Github user rdblue commented on a diff in the pull request:
https://github.com/apache/spark/pull/19269#discussion_r139832973
--- Diff:
sql/core/src/main/java/org/apache/spark/sql/sources/v2/writer/DataSourceV2Writer.java
---
@@ -0,0 +1,71 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.sql.sources.v2.writer;
+
+import org.apache.spark.annotation.InterfaceStability;
+import org.apache.spark.sql.Row;
+import org.apache.spark.sql.SaveMode;
+import org.apache.spark.sql.sources.v2.DataSourceV2Options;
+import org.apache.spark.sql.sources.v2.WriteSupport;
+import org.apache.spark.sql.types.StructType;
+
+/**
+ * A data source writer that is returned by
+ * {@link WriteSupport#createWriter(StructType, SaveMode,
DataSourceV2Options)}.
+ * It can mix in various writing optimization interfaces to speed up the
data saving. The actual
+ * writing logic is delegated to {@link WriteTask} that is returned by
{@link #createWriteTask()}.
+ *
+ * The writing procedure is:
+ * 1. Create a write task by {@link #createWriteTask()}, serialize and
send it to all the
+ * partitions of the input data(RDD).
+ * 2. For each partition, create a data writer with the write task, and
write the data of the
+ * partition with this writer. If all the data are written
successfully, call
+ * {@link DataWriter#commit()}. If exception happens during the
writing, call
+ * {@link DataWriter#abort()}. This step may repeat several times as
Spark will retry failed
+ * tasks.
+ * 3. Wait until all the writers/partitions are finished, i.e., either
committed or aborted. If
+ * all partitions are written successfully, call {@link
#commit(WriterCommitMessage[])}. If
+ * some partitions failed and aborted, call {@link #abort()}.
--- End diff --
The main reason why I wanted a separate SPIP for the write path was this
point in the doc:
> Ideally partitioning/bucketing concept should not be exposed in the Data
Source API V2, because they are just techniques for data skipping and
pre-partitioning. However, these 2 concepts are already widely used in Spark,
e.g. DataFrameWriter.partitionBy and DDL syntax like ADD PARTITION. To be
consistent, we need to add partitioning/bucketing to Data Source V2, so that
the implementations can be able to specify partitioning/bucketing for
read/write.
There's a lot in there that's worth thinking about and possibly changing:
1. Ideally, the DataSourceV2 API wouldn't support bucketing/partitioning
2. The current DataFrameWriter API is what we should continue to support
3. Implementations should supply bucketing and partitioning for writes
because of 2
**Bucketing/partitioning**: It comes down to the level at which this API is
going to be used. It looks like this API currently ignores bucketing and
partitioning (unless my read through was too quick). I think I agree that in
the long term that's a good thing, but we need ways for a data source to tell
Spark about its requirements for incoming data.
In the current version, it looks like Spark would know how to prepare data
for writers outside of this API (rather than including support as suggested by
point 3). When writing a partitioned table, Spark would get the partitioning
from the table definition in the metastore and automatically sort by partition
columns. Is that right?
I'd like to move the data store's requirements behind this API. For
example, writing to HBase files directly requires sorting by key first. We
don't want to do the sort in the writer because it may duplicate work (and
isn't captured in the physical plan), and we also don't want to require Spark
to know about the requirements of the HBase data store, or any other specific
implementation.
**DataFrameWriter API**: I'd like to talk about separating the API for
table definitions and writes, but not necessarily as part of this work. The
SPIP should clearly state whether that's part of the scope for this API, and
how the proposed API works for both if that work is going to be done in a
future proposal. A good example of that is how bucketing/partitioning will be
handled.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
