();

Metrics

The following Flink metrics are provided by the Flink Iceberg sink.

Parallel writer metrics are added under the sub group of IcebergStreamWriter. They should have the following key-value tags.

• table: full table name (like iceberg.my_db.my_table)
• subtask_index: writer subtask index starting from 0

Metric name	Metric type	Description
lastFlushDurationMs	Gauge	The duration (in milli) that writer subtasks take to flush and upload the files during checkpoint.
flushedDataFiles	Counter	Number of data files flushed and uploaded.
flushedDeleteFiles	Counter	Number of delete files flushed and uploaded.
flushedReferencedDataFiles	Counter	Number of data files referenced by the flushed delete files.
dataFilesSizeHistogram	Histogram	Histogram distribution of data file sizes (in bytes).
deleteFilesSizeHistogram	Histogram	Histogram distribution of delete file sizes (in bytes).

Committer metrics are added under the sub group of IcebergFilesCommitter. They should have the following key-value tags.

• table: full table name (like iceberg.my_db.my_table)

Metric name	Metric type	Description
lastCheckpointDurationMs	Gauge	The duration (in milli) that the committer operator checkpoints its state.
lastCommitDurationMs	Gauge	The duration (in milli) that the Iceberg table commit takes.
committedDataFilesCount	Counter	Number of data files committed.
committedDataFilesRecordCount	Counter	Number of records contained in the committed data files.
committedDataFilesByteCount	Counter	Number of bytes contained in the committed data files.
committedDeleteFilesCount	Counter	Number of delete files committed.
committedDeleteFilesRecordCount	Counter	Number of records contained in the committed delete files.
committedDeleteFilesByteCount	Counter	Number of bytes contained in the committed delete files.
elapsedSecondsSinceLastSuccessfulCommit	Gauge	Elapsed time (in seconds) since last successful Iceberg commit.

elapsedSecondsSinceLastSuccessfulCommit is an ideal alerting metric to detect failed or missing Iceberg commits.

• Iceberg commit happened after successful Flink checkpoint in the notifyCheckpointComplete callback. It could happen that Iceberg commits failed (for whatever reason), while Flink checkpoints succeeding.
• It could also happen that notifyCheckpointComplete wasn't triggered (for whatever bug). As a result, there won't be any Iceberg commits attempted.

If the checkpoint interval (and expected Iceberg commit interval) is 5 minutes, set up alert with rule like elapsedSecondsSinceLastSuccessfulCommit > 60 minutes to detect failed or missing Iceberg commits in the past hour.

Options

Write options

Flink write options are passed when configuring the FlinkSink, like this:

FlinkSink.Builder builder = FlinkSink.forRow(dataStream, SimpleDataUtil.FLINK_SCHEMA)
    .table(table)
    .tableLoader(tableLoader)
    .set("write-format", "orc")
    .set(FlinkWriteOptions.OVERWRITE_MODE, "true");

For Flink SQL, write options can be passed in via SQL hints like this:

INSERT INTO tableName /*+ OPTIONS('upsert-enabled'='true') */
...

Check out all the options here: write-options

Distribution mode

Flink streaming writer supports both HASH and RANGE distribution mode. You can enable it via FlinkSink#Builder#distributionMode(DistributionMode) or via write-options.

Hash distribution

HASH distribution shuffles data by partition key (partitioned table) or equality fields (non-partitioned table). It simply leverages Flink's DataStream#keyBy to distribute the data.

HASH distribution has a few limitations.

• It doesn't handle skewed data well. E.g. some partitions have a lot more data than others.
• It can result in unbalanced traffic distribution if cardinality of the partition key or equality fields is low as demonstrated by [PR 4228](https://github.com/apache/iceberg/pull/4228).
• Writer parallelism is limited to the cardinality of the hash key. If the cardinality is 10, only at most 10 writer tasks would get the traffic. Having higher writer parallelism (even if traffic volume requires) won't help.

Range distribution (experimental)

RANGE distribution shuffles data by partition key or sort order via a custom range partitioner. Range distribution collects traffic statistics to guide the range partitioner to evenly distribute traffic to writer tasks.

Range distribution only shuffle the data via range partitioner. Rows are not sorted within a data file, which Flink streaming writer doesn't support yet.

Use cases

RANGE distribution can be applied to an Iceberg table that either is partitioned or has SortOrder defined. For a partitioned table without SortOrder, partition columns are used as sort order. If SortOrder is explicitly defined for the table, it is used by the range partitioner.

Range distribution can handle skewed data. E.g.

• Table is partitioned by event time. Typically, recent hours have more data, while the long-tail hours have less and less data.
• Table is partitioned by country code, where some countries (like US) have a lot more traffic and smaller countries have a lot less data
• Table is partitioned by event type, where some types have a lot more data than others.

Range distribution can also cluster data on non-partition columns. E.g., table is partitioned hourly on ingestion time. Queries often include predicate on a non-partition column like device_id or country_code. Range partition would improve the query performance by clustering on the non-partition column when table SortOrder is defined with the non-partition column.

Traffic statistics

Statistics are collected by every shuffle operator subtask and aggregated by the coordinator for every checkpoint cycle. Aggregated statistics are broadcast to all subtasks and applied to the range partitioner in the next checkpoint. So it may take up to two checkpoint cycles to detect traffic distribution change and apply the new statistics to range partitioner.

Range distribution can work with low cardinality (like country_code) or high cardinality (like device_id) scenarios.

• For low cardinality scenario (like hundreds or thousands), HashMap is used to track traffic distribution for every key. If a new sort key value shows up, range partitioner would just round-robin it to the writer tasks before traffic distribution has been learned about the new key.
• For high cardinality scenario (like millions or billions), uniform random sampling (reservoir sampling) is used to compute range bounds that split the sort key space evenly. It keeps the memory footprint and network exchange low. Reservoir sampling work well if key distribution is relatively even. If a single hot key has unbalanced large share of the traffic, range split by uniform sampling probably won't work very well.

Usage

Here is how to enable range distribution in Java. There are two optional advanced configs. Default should work well for most cases. See write-options for details.

FlinkSink.forRowData(input)
    ...
    .distributionMode(DistributionMode.RANGE)
    .rangeDistributionStatisticsType(StatisticsType.Auto)
    .rangeDistributionSortKeyBaseWeight(0.0d)
    .append();

Overhead

Data shuffling (hash or range) has computational overhead of serialization/deserialization and network I/O. Expect some increase of CPU utilization.

Range distribution also collect and aggregate data distribution statistics. That would also incur some CPU overhead. Memory overhead is typically small if using default statistics type of Auto. Don't use Map statistics type if key cardinality is high. That could result in significant memory footprint and large network exchange for statistics aggregation.

Notes

Flink streaming write jobs rely on snapshot summary to keep the last committed checkpoint ID, and store uncommitted data as temporary files. Therefore, expiring snapshots and deleting orphan files could possibly corrupt the state of the Flink job. To avoid that, make sure to keep the last snapshot created by the Flink job (which can be identified by the flink.job-id property in the summary), and only delete orphan files that are old enough.