|
| 1 | +<!-- |
| 2 | +Licensed to the Apache Software Foundation (ASF) under one |
| 3 | +or more contributor license agreements. See the NOTICE file |
| 4 | +distributed with this work for additional information |
| 5 | +regarding copyright ownership. The ASF licenses this file |
| 6 | +to you under the Apache License, Version 2.0 (the |
| 7 | +"License"); you may not use this file except in compliance |
| 8 | +with the License. You may obtain a copy of the License at |
| 9 | +
|
| 10 | + http://www.apache.org/licenses/LICENSE-2.0 |
| 11 | +
|
| 12 | +Unless required by applicable law or agreed to in writing, |
| 13 | +software distributed under the License is distributed on an |
| 14 | +"AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
| 15 | +KIND, either express or implied. See the License for the |
| 16 | +specific language governing permissions and limitations |
| 17 | +under the License. |
| 18 | +--> |
| 19 | + |
| 20 | +# Logical plans via datafusion-proto |
| 21 | + |
| 22 | +`SessionContext.fromProto(byte[])` accepts a serialized DataFusion |
| 23 | +`LogicalPlanNode` and returns a lazy `DataFrame`. This is useful when you |
| 24 | +already have a plan produced by another DataFusion-aware tool, or when |
| 25 | +you want to construct the plan programmatically with finer-grained |
| 26 | +control than the `sql` or DataFrame APIs. |
| 27 | + |
| 28 | +The protobuf Java classes are generated by the build into the |
| 29 | +`org.apache.datafusion.protobuf` (plan and expression nodes) and |
| 30 | +`datafusion_common` (scalar values, schema, column references, file |
| 31 | +formats) packages. The Maven build downloads pinned `.proto` files from |
| 32 | +the matching upstream DataFusion tag on first build, then generates the |
| 33 | +Java classes locally — see the |
| 34 | +[Contributor Guide](../contributor-guide/updating-datafusion-version.md) |
| 35 | +for how to bump the version. |
| 36 | + |
| 37 | +## A minimal plan |
| 38 | + |
| 39 | +The smallest interesting plan is a projection of a literal over an |
| 40 | +empty input. It is useful as a sanity check and exercises serialization |
| 41 | +end-to-end without touching any storage. |
| 42 | + |
| 43 | +```java |
| 44 | +import org.apache.arrow.memory.RootAllocator; |
| 45 | +import org.apache.arrow.vector.IntVector; |
| 46 | +import org.apache.arrow.vector.VectorSchemaRoot; |
| 47 | +import org.apache.arrow.vector.ipc.ArrowReader; |
| 48 | +import org.apache.datafusion.DataFrame; |
| 49 | +import org.apache.datafusion.SessionContext; |
| 50 | +import org.apache.datafusion.protobuf.EmptyRelationNode; |
| 51 | +import org.apache.datafusion.protobuf.LogicalExprNode; |
| 52 | +import org.apache.datafusion.protobuf.LogicalPlanNode; |
| 53 | +import org.apache.datafusion.protobuf.ProjectionNode; |
| 54 | + |
| 55 | +import datafusion_common.DatafusionCommon; |
| 56 | + |
| 57 | +LogicalPlanNode plan = |
| 58 | + LogicalPlanNode.newBuilder() |
| 59 | + .setProjection( |
| 60 | + ProjectionNode.newBuilder() |
| 61 | + .setInput( |
| 62 | + LogicalPlanNode.newBuilder() |
| 63 | + .setEmptyRelation( |
| 64 | + EmptyRelationNode.newBuilder().setProduceOneRow(true).build()) |
| 65 | + .build()) |
| 66 | + .addExpr( |
| 67 | + LogicalExprNode.newBuilder() |
| 68 | + .setLiteral( |
| 69 | + DatafusionCommon.ScalarValue.newBuilder().setInt32Value(1).build()) |
| 70 | + .build()) |
| 71 | + .build()) |
| 72 | + .build(); |
| 73 | + |
| 74 | +try (var allocator = new RootAllocator(); |
| 75 | + SessionContext ctx = new SessionContext(); |
| 76 | + DataFrame df = ctx.fromProto(plan.toByteArray()); |
| 77 | + ArrowReader reader = df.collect(allocator)) { |
| 78 | + reader.loadNextBatch(); |
| 79 | + VectorSchemaRoot batch = reader.getVectorSchemaRoot(); |
| 80 | + IntVector col = (IntVector) batch.getVector(0); |
| 81 | + System.out.println(col.get(0)); // 1 |
| 82 | +} |
| 83 | +``` |
| 84 | + |
| 85 | +`fromProto` performs the same logical-planning, optimization, and |
| 86 | +physical-planning pipeline as `sql`; the result is a lazy |
| 87 | +[`DataFrame`](dataframe.md) that only executes when you pull results. |
| 88 | + |
| 89 | +## Scanning a Parquet file via ListingTableScanNode |
| 90 | + |
| 91 | +A `ListingTableScanNode` reads one or more files of the same format |
| 92 | +from disk. Unlike `registerParquet`, it does not require the table to |
| 93 | +be in the catalog — the scan node carries everything DataFusion needs: |
| 94 | +the file paths, the schema, the projection, the file format, and the |
| 95 | +target partition count. |
| 96 | + |
| 97 | +The scan node's `schema` field is a `datafusion_common.Schema`, not an |
| 98 | +Arrow `Schema`. Convert between the two with the helper in |
| 99 | +`org.apache.datafusion.proto.SchemaConverter`: |
| 100 | + |
| 101 | +```java |
| 102 | +import org.apache.arrow.vector.types.pojo.Schema; |
| 103 | +import org.apache.datafusion.proto.SchemaConverter; |
| 104 | + |
| 105 | +Schema arrow = ctx.tableSchema("lineitem"); |
| 106 | +DatafusionCommon.Schema schemaProto = SchemaConverter.toProto(arrow); |
| 107 | +``` |
| 108 | + |
| 109 | +The full example: register the file once to introspect its schema, |
| 110 | +then build a plan that scans the same file, sorts by `l_orderkey`, |
| 111 | +and fetches the first row. This is equivalent to |
| 112 | +`SELECT l_orderkey FROM lineitem ORDER BY l_orderkey LIMIT 1`. |
| 113 | + |
| 114 | +```java |
| 115 | +import org.apache.datafusion.protobuf.BareTableReference; |
| 116 | +import org.apache.datafusion.protobuf.ListingTableScanNode; |
| 117 | +import org.apache.datafusion.protobuf.ProjectionColumns; |
| 118 | +import org.apache.datafusion.protobuf.SortExprNode; |
| 119 | +import org.apache.datafusion.protobuf.SortNode; |
| 120 | +import org.apache.datafusion.protobuf.TableReference; |
| 121 | + |
| 122 | +String path = "/path/to/lineitem.parquet"; |
| 123 | + |
| 124 | +try (var allocator = new RootAllocator(); |
| 125 | + SessionContext ctx = new SessionContext()) { |
| 126 | + |
| 127 | + ctx.registerParquet("lineitem", path); |
| 128 | + DatafusionCommon.Schema schemaProto = |
| 129 | + SchemaConverter.toProto(ctx.tableSchema("lineitem")); |
| 130 | + |
| 131 | + LogicalExprNode orderKeyCol = |
| 132 | + LogicalExprNode.newBuilder() |
| 133 | + .setColumn(DatafusionCommon.Column.newBuilder().setName("l_orderkey").build()) |
| 134 | + .build(); |
| 135 | + |
| 136 | + LogicalPlanNode plan = |
| 137 | + LogicalPlanNode.newBuilder() |
| 138 | + .setSort( |
| 139 | + SortNode.newBuilder() |
| 140 | + .setInput( |
| 141 | + LogicalPlanNode.newBuilder() |
| 142 | + .setListingScan( |
| 143 | + ListingTableScanNode.newBuilder() |
| 144 | + .setTableName( |
| 145 | + TableReference.newBuilder() |
| 146 | + .setBare( |
| 147 | + BareTableReference.newBuilder() |
| 148 | + .setTable("lineitem") |
| 149 | + .build()) |
| 150 | + .build()) |
| 151 | + .addPaths(path) |
| 152 | + .setFileExtension(".parquet") |
| 153 | + .setSchema(schemaProto) |
| 154 | + .setProjection( |
| 155 | + ProjectionColumns.newBuilder() |
| 156 | + .addColumns("l_orderkey") |
| 157 | + .build()) |
| 158 | + .setParquet( |
| 159 | + DatafusionCommon.ParquetFormat.getDefaultInstance()) |
| 160 | + .setTargetPartitions(1) |
| 161 | + .build()) |
| 162 | + .build()) |
| 163 | + .addExpr( |
| 164 | + SortExprNode.newBuilder() |
| 165 | + .setExpr(orderKeyCol) |
| 166 | + .setAsc(true) |
| 167 | + .setNullsFirst(false) |
| 168 | + .build()) |
| 169 | + .setFetch(1) |
| 170 | + .build()) |
| 171 | + .build(); |
| 172 | + |
| 173 | + try (DataFrame df = ctx.fromProto(plan.toByteArray()); |
| 174 | + ArrowReader reader = df.collect(allocator)) { |
| 175 | + reader.loadNextBatch(); |
| 176 | + // ... |
| 177 | + } |
| 178 | +} |
| 179 | +``` |
| 180 | + |
| 181 | +## When to use proto plans |
| 182 | + |
| 183 | +The `sql` and DataFrame APIs are the right choice for most workloads. |
| 184 | +Reach for `fromProto` when you need one of: |
| 185 | + |
| 186 | +- **Cross-tool interop.** Accept plans produced by another |
| 187 | + DataFusion-based system (a planner, a scheduler, a query frontend). |
| 188 | +- **Programmatic plan construction.** Build the plan node tree directly |
| 189 | + instead of going through SQL parsing, useful for tools that compile |
| 190 | + their own surface language to DataFusion. |
| 191 | +- **Plan persistence.** Serialize a plan to bytes, store or transmit |
| 192 | + it, and execute it later — possibly in a different process or on a |
| 193 | + different machine. |
| 194 | + |
| 195 | +## Schema conversion support |
| 196 | + |
| 197 | +`SchemaConverter.toProto` and `SchemaConverter.fromProto` support the |
| 198 | +primitive Arrow types this project's tests exercise: `Bool`, signed and |
| 199 | +unsigned integer types 8 through 64 bits, `Float32`, `Float64`, `Utf8`, |
| 200 | +`Utf8View`, `LargeUtf8`, `Date32`, and `Decimal128`. Anything else |
| 201 | +raises `UnsupportedOperationException` naming the offending type. |
0 commit comments