@VisibleForTesting
@SuppressWarnings("unchecked")
<E> FileAppender<E> newAppender(Path temp) {
Format format = descriptor.getFormat();
if (Formats.PARQUET.equals(format)) {
// by default, Parquet is not durable
if (DescriptorUtil.isDisabled(FileSystemProperties.NON_DURABLE_PARQUET_PROP, descriptor)) {
return (FileAppender<E>)
new DurableParquetAppender(
fs, temp, descriptor.getSchema(), conf, descriptor.getCompressionType());
} else {
return (FileAppender<E>)
new ParquetAppender(
fs, temp, descriptor.getSchema(), conf, descriptor.getCompressionType());
}
} else if (Formats.AVRO.equals(format)) {
return new AvroAppender<E>(fs, temp, descriptor.getSchema(), descriptor.getCompressionType());
} else if (Formats.CSV.equals(format)
&& DescriptorUtil.isEnabled(FileSystemProperties.ALLOW_CSV_PROP, descriptor)) {
return new CSVAppender<E>(fs, temp, descriptor);
} else {
this.state = ReaderWriterState.ERROR;
throw new UnknownFormatException("Unknown format " + descriptor);
}
}
@Override
public final void initialize() {
Preconditions.checkState(
state.equals(ReaderWriterState.NEW), "Unable to open a writer from state:%s", state);
// ensure the directory exists
try {
fs.mkdirs(directory);
} catch (IOException ex) {
this.state = ReaderWriterState.ERROR;
throw new DatasetIOException("Failed to create path " + directory, ex);
}
// initialize paths
this.finalPath = new Path(directory, uniqueFilename(descriptor.getFormat()));
this.tempPath = tempFilename(finalPath);
this.appender = newAppender(tempPath);
try {
appender.open();
} catch (IOException e) {
this.state = ReaderWriterState.ERROR;
throw new DatasetIOException("Failed to open appender " + appender, e);
}
this.count = 0;
LOG.debug("Opened appender {} for {}", appender, finalPath);
this.state = ReaderWriterState.OPEN;
}
@Override
public final void initialize() {
Preconditions.checkState(
state.equals(ReaderWriterState.NEW), "Unable to open a writer from state:%s", state);
ValidationException.check(
isSupportedFormat(descriptor), "Not a supported format: %s", descriptor.getFormat());
// ensure the directory exists
try {
fs.mkdirs(directory);
} catch (RuntimeException e) {
this.state = ReaderWriterState.ERROR;
throw new DatasetOperationException(e, "Failed to create path %s", directory);
} catch (IOException ex) {
this.state = ReaderWriterState.ERROR;
throw new DatasetIOException("Failed to create path " + directory, ex);
}
// initialize paths
try {
this.finalPath = new Path(directory, uniqueFilename(descriptor.getFormat()));
this.tempPath = tempFilename(finalPath);
} catch (RuntimeException e) {
this.state = ReaderWriterState.ERROR;
throw new DatasetOperationException(e, "Failed to initialize file paths under %s", directory);
}
try {
this.appender = newAppender(tempPath);
appender.open();
} catch (RuntimeException e) {
this.state = ReaderWriterState.ERROR;
throw new DatasetOperationException(e, "Failed to open appender %s", appender);
} catch (IOException e) {
this.state = ReaderWriterState.ERROR;
throw new DatasetIOException("Failed to open appender " + appender, e);
}
this.count = 0;
LOG.info("Opened output appender {} for {}", appender, finalPath);
this.state = ReaderWriterState.OPEN;
}
@Override
public void merge(FileSystemDataset<E> update) {
DatasetDescriptor updateDescriptor = update.getDescriptor();
if (!updateDescriptor.getFormat().equals(descriptor.getFormat())) {
throw new DatasetRepositoryException(
"Cannot merge dataset format "
+ updateDescriptor.getFormat()
+ " with format "
+ descriptor.getFormat());
}
if (updateDescriptor.isPartitioned() != descriptor.isPartitioned()) {
throw new DatasetRepositoryException(
"Cannot merge an unpartitioned dataset with a " + " partitioned one or vice versa.");
} else if (updateDescriptor.isPartitioned()
&& descriptor.isPartitioned()
&& !updateDescriptor.getPartitionStrategy().equals(descriptor.getPartitionStrategy())) {
throw new DatasetRepositoryException(
"Cannot merge dataset partition strategy "
+ updateDescriptor.getPartitionStrategy()
+ " with "
+ descriptor.getPartitionStrategy());
}
if (!updateDescriptor.getSchema().equals(descriptor.getSchema())) {
throw new DatasetRepositoryException(
"Cannot merge dataset schema "
+ updateDescriptor.getFormat()
+ " with schema "
+ descriptor.getFormat());
}
Set<String> addedPartitions = Sets.newHashSet();
for (Path path : update.pathIterator()) {
URI relativePath = update.getDirectory().toUri().relativize(path.toUri());
Path newPath = new Path(directory, new Path(relativePath));
Path newPartitionDirectory = newPath.getParent();
try {
if (!fileSystem.exists(newPartitionDirectory)) {
fileSystem.mkdirs(newPartitionDirectory);
}
logger.debug("Renaming {} to {}", path, newPath);
boolean renameOk = fileSystem.rename(path, newPath);
if (!renameOk) {
throw new DatasetException(
"Dataset merge failed during rename of " + path + " to " + newPath);
}
} catch (IOException e) {
throw new DatasetIOException("Dataset merge failed", e);
}
if (descriptor.isPartitioned() && partitionListener != null) {
String partition = newPartitionDirectory.toString();
if (!addedPartitions.contains(partition)) {
partitionListener.partitionAdded(name, partition);
addedPartitions.add(partition);
}
}
}
}
/**
* Checks that the {@code existing} {@link DatasetDescriptor} is compatible with {@code test}.
*
* @param existing the current {@code DatasetDescriptor} for a dataset
* @param test a new {@code DatasetDescriptor} for the same dataset
*/
public static void checkCompatible(DatasetDescriptor existing, DatasetDescriptor test) {
checkNotChanged("format", existing.getFormat(), test.getFormat());
checkNotChanged("partitioning", existing.isPartitioned(), test.isPartitioned());
if (existing.isPartitioned()) {
checkNotChanged(
"partition strategy", existing.getPartitionStrategy(), test.getPartitionStrategy());
}
// check can read records written with old schema using new schema
Schema oldSchema = existing.getSchema();
Schema testSchema = test.getSchema();
if (!SchemaValidationUtil.canRead(oldSchema, testSchema)) {
throw new IncompatibleSchemaException(
"Schema cannot read data "
+ "written using existing schema. Schema: "
+ testSchema.toString(true)
+ "\nExisting schema: "
+ oldSchema.toString(true));
}
}
static <E> FileSystemWriter<E> newWriter(FileSystem fs, Path path, DatasetDescriptor descriptor) {
Format format = descriptor.getFormat();
if (Formats.PARQUET.equals(format)) {
// by default, Parquet is not durable
if (DescriptorUtil.isDisabled(FileSystemProperties.NON_DURABLE_PARQUET_PROP, descriptor)) {
return new IncrementalWriter<E>(fs, path, descriptor);
} else {
return new FileSystemWriter<E>(fs, path, descriptor);
}
} else if (Formats.AVRO.equals(format) || Formats.CSV.equals(format)) {
return new IncrementalWriter<E>(fs, path, descriptor);
} else {
return new FileSystemWriter<E>(fs, path, descriptor);
}
}
static <E> PartitionedDatasetWriter<E, ?> newWriter(FileSystemView<E> view) {
DatasetDescriptor descriptor = view.getDataset().getDescriptor();
Format format = descriptor.getFormat();
if (Formats.PARQUET.equals(format)) {
// by default, Parquet is not durable
if (DescriptorUtil.isDisabled(FileSystemProperties.NON_DURABLE_PARQUET_PROP, descriptor)) {
return new IncrementalPartitionedDatasetWriter<E>(view);
} else {
return new NonDurablePartitionedDatasetWriter<E>(view);
}
} else if (Formats.AVRO.equals(format) || Formats.CSV.equals(format)) {
return new IncrementalPartitionedDatasetWriter<E>(view);
} else {
return new NonDurablePartitionedDatasetWriter<E>(view);
}
}
@Override
public void initialize() {
Preconditions.checkState(
state.equals(ReaderWriterState.NEW), "Unable to open a writer from state:%s", state);
DatasetDescriptor descriptor = view.getDataset().getDescriptor();
ValidationException.check(
FileSystemWriter.isSupportedFormat(descriptor),
"Not a supported format: %s",
descriptor.getFormat());
LOG.debug("Opening partitioned dataset writer w/strategy:{}", partitionStrategy);
cachedWriters =
CacheBuilder.newBuilder()
.maximumSize(maxWriters)
.removalListener(new DatasetWriterCloser<E>())
.build(createCacheLoader());
state = ReaderWriterState.OPEN;
}
@SuppressWarnings("unchecked")
private <E> FileAppender<E> newAppender(Path temp) {
Format format = descriptor.getFormat();
if (Formats.PARQUET.equals(format)) {
// by default, guarantee durability with the more costly writer
if (DescriptorUtil.isEnabled(FileSystemProperties.NON_DURABLE_PARQUET_PROP, descriptor)) {
return (FileAppender<E>) new ParquetAppender(fs, temp, descriptor.getSchema(), conf, true);
} else {
return (FileAppender<E>)
new DurableParquetAppender(fs, temp, descriptor.getSchema(), conf, true);
}
} else if (Formats.AVRO.equals(format)) {
return new AvroAppender<E>(fs, temp, descriptor.getSchema(), true);
} else if (Formats.CSV.equals(format)
&& DescriptorUtil.isEnabled(FileSystemProperties.ALLOW_CSV_PROP, descriptor)) {
return new CSVAppender<E>(fs, temp, descriptor);
} else {
this.state = ReaderWriterState.ERROR;
throw new DatasetWriterException("Unknown format " + descriptor);
}
}
@Override
public <E> Dataset<E> update(String name, DatasetDescriptor descriptor) {
Preconditions.checkArgument(name != null, "Dataset name cannot be null");
Preconditions.checkArgument(descriptor != null, "DatasetDescriptro cannot be null");
DatasetDescriptor oldDescriptor = metadataProvider.load(name);
// oldDescriptor is valid if load didn't throw NoSuchDatasetException
if (!oldDescriptor.getFormat().equals(descriptor.getFormat())) {
throw new DatasetRepositoryException(
"Cannot change dataset format from "
+ oldDescriptor.getFormat()
+ " to "
+ descriptor.getFormat());
}
final URI oldLocation = oldDescriptor.getLocation();
if ((oldLocation != null) && !(oldLocation.equals(descriptor.getLocation()))) {
throw new DatasetRepositoryException("Cannot change the dataset's location");
}
if (oldDescriptor.isPartitioned() != descriptor.isPartitioned()) {
throw new DatasetRepositoryException(
"Cannot change an unpartitioned dataset to " + " partitioned or vice versa.");
} else if (oldDescriptor.isPartitioned()
&& descriptor.isPartitioned()
&& !oldDescriptor.getPartitionStrategy().equals(descriptor.getPartitionStrategy())) {
throw new DatasetRepositoryException(
"Cannot change partition strategy from "
+ oldDescriptor.getPartitionStrategy()
+ " to "
+ descriptor.getPartitionStrategy());
}
// check can read records written with old schema using new schema
final Schema oldSchema = oldDescriptor.getSchema();
final Schema newSchema = descriptor.getSchema();
if (!SchemaValidationUtil.canRead(oldSchema, newSchema)) {
throw new IncompatibleSchemaException(
"New schema cannot read data "
+ "written using "
+ "old schema. New schema: "
+ newSchema.toString(true)
+ "\nOld schema: "
+ oldSchema.toString(true));
}
DatasetDescriptor updatedDescriptor = metadataProvider.update(name, descriptor);
updatedDescriptor = addRepositoryUri(updatedDescriptor);
logger.debug(
"Updated dataset:{} schema:{} datasetPath:{}",
new Object[] {
name, updatedDescriptor.getSchema(), updatedDescriptor.getLocation().toString()
});
return new FileSystemDataset.Builder()
.name(name)
.configuration(conf)
.descriptor(updatedDescriptor)
.partitionKey(
updatedDescriptor.isPartitioned()
? org.kitesdk.data.impl.Accessor.getDefault().newPartitionKey()
: null)
.partitionListener(getPartitionListener())
.build();
}
static boolean isSupportedFormat(DatasetDescriptor descriptor) {
Format format = descriptor.getFormat();
return (SUPPORTED_FORMATS.contains(format)
|| (Formats.CSV.equals(format)
&& DescriptorUtil.isEnabled(FileSystemProperties.ALLOW_CSV_PROP, descriptor)));
}