为了让前面规划的互联网+数据平台能有效对电子商务数据进行管理及实现大数据统计功能,必须在平台上再增加一个MongDB-Engine:数据平台用户通过传入一种Context来指示MongoDB-Engine运算。与前面JDBC-Engine和Cassandra-Engine通过文本式传递指令不同的是:MangoDB没有一套像SQL或CQL这样的文本式编程语言。但MangoDB基本上都是通过Bson类型的参数进行运算的,Bson是个java interface:
/**
* An interface for types that are able to render themselves into a [email protected] BsonDocument}.
*
* @since 3.0
*/
public interface Bson {
/**
* Render the filter into a BsonDocument.
*
* @param documentClass the document class in scope for the collection. This parameter may be ignored, but it may be used to alter
* the structure of the returned [email protected] BsonDocument} based on some knowledge of the document class.
* @param codecRegistry the codec registry. This parameter may be ignored, but it may be used to look up [email protected] Codec} instances for
* the document class or any other related class.
* @param <TDocument> the type of the document class
* @return the BsonDocument
*/
<TDocument> BsonDocument toBsonDocument(Class<TDocument> documentClass, CodecRegistry codecRegistry);
}
任何实现Bson的类型都可以通过toBsonDocument来进行Bson到Document的转换。下面是Filter类型的例子:
private static final class OperatorFilter<TItem> implements Bson {
private final String operatorName;
private final String fieldName;
private final TItem value;
OperatorFilter(final String operatorName, final String fieldName, final TItem value) {
this.operatorName = notNull("operatorName", operatorName);
this.fieldName = notNull("fieldName", fieldName);
this.value = value;
}
@Override
public <TDocument> BsonDocument toBsonDocument(final Class<TDocument> documentClass, final CodecRegistry codecRegistry) {
BsonDocumentWriter writer = new BsonDocumentWriter(new BsonDocument());
writer.writeStartDocument();
writer.writeName(fieldName);
writer.writeStartDocument();
writer.writeName(operatorName);
encodeValue(writer, value, codecRegistry);
writer.writeEndDocument();
writer.writeEndDocument();
return writer.getDocument();
}
@Override
public String toString() {
return operatorFilterToString(fieldName, operatorName, value);
}
}
对于MongoDB-Engine,我们需要实现下面这几大类功能才能满足平台要求,包括:
count
disctinct
aggregate
mapReduce
bulkWrite
insert
delete
replace
update
具体函数细节和功能描述如下:
count:
/**
* Counts the number of documents in the collection.
* @return a Observable with a single element indicating the number of documents
*/
def count(): SingleObservable[Long] = observeLong(wrapped.count(_: SingleResultCallback[java.lang.Long]))
/**
* Counts the number of documents in the collection according to the given options.
* @param filter the query filter
* @param options the options describing the count
* @return a Observable with a single element indicating the number of documents
*/
def count(filter: Bson, options: CountOptions): SingleObservable[Long] =
observeLong(wrapped.count(filter, options, _: SingleResultCallback[java.lang.Long]))
distinct:
/**
* Gets the distinct values of the specified field name.
* @param fieldName the field name
* @tparam C the target type of the observable.
* @return a Observable emitting the sequence of distinct values
*/
def distinct[C](fieldName: String)(implicit ct: ClassTag[C]): DistinctObservable[C] =
DistinctObservable(wrapped.distinct(fieldName, ct))
/**
* Gets the distinct values of the specified field name.
* @param fieldName the field name
* @param filter the query filter
* @tparam C the target type of the observable.
* @return a Observable emitting the sequence of distinct values
*/
def distinct[C](fieldName: String, filter: Bson)(implicit ct: ClassTag[C]): DistinctObservable[C] =
DistinctObservable(wrapped.distinct(fieldName, filter, ct))
find:
/**
* Finds all documents in the collection.
* @tparam C the target document type of the observable.
* @return the find Observable
*/
def find[C]()(implicit e: C DefaultsTo TResult, ct: ClassTag[C]): FindObservable[C] =
FindObservable(wrapped.find[C](ct))
/**
* Finds all documents in the collection.
* @param filter the query filter
* @tparam C the target document type of the observable.
* @return the find Observable
*/
def find[C](filter: Bson)(implicit e: C DefaultsTo TResult, ct: ClassTag[C]): FindObservable[C] =
FindObservable(wrapped.find(filter, ct))
aggregate:
/**
* Aggregates documents according to the specified aggregation pipeline.
* @param pipeline the aggregate pipeline
* @return a Observable containing the result of the aggregation operation
* [[http://docs.mongodb.org/manual/aggregation/ Aggregation]]
*/
def aggregate[C](pipeline: Seq[Bson])(implicit e: C DefaultsTo TResult, ct: ClassTag[C]): AggregateObservable[C] =
AggregateObservable(wrapped.aggregate[C](pipeline.asJava, ct))
mapReduce:
/**
* Aggregates documents according to the specified map-reduce function.
* @param mapFunction A JavaScript function that associates or "maps" a value with a key and emits the key and value pair.
* @param reduceFunction A JavaScript function that "reduces" to a single object all the values associated with a particular key.
* @tparam C the target document type of the observable.
* @return a Observable containing the result of the map-reduce operation
*/
def mapReduce[C](mapFunction: String, reduceFunction: String)(implicit e: C DefaultsTo TResult, ct: ClassTag[C]): MapReduceObservable[C] =
MapReduceObservable(wrapped.mapReduce(mapFunction, reduceFunction, ct))
bulkWrite:
/**
* Executes a mix of inserts, updates, replaces, and deletes.
* @param requests the writes to execute
* @return a Observable with a single element the BulkWriteResult
*/
def bulkWrite(requests: Seq[_ <: WriteModel[_ <: TResult]]): SingleObservable[BulkWriteResult] =
observe(wrapped.bulkWrite(
requests.asJava.asInstanceOf[util.List[_ <: WriteModel[_ <: TResult]]],
_: SingleResultCallback[BulkWriteResult]
))
/**
* Executes a mix of inserts, updates, replaces, and deletes.
* @param requests the writes to execute
* @param options the options to apply to the bulk write operation
* @return a Observable with a single element the BulkWriteResult
*/
def bulkWrite(requests: Seq[_ <: WriteModel[_ <: TResult]], options: BulkWriteOptions): SingleObservable[BulkWriteResult] =
observe(wrapped.bulkWrite(
requests.asJava.asInstanceOf[util.List[_ <: WriteModel[_ <: TResult]]],
options,
_: SingleResultCallback[BulkWriteResult]
))
insert:
/**
* Inserts the provided document. If the document is missing an identifier, the driver should generate one.
* @param document the document to insert
* @return a Observable with a single element indicating when the operation has completed or with either a
* com.mongodb.DuplicateKeyException or com.mongodb.MongoException
*/
def insertOne(document: TResult): SingleObservable[Completed] = observeCompleted(wrapped.insertOne(document, _: SingleResultCallback[Void]))
/**
* Inserts the provided document. If the document is missing an identifier, the driver should generate one.
* @param document the document to insert
* @param options the options to apply to the operation
* @return a Observable with a single element indicating when the operation has completed or with either a
* com.mongodb.DuplicateKeyException or com.mongodb.MongoException
*/
def insertOne(document: TResult, options: InsertOneOptions): SingleObservable[Completed] =
observeCompleted(wrapped.insertOne(document, options, _: SingleResultCallback[Void]))
/**
* Inserts a batch of documents. The preferred way to perform bulk inserts is to use the BulkWrite API. However, when talking with
* server 2.6, using this method will be faster due to constraints in the bulk API related to error handling.
* @param documents the documents to insert
* @return a Observable with a single element indicating when the operation has completed or with either a
* com.mongodb.DuplicateKeyException or com.mongodb.MongoException
*/
def insertMany(documents: Seq[_ <: TResult]): SingleObservable[Completed] =
observeCompleted(wrapped.insertMany(documents.asJava, _: SingleResultCallback[Void]))
/**
* Inserts a batch of documents. The preferred way to perform bulk inserts is to use the BulkWrite API. However, when talking with
* server 2.6, using this method will be faster due to constraints in the bulk API related to error handling.
* @param documents the documents to insert
* @param options the options to apply to the operation
* @return a Observable with a single element indicating when the operation has completed or with either a
* com.mongodb.DuplicateKeyException or com.mongodb.MongoException
*/
def insertMany(documents: Seq[_ <: TResult], options: InsertManyOptions): SingleObservable[Completed] =
observeCompleted(wrapped.insertMany(documents.asJava, options, _: SingleResultCallback[Void]))
delete:
/**
* Removes at most one document from the collection that matches the given filter. If no documents match, the collection is not
* modified.
* @param filter the query filter to apply the the delete operation
* @return a Observable with a single element the DeleteResult or with an com.mongodb.MongoException
*/
def deleteOne(filter: Bson): SingleObservable[DeleteResult] = observe(wrapped.deleteOne(filter, _: SingleResultCallback[DeleteResult]))
/**
* Removes at most one document from the collection that matches the given filter. If no documents match, the collection is not
* modified.
* @param filter the query filter to apply the the delete operation
* @param options the options to apply to the delete operation
* @return a Observable with a single element the DeleteResult or with an com.mongodb.MongoException
*/
def deleteOne(filter: Bson, options: DeleteOptions): SingleObservable[DeleteResult] =
observe(wrapped.deleteOne(filter, options, _: SingleResultCallback[DeleteResult]))
/**
* Removes all documents from the collection that match the given query filter. If no documents match, the collection is not modified.
* @param filter the query filter to apply the the delete operation
* @return a Observable with a single element the DeleteResult or with an com.mongodb.MongoException
*/
def deleteMany(filter: Bson): SingleObservable[DeleteResult] = observe(wrapped.deleteMany(filter, _: SingleResultCallback[DeleteResult]))
/**
* Removes all documents from the collection that match the given query filter. If no documents match, the collection is not modified.
* @param filter the query filter to apply the the delete operation
* @param options the options to apply to the delete operation
* @return a Observable with a single element the DeleteResult or with an com.mongodb.MongoException
*/
def deleteMany(filter: Bson, options: DeleteOptions): SingleObservable[DeleteResult] =
observe(wrapped.deleteMany(filter, options, _: SingleResultCallback[DeleteResult]))
replace:
/**
* Replace a document in the collection according to the specified arguments.
* [[http://docs.mongodb.org/manual/tutorial/modify-documents/#replace-the-document Replace]]
* @param filter the query filter to apply the the replace operation
* @param replacement the replacement document
* @return a Observable with a single element the UpdateResult
*/
def replaceOne(filter: Bson, replacement: TResult): SingleObservable[UpdateResult] =
observe(wrapped.replaceOne(filter, replacement, _: SingleResultCallback[UpdateResult]))
/**
* Replace a document in the collection according to the specified arguments.
* @param filter the query filter to apply the the replace operation
* @param replacement the replacement document
* @param options the options to apply to the replace operation
* @return a Observable with a single element the UpdateResult
*/
def replaceOne(filter: Bson, replacement: TResult, options: UpdateOptions): SingleObservable[UpdateResult] =
observe(wrapped.replaceOne(filter, replacement, options, _: SingleResultCallback[UpdateResult]))
update:
/**
* Update a single document in the collection according to the specified arguments.
* @param filter a document describing the query filter, which may not be null. This can be of any type for which a `Codec` is
* registered
* @param update a document describing the update, which may not be null. The update to apply must include only update operators. This
* can be of any type for which a `Codec` is registered
* @return a Observable with a single element the UpdateResult
*/
def updateOne(filter: Bson, update: Bson): SingleObservable[UpdateResult] =
observe(wrapped.updateOne(filter, update, _: SingleResultCallback[UpdateResult]))
/**
* Update a single document in the collection according to the specified arguments.
* @param filter a document describing the query filter, which may not be null. This can be of any type for which a `Codec` is
* registered
* @param update a document describing the update, which may not be null. The update to apply must include only update operators. This
* can be of any type for which a `Codec` is registered
* @param options the options to apply to the update operation
* @return a Observable with a single element the UpdateResult
*/
def updateOne(filter: Bson, update: Bson, options: UpdateOptions): SingleObservable[UpdateResult] =
observe(wrapped.updateOne(filter, update, options, _: SingleResultCallback[UpdateResult]))
/**
* Update documents in the collection according to the specified arguments.
* @param filter a document describing the query filter, which may not be null. This can be of any type for which a `Codec` is
* registered
* @param update a document describing the update, which may not be null. The update to apply must include only update operators. This
* can be of any type for which a `Codec` is registered
* @return a Observable with a single element the UpdateResult
*/
def updateMany(filter: Bson, update: Bson): SingleObservable[UpdateResult] =
observe(wrapped.updateMany(filter, update, _: SingleResultCallback[UpdateResult]))
可以看到:函数传入参数大致两个类型:Bson、XXOptions。我们还需要适度增加一些数据库管理功能,包括:
createCollection
listCollection
dropCollection
createIndex
dropIndex
createView
具体函数细节和功能描述如下:
dropCollection
/**
* Drops this collection from the Database.
*/
def drop(): SingleObservable[Completed] = observeCompleted(wrapped.drop(_: SingleResultCallback[Void]))
createIndex
/**
* @param key an object describing the index key(s), which may not be null. This can be of any type for which a `Codec` is
* registered
* @return a Observable with a single element indicating when the operation has completed
*/
def createIndex(key: Bson): SingleObservable[String] =
observe(wrapped.createIndex(key, _: SingleResultCallback[String]))
/**
* @param key an object describing the index key(s), which may not be null. This can be of any type for which a `Codec` is
* registered
* @param options the options for the index
* @return a Observable with a single element indicating when the operation has completed
*/
def createIndex(key: Bson, options: IndexOptions): SingleObservable[String] =
observe(wrapped.createIndex(key, options, _: SingleResultCallback[String]))
dropIndex
/**
* Drops the given index.
* @param indexName the name of the index to remove
* @return a Observable with a single element indicating when the operation has completed
*/
def dropIndex(indexName: String): SingleObservable[Completed] = observeCompleted(wrapped.dropIndex(indexName, _: SingleResultCallback[Void]))
/**
* Drops the given index.
* @param indexName the name of the index to remove
* @param dropIndexOptions options to use when dropping indexes
* @return a Observable with a single element indicating when the operation has completed
*/
def dropIndex(indexName: String, dropIndexOptions: DropIndexOptions): SingleObservable[Completed] =
observeCompleted(wrapped.dropIndex(indexName, dropIndexOptions, _: SingleResultCallback[Void]))
/**
* Drops the index given the keys used to create it.
* @param keys the keys of the index to remove
* @return a Observable with a single element indicating when the operation has completed
*/
def dropIndex(keys: Bson): SingleObservable[Completed] = observeCompleted(wrapped.dropIndex(keys, _: SingleResultCallback[Void]))
/**
* Drops the index given the keys used to create it.
* @param keys the keys of the index to remove
* @param dropIndexOptions options to use when dropping indexes
* @return a Observable with a single element indicating when the operation has completed
*/
def dropIndex(keys: Bson, dropIndexOptions: DropIndexOptions): SingleObservable[Completed] =
observeCompleted(wrapped.dropIndex(keys, dropIndexOptions, _: SingleResultCallback[Void]))
/**
* Drop all the indexes on this collection, except for the default on _id.
* @return a Observable with a single element indicating when the operation has completed
*/
def dropIndexes(): SingleObservable[Completed] =
observeCompleted(wrapped.dropIndexes(_: SingleResultCallback[Void]))
/**
* Drop all the indexes on this collection, except for the default on _id.
* @param dropIndexOptions options to use when dropping indexes
* @return a Observable with a single element indicating when the operation has completed
*/
def dropIndexes(dropIndexOptions: DropIndexOptions): SingleObservable[Completed] =
observeCompleted(wrapped.dropIndexes(dropIndexOptions, _: SingleResultCallback[Void]))
listCollection
/**
* Finds all the collections in this database.
* @tparam TResult the target document type of the iterable.
* @return the fluent list collections interface
*/
def listCollections[TResult]()(implicit e: TResult DefaultsTo Document, ct: ClassTag[TResult]): ListCollectionsObservable[TResult] = ListCollectionsObservable(wrapped.listCollections(ct))
createCollection
/**
* Create a new collection with the given name.
* @param collectionName the name for the new collection to create
* @return a Observable identifying when the collection has been created
*/
def createCollection(collectionName: String): SingleObservable[Completed] =
observeCompleted(wrapped.createCollection(collectionName, _: SingleResultCallback[Void]))
/**
* Create a new collection with the selected options
* @param collectionName the name for the new collection to create
* @param options various options for creating the collection
* @return a Observable identifying when the collection has been created
*/
def createCollection(collectionName: String, options: CreateCollectionOptions): SingleObservable[Completed] =
observeCompleted(wrapped.createCollection(collectionName, options, _: SingleResultCallback[Void]))
createView
/**
* Creates a view with the given name, backing collection/view name, and aggregation pipeline that defines the view.
* @param viewName the name of the view to create
* @param viewOn the backing collection/view for the view
* @param pipeline the pipeline that defines the view
*/
def createView(viewName: String, viewOn: String, pipeline: Seq[Bson]): SingleObservable[Completed] =
observeCompleted(wrapped.createView(viewName, viewOn, pipeline.asJava, _: SingleResultCallback[Void]))
/**
* Creates a view with the given name, backing collection/view name, aggregation pipeline, and options that defines the view.
* @param viewName the name of the view to create
* @param viewOn the backing collection/view for the view
* @param pipeline the pipeline that defines the view
* @param createViewOptions various options for creating the view
*/
def createView(viewName: String, viewOn: String, pipeline: Seq[Bson], createViewOptions: CreateViewOptions): SingleObservable[Completed] =
observeCompleted(wrapped.createView(viewName, viewOn, pipeline.asJava, createViewOptions, _: SingleResultCallback[Void]))
下面就是根据以上需求分析初步做出的功能框架设计方案:
import org.bson.conversions.Bson
import org.mongodb.scala._
trait MGOCommands
object MGOCommands {
case class Count(filter: Option[Bson], options: Option[Any]) extends MGOCommands
case class Distict(fieldName: String, filter: Option[Bson]) extends MGOCommands
case class Find(filter: Option[Bson]) extends MGOCommands
case class Aggregate(pipeLine: Seq[Bson]) extends MGOCommands
case class MapReduce(mapFunction: String, reduceFunction: String) extends MGOCommands
case class Insert(newdocs: Seq[Document], options: Option[Any]) extends MGOCommands
case class Delete(filter: Bson, options: Option[Any], onlyOne: Boolean = false) extends MGOCommands
case class Replace(filter: Bson, replacement: Document, options: Option[Any]) extends MGOCommands
case class Update(filter: Bson, update: Bson, options: Option[Any]) extends MGOCommands
case class BulkWrite(commands: Seq[MGOCommands], options: Option[Any]) extends MGOCommands
}
object MGOAdmins {
case class DropCollection(collName: String) extends MGOCommands
case class CreateCollection(collName: String, options: Option[Any]) extends MGOCommands
case class ListCollection(dbName: String) extends MGOCommands
case class CreateView(viewName: String, viewOn: String, pipeline: Seq[Bson], options: Option[Any]) extends MGOCommands
case class CreateIndex(filter: Option[Bson], options: Option[Any]) extends MGOCommands
case class DropIndexByName(indexName: String, options: Option[Any]) extends MGOCommands
case class DropIndexByKey(key: Bson, options: Option[Any]) extends MGOCommands
case class DropAllIndexes(options: Option[Any]) extends MGOCommands
}
case class MGOContext (
dbName: String,
collName: String,
action: MGOCommands = null
) { ctx =>
def setDbName(name: String): MGOContext = ctx.copy(dbName = name)
def setCollName(name: String): MGOContext = ctx.copy(collName = name)
def setCommand(cmd: MGOCommands): MGOContext = ctx.copy(action = cmd)
}
object MGOContext {
def apply(db: String, coll: String) = new MGOContext(db,coll)
def apply(db: String, coll: String, command: MGOCommands) =
new MGOContext(db,coll,command)
}
原创文章,作者:ItWorker,如若转载,请注明出处:https://blog.ytso.com/12807.html