|
| 1 | +package krangl |
| 2 | + |
| 3 | +import org.apache.arrow.memory.BufferAllocator |
| 4 | +import org.apache.arrow.memory.RootAllocator |
| 5 | +import org.apache.arrow.vector.BaseFixedWidthVector |
| 6 | +import org.apache.arrow.vector.BigIntVector |
| 7 | +import org.apache.arrow.vector.BitVector |
| 8 | +import org.apache.arrow.vector.Float4Vector |
| 9 | +import org.apache.arrow.vector.Float8Vector |
| 10 | +import org.apache.arrow.vector.IntVector |
| 11 | +import org.apache.arrow.vector.SmallIntVector |
| 12 | +import org.apache.arrow.vector.TinyIntVector |
| 13 | +import org.apache.arrow.vector.VarCharVector |
| 14 | +import org.apache.arrow.vector.VectorSchemaRoot |
| 15 | +import org.apache.arrow.vector.ipc.ArrowFileReader |
| 16 | +import org.apache.arrow.vector.ipc.ArrowFileWriter |
| 17 | +import org.apache.arrow.vector.types.FloatingPointPrecision |
| 18 | +import org.apache.arrow.vector.types.pojo.ArrowType |
| 19 | +import org.apache.arrow.vector.types.pojo.Schema |
| 20 | +import org.apache.arrow.vector.util.ByteArrayReadableSeekableByteChannel |
| 21 | +import org.apache.arrow.vector.util.Text |
| 22 | +import java.io.ByteArrayOutputStream |
| 23 | +import java.io.File |
| 24 | +import java.nio.channels.* |
| 25 | +import java.nio.file.StandardOpenOption |
| 26 | +import java.util.* |
| 27 | + |
| 28 | +internal fun unwrapStringArrayFromArrow(vector: VarCharVector): ArrayList<String?> { |
| 29 | + val result = ArrayList<String?>() |
| 30 | + for (i in 0 until vector.valueCount) { |
| 31 | + result.add(vector.getObject(i)?.toString()) |
| 32 | + } |
| 33 | + return result |
| 34 | +} |
| 35 | + |
| 36 | +internal inline fun <reified ELEMENT_TYPE>unwrapNumericVectorFromArrow(vector: BaseFixedWidthVector, elementClass: Class<ELEMENT_TYPE>): List<ELEMENT_TYPE?> { |
| 37 | + val elements = vector.valueCount |
| 38 | + val outVector = ArrayList<ELEMENT_TYPE?>(elements) |
| 39 | + for (i in 0 until elements) { |
| 40 | + outVector.add(vector.getObject(i) as ELEMENT_TYPE?) |
| 41 | + } |
| 42 | + return outVector |
| 43 | +} |
| 44 | + |
| 45 | +internal fun unwrapBooleanArrayFromArrow(vector: BitVector): ArrayList<Boolean?> { |
| 46 | + val result = ArrayList<Boolean?>() |
| 47 | + for (i in 0 until vector.valueCount) { |
| 48 | + result.add(vector.getObject(i)) |
| 49 | + } |
| 50 | + return result |
| 51 | +} |
| 52 | + |
| 53 | +fun DataFrame.Companion.arrowReader() = ArrowReader() |
| 54 | + |
| 55 | +class ArrowReader() { |
| 56 | + /** |
| 57 | + * Internal low-level function. |
| 58 | + * Use this function if you are working with [VectorSchemaRoot]s directly in your project. |
| 59 | + */ |
| 60 | + fun fromVectorSchemaRoot(vectorSchemaRoot: VectorSchemaRoot): DataFrame { |
| 61 | + val kranglVectors = vectorSchemaRoot.fieldVectors.map { fieldVector -> |
| 62 | + when (fieldVector.field.type) { |
| 63 | + is ArrowType.FixedSizeList, is ArrowType.List -> { |
| 64 | + throw Exception("Matrices are not supported yet") |
| 65 | + } |
| 66 | + is ArrowType.Utf8 -> { |
| 67 | + StringCol(fieldVector.name, unwrapStringArrayFromArrow(fieldVector as VarCharVector)) |
| 68 | + } |
| 69 | + is ArrowType.Int -> { |
| 70 | + val bitWidth = (fieldVector.field.type as ArrowType.Int).bitWidth |
| 71 | + when (bitWidth) { |
| 72 | + 8 -> IntCol(fieldVector.name, unwrapNumericVectorFromArrow(fieldVector as TinyIntVector, Int::class.java)) |
| 73 | + 16 -> IntCol(fieldVector.name, unwrapNumericVectorFromArrow(fieldVector as SmallIntVector, Int::class.java)) |
| 74 | + 32 -> IntCol(fieldVector.name, unwrapNumericVectorFromArrow(fieldVector as IntVector, Int::class.java)) |
| 75 | + 64 -> LongCol(fieldVector.name, unwrapNumericVectorFromArrow(fieldVector as BigIntVector, Long::class.java)) |
| 76 | + else -> throw java.lang.Exception("Incorrect Int.bitWidth ($bitWidth, should never happen)") |
| 77 | + } |
| 78 | + } |
| 79 | + is ArrowType.FloatingPoint -> { |
| 80 | + val precision = (fieldVector.field.type as ArrowType.FloatingPoint).precision |
| 81 | + when (precision) { |
| 82 | + FloatingPointPrecision.HALF -> java.lang.Exception("HALF float not supported") |
| 83 | + FloatingPointPrecision.SINGLE -> DoubleCol(fieldVector.name, unwrapNumericVectorFromArrow(fieldVector as Float4Vector, Double::class.java)) |
| 84 | + FloatingPointPrecision.DOUBLE -> DoubleCol(fieldVector.name, unwrapNumericVectorFromArrow(fieldVector as Float8Vector, Double::class.java)) |
| 85 | + else -> throw java.lang.Exception("Incorrect FloatingPoint.precision ($precision, should never happen)") |
| 86 | + } |
| 87 | + } |
| 88 | + is ArrowType.Bool -> { |
| 89 | + BooleanCol(fieldVector.name, unwrapBooleanArrayFromArrow(fieldVector as BitVector)) |
| 90 | + } |
| 91 | + else -> { |
| 92 | + throw Exception("${fieldVector.field.type.typeID.name} is not supported yet") |
| 93 | + } |
| 94 | + } |
| 95 | + } |
| 96 | + |
| 97 | + return dataFrameOf(*(kranglVectors as List<DataCol>).toTypedArray()) |
| 98 | + } |
| 99 | + |
| 100 | + /** |
| 101 | + * Read [VectorSchemaRoot] from existing [channel] and convert it to [DataFrame]. |
| 102 | + * Use this function if you want to manage channels yourself, make in-memory IPC sharing and so on. |
| 103 | + * If [allocator] is null, it will be created and closed inside. |
| 104 | + */ |
| 105 | + fun readFromChannel(channel: SeekableByteChannel, allocator: BufferAllocator?): DataFrame { |
| 106 | + fun readFromChannelAllocating(channel: SeekableByteChannel, allocator: BufferAllocator?): DataFrame { |
| 107 | + ArrowFileReader(channel, allocator).use { reader -> |
| 108 | + reader.loadNextBatch() |
| 109 | + return fromVectorSchemaRoot(reader.vectorSchemaRoot) |
| 110 | + } |
| 111 | + } |
| 112 | + if (allocator == null ) { |
| 113 | + RootAllocator().use { newAllocator -> |
| 114 | + return readFromChannelAllocating(channel, newAllocator) |
| 115 | + } |
| 116 | + } else { |
| 117 | + return readFromChannelAllocating(channel, allocator) |
| 118 | + } |
| 119 | + } |
| 120 | + |
| 121 | + /** |
| 122 | + * Read [VectorSchemaRoot] from ByteArray and convert it to [DataFrame]. |
| 123 | + */ |
| 124 | + fun fromByteArray(byteArray: ByteArray): DataFrame { |
| 125 | + return readFromChannel(ByteArrayReadableSeekableByteChannel(byteArray), null) |
| 126 | + } |
| 127 | + |
| 128 | + /** |
| 129 | + * Read [VectorSchemaRoot] from [file] by and convert it to [DataFrame]. |
| 130 | + */ |
| 131 | + fun fromFile(file: File): DataFrame { |
| 132 | + if (!file.exists()) { |
| 133 | + throw Exception("${file.path} does not exist") |
| 134 | + } |
| 135 | + if (file.isDirectory) { |
| 136 | + throw Exception("${file.path} is directory") |
| 137 | + } |
| 138 | + FileChannel.open( |
| 139 | + file.toPath(), |
| 140 | + StandardOpenOption.READ |
| 141 | + ).use { channel -> |
| 142 | + return readFromChannel(channel, null) |
| 143 | + } |
| 144 | + } |
| 145 | + |
| 146 | + /** |
| 147 | + * Read [VectorSchemaRoot] from file by [path] and convert it to [DataFrame]. |
| 148 | + */ |
| 149 | + fun fromFile(path: String): DataFrame { |
| 150 | + return fromFile(File(path)) |
| 151 | + } |
| 152 | +} |
| 153 | + |
| 154 | +fun DataFrame.arrowWriter() = ArrowWriter(this) |
| 155 | + |
| 156 | +class ArrowWriter(val dataFrame: DataFrame) { |
| 157 | + internal fun fromStringCol(column: StringCol, allocator: BufferAllocator): VarCharVector { |
| 158 | + val fieldVector = VarCharVector(column.name, allocator) |
| 159 | + fieldVector.allocateNew(column.length) |
| 160 | + column.values.forEachIndexed { index, value -> |
| 161 | + if (value == null) { |
| 162 | + fieldVector.setNull(index) |
| 163 | + } else { |
| 164 | + fieldVector.setSafe(index, Text(value)) |
| 165 | + } |
| 166 | + } |
| 167 | + fieldVector.valueCount = column.length |
| 168 | + return fieldVector |
| 169 | + } |
| 170 | + |
| 171 | + internal fun fromBooleanCol(column: BooleanCol, allocator: BufferAllocator): BitVector { |
| 172 | + val fieldVector = BitVector(column.name, allocator) |
| 173 | + fieldVector.allocateNew(column.length) |
| 174 | + column.values.forEachIndexed { index, value -> |
| 175 | + if (value == null) { |
| 176 | + fieldVector.setNull(index) |
| 177 | + } else { |
| 178 | + fieldVector.setSafe(index, if (value) 1 else 0) |
| 179 | + } |
| 180 | + } |
| 181 | + fieldVector.valueCount = column.length |
| 182 | + return fieldVector |
| 183 | + } |
| 184 | + |
| 185 | + internal fun fromIntCol(column: IntCol, allocator: BufferAllocator): IntVector { |
| 186 | + val fieldVector = IntVector(column.name, allocator) |
| 187 | + fieldVector.allocateNew(column.length) |
| 188 | + column.values.forEachIndexed { index, value -> |
| 189 | + if (value == null) { |
| 190 | + fieldVector.setNull(index) |
| 191 | + } else { |
| 192 | + fieldVector.setSafe(index, value) |
| 193 | + } |
| 194 | + } |
| 195 | + fieldVector.valueCount = column.length |
| 196 | + return fieldVector |
| 197 | + } |
| 198 | + |
| 199 | + internal fun fromLongCol(column: LongCol, allocator: BufferAllocator): BigIntVector { |
| 200 | + val fieldVector = BigIntVector(column.name, allocator) |
| 201 | + fieldVector.allocateNew(column.length) |
| 202 | + column.values.forEachIndexed { index, value -> |
| 203 | + if (value == null) { |
| 204 | + fieldVector.setNull(index) |
| 205 | + } else { |
| 206 | + fieldVector.setSafe(index, value) |
| 207 | + } |
| 208 | + } |
| 209 | + fieldVector.valueCount = column.length |
| 210 | + return fieldVector |
| 211 | + } |
| 212 | + |
| 213 | + internal fun fromDoubleCol(column: DoubleCol, allocator: BufferAllocator): Float8Vector { |
| 214 | + val fieldVector = Float8Vector(column.name, allocator) |
| 215 | + fieldVector.allocateNew(column.length) |
| 216 | + column.values.forEachIndexed { index, value -> |
| 217 | + if (value == null) { |
| 218 | + fieldVector.setNull(index) |
| 219 | + } else { |
| 220 | + fieldVector.setSafe(index, value) |
| 221 | + } |
| 222 | + } |
| 223 | + fieldVector.valueCount = column.length |
| 224 | + return fieldVector |
| 225 | + } |
| 226 | + |
| 227 | + internal fun fromAnyCol(column: AnyCol, allocator: BufferAllocator): VarCharVector { |
| 228 | + val fieldVector = VarCharVector(column.name, allocator) |
| 229 | + fieldVector.allocateNew(column.length) |
| 230 | + column.values.forEachIndexed { index, value -> |
| 231 | + if (value == null) { |
| 232 | + fieldVector.setNull(index) |
| 233 | + } else { |
| 234 | + fieldVector.setSafe(index, Text(value.toString())) |
| 235 | + } |
| 236 | + } |
| 237 | + fieldVector.valueCount = column.length |
| 238 | + return fieldVector |
| 239 | + } |
| 240 | + |
| 241 | + /** |
| 242 | + * Internal low-level function. |
| 243 | + * Use this function if you are working with [VectorSchemaRoot]s and [BufferAllocator]s directly in your project. |
| 244 | + */ |
| 245 | + fun allocateVectorSchemaRoot(allocator: BufferAllocator): VectorSchemaRoot { |
| 246 | + val arrowVectors = dataFrame.cols.map { column -> |
| 247 | + when (column) { |
| 248 | + is StringCol -> fromStringCol(column, allocator) |
| 249 | + is BooleanCol -> fromBooleanCol(column, allocator) |
| 250 | + is IntCol -> fromIntCol(column, allocator) |
| 251 | + is LongCol -> fromLongCol(column, allocator) |
| 252 | + is DoubleCol -> fromDoubleCol(column, allocator) |
| 253 | + is AnyCol -> fromAnyCol(column, allocator) |
| 254 | + else -> { |
| 255 | + throw Exception("Unknown column type ${column.javaClass.canonicalName}") |
| 256 | + } |
| 257 | + } |
| 258 | + } |
| 259 | + return VectorSchemaRoot(arrowVectors) |
| 260 | + } |
| 261 | + |
| 262 | + /** |
| 263 | + * Export [dataFrame] to [VectorSchemaRoot] and write it to any existing [channel]. |
| 264 | + * Use this function if you want to manage channels yourself, make in-memory IPC sharing and so on |
| 265 | + */ |
| 266 | + fun writeToChannel(channel: WritableByteChannel) { |
| 267 | + RootAllocator().use { allocator -> |
| 268 | + this.allocateVectorSchemaRoot(allocator).use { vectorSchemaRoot -> |
| 269 | + ArrowFileWriter(vectorSchemaRoot, null, channel).use { writer -> |
| 270 | + writer.writeBatch(); |
| 271 | + } |
| 272 | + } |
| 273 | + } |
| 274 | + } |
| 275 | + |
| 276 | + /** |
| 277 | + * Export [dataFrame] to [VectorSchemaRoot] and write it to new ByteArray. |
| 278 | + */ |
| 279 | + fun toByteArray(): ByteArray { |
| 280 | + ByteArrayOutputStream().use { byteArrayStream -> |
| 281 | + Channels.newChannel(byteArrayStream).use { channel -> |
| 282 | + writeToChannel(channel) |
| 283 | + return byteArrayStream.toByteArray() |
| 284 | + } |
| 285 | + } |
| 286 | + } |
| 287 | + |
| 288 | + /** |
| 289 | + * Export [dataFrame] to [VectorSchemaRoot] and write it to new or existing [file]. |
| 290 | + * Temporary file is created if [file] argument is null. |
| 291 | + */ |
| 292 | + fun toFile(file: File?): File { |
| 293 | + val saveToFile = file ?: File.createTempFile("DataFrame", ".arrow") |
| 294 | + |
| 295 | + FileChannel.open( |
| 296 | + saveToFile.toPath(), |
| 297 | + StandardOpenOption.WRITE, |
| 298 | + StandardOpenOption.CREATE |
| 299 | + ).use { channel -> |
| 300 | + channel.truncate(0) |
| 301 | + writeToChannel(channel) |
| 302 | + } |
| 303 | + return saveToFile |
| 304 | + } |
| 305 | + |
| 306 | + /** |
| 307 | + * Export [dataFrame] to [VectorSchemaRoot] and write it to new or existing file by [path]. |
| 308 | + * Temporary file is created if [path] argument is null. |
| 309 | + */ |
| 310 | + fun toFile(path: String?): File { |
| 311 | + val saveToFile = if (path != null) { |
| 312 | + File(path) |
| 313 | + } else { |
| 314 | + File.createTempFile("DataFrame", ".arrow") |
| 315 | + } |
| 316 | + return toFile(saveToFile) |
| 317 | + } |
| 318 | +} |
0 commit comments