rodmeneses commented on code in PR #12071:
URL: https://github.com/apache/iceberg/pull/12071#discussion_r2112131384
##########
flink/v1.20/flink/src/main/java/org/apache/iceberg/flink/sink/IcebergSink.java:
##########
@@ -645,72 +711,135 @@ private DataStream<RowData>
distributeDataStream(DataStream<RowData> input) {
DistributionMode mode = flinkWriteConf.distributionMode();
Schema schema = table.schema();
PartitionSpec spec = table.spec();
+ SortOrder sortOrder = table.sortOrder();
+
LOG.info("Write distribution mode is '{}'", mode.modeName());
switch (mode) {
case NONE:
- if (equalityFieldIds.isEmpty()) {
- return input;
- } else {
- LOG.info("Distribute rows by equality fields, because there are
equality fields set");
- return input.keyBy(new EqualityFieldKeySelector(schema,
flinkRowType, equalityFieldIds));
- }
-
+ return distributeDataStreamByNoneDistributionMode(input, schema);
case HASH:
- if (equalityFieldIds.isEmpty()) {
- if (table.spec().isUnpartitioned()) {
- LOG.warn(
- "Fallback to use 'none' distribution mode, because there are
no equality fields set "
- + "and table is unpartitioned");
- return input;
- } else {
- if (BucketPartitionerUtil.hasOneBucketField(spec)) {
- return input.partitionCustom(
- new BucketPartitioner(spec),
- new BucketPartitionKeySelector(spec, schema, flinkRowType));
- } else {
- return input.keyBy(new PartitionKeySelector(spec, schema,
flinkRowType));
- }
- }
- } else {
- if (spec.isUnpartitioned()) {
- LOG.info(
- "Distribute rows by equality fields, because there are
equality fields set "
- + "and table is unpartitioned");
- return input.keyBy(
- new EqualityFieldKeySelector(schema, flinkRowType,
equalityFieldIds));
- } else {
- for (PartitionField partitionField : spec.fields()) {
- Preconditions.checkState(
- equalityFieldIds.contains(partitionField.sourceId()),
- "In 'hash' distribution mode with equality fields set,
partition field '%s' "
- + "should be included in equality fields: '%s'",
- partitionField,
- equalityFieldColumns);
- }
- return input.keyBy(new PartitionKeySelector(spec, schema,
flinkRowType));
- }
- }
-
+ return distributeDataStreamByHashDistributionMode(input, schema, spec);
case RANGE:
- if (equalityFieldIds.isEmpty()) {
- LOG.warn(
- "Fallback to use 'none' distribution mode, because there are no
equality fields set "
- + "and {}=range is not supported yet in flink",
- WRITE_DISTRIBUTION_MODE);
- return input;
- } else {
- LOG.info(
- "Distribute rows by equality fields, because there are equality
fields set "
- + "and{}=range is not supported yet in flink",
- WRITE_DISTRIBUTION_MODE);
- return input.keyBy(new EqualityFieldKeySelector(schema,
flinkRowType, equalityFieldIds));
- }
-
+ return distributeDataStreamByRangeDistributionMode(input, schema,
spec, sortOrder);
default:
throw new RuntimeException("Unrecognized " + WRITE_DISTRIBUTION_MODE +
": " + mode);
}
}
+ private DataStream<RowData> distributeDataStreamByNoneDistributionMode(
+ DataStream<RowData> input, Schema iSchema) {
+ if (equalityFieldIds.isEmpty()) {
+ return input;
+ } else {
+ LOG.info("Distribute rows by equality fields, because there are equality
fields set");
+ return input.keyBy(new EqualityFieldKeySelector(iSchema, flinkRowType,
equalityFieldIds));
+ }
+ }
+
+ private DataStream<RowData> distributeDataStreamByHashDistributionMode(
+ DataStream<RowData> input, Schema iSchema, PartitionSpec partitionSpec) {
+ if (equalityFieldIds.isEmpty()) {
+ if (partitionSpec.isUnpartitioned()) {
+ LOG.warn(
+ "Fallback to use 'none' distribution mode, because there are no
equality fields set "
+ + "and table is unpartitioned");
+ return input;
+ } else {
+ return input.keyBy(new PartitionKeySelector(partitionSpec, iSchema,
flinkRowType));
+ }
+ } else {
+ if (partitionSpec.isUnpartitioned()) {
+ LOG.info(
+ "Distribute rows by equality fields, because there are equality
fields set "
+ + "and table is unpartitioned");
+ return input.keyBy(new EqualityFieldKeySelector(iSchema, flinkRowType,
equalityFieldIds));
+ } else {
+ for (PartitionField partitionField : partitionSpec.fields()) {
+ Preconditions.checkState(
+ equalityFieldIds.contains(partitionField.sourceId()),
+ "In 'hash' distribution mode with equality fields set, partition
field '%s' "
+ + "should be included in equality fields: '%s'",
+ partitionField,
+ equalityFieldColumns);
+ }
+ return input.keyBy(new PartitionKeySelector(partitionSpec, iSchema,
flinkRowType));
+ }
+ }
+ }
+
+ private DataStream<RowData> distributeDataStreamByRangeDistributionMode(
+ DataStream<RowData> input,
+ Schema iSchema,
+ PartitionSpec partitionSpec,
+ SortOrder sortOrderParam) {
+
+ int writerParallelism =
Review Comment:
Hi @stevenzwu . Thanks for the clarification. I think I know what you mean.
On FlinkSink, even without talking about RANGE distribution, the parallelism of
the v1 sink by default will be the same as the input source parallelism.
This is a good approach, because it encourage chaining.
However, we dont have that logic on the IcebergSink v2. I think we should
have the same logic there. I could do that in another PR and the follow up with
this one. What do you think? Thanks
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
