@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);
}
}
}
}
@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 <E> Dataset<E> create(String name, DatasetDescriptor descriptor) {
Preconditions.checkArgument(name != null, "Name can not be null");
Preconditions.checkArgument(descriptor != null, "Descriptor can not be null");
Preconditions.checkArgument(
descriptor.getLocation() == null,
"Descriptor location cannot be set; " + "it is assigned by the MetadataProvider");
DatasetDescriptor newDescriptor = metadataProvider.create(name, descriptor);
newDescriptor = addRepositoryUri(newDescriptor);
final URI location = newDescriptor.getLocation();
if (location == null) {
throw new DatasetRepositoryException(
"[BUG] MetadataProvider did not assign a location to dataset:" + name);
}
ensureExists(newDescriptor, conf);
logger.debug(
"Created dataset:{} schema:{} datasetPath:{}",
new Object[] {name, newDescriptor.getSchema(), location.toString()});
return new FileSystemDataset.Builder()
.name(name)
.configuration(conf)
.descriptor(newDescriptor)
.partitionKey(
newDescriptor.isPartitioned()
? org.kitesdk.data.impl.Accessor.getDefault().newPartitionKey()
: null)
.partitionListener(getPartitionListener())
.build();
}
private GetStorageKey(View<E> view) {
DatasetDescriptor descriptor = view.getDataset().getDescriptor();
// get serializable versions of transient objects
this.strategyString = descriptor.getPartitionStrategy().toString(false /* no white space */);
this.schemaString = descriptor.getSchema().toString(false /* no white space */);
this.type = view.getType();
if (view instanceof AbstractRefinableView) {
this.constraints = ((AbstractRefinableView) view).getConstraints().toQueryMap();
} else {
this.constraints = null;
}
}
@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);
}
}
/**
* 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));
}
}
/**
* Precondition-style validation that the DatasetDescriptor is compatible.
*
* @param descriptor a {@link DatasetDescriptor}
*/
public static void checkDescriptor(DatasetDescriptor descriptor) {
Preconditions.checkNotNull(descriptor, "Descriptor cannot be null");
Schema schema = descriptor.getSchema();
checkSchema(schema);
if (descriptor.isPartitioned()) {
// marked as [BUG] because this is checked in DatasetDescriptor
Preconditions.checkArgument(
schema.getType() == Schema.Type.RECORD,
"[BUG] Partitioned datasets must have record schemas");
Set<String> names = Sets.newHashSet();
for (Schema.Field field : schema.getFields()) {
names.add(field.name());
}
List<String> incompatible = Lists.newArrayList();
List<String> duplicates = Lists.newArrayList();
for (FieldPartitioner fp : descriptor.getPartitionStrategy().getFieldPartitioners()) {
String name = fp.getName();
if (!isCompatibleName(name)) {
incompatible.add(name);
} else if (names.contains(name)) {
duplicates.add(name);
} else {
names.add(name);
}
}
Preconditions.checkState(
incompatible.isEmpty(),
"Hive incompatible: partition names are not alphanumeric (plus '_'): %s",
Joiner.on(", ").join(incompatible));
Preconditions.checkState(
duplicates.isEmpty(),
"Hive incompatible: partition names duplicate data fields: %s",
Joiner.on(", ").join(duplicates));
}
}
@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();
}
