Stream Decorator Reference

The cartesianProduct function turns a single tuple with a multi-valued field (i.e., an array) into multiple tuples, one for each value in the array field. That is, given a single tuple containing an array of N values for fieldA, the cartesianProduct function will output N tuples, each with one value from the original tuple’s array. In essence, you can flatten arrays for further processing.

For example, using cartesianProduct you can turn this tuple:

into the following 3 tuples:

The classify function classifies tuples using a logistic regression text classification model. It was designed specifically to work with models trained using the train function. The classify function uses the model function to retrieve a stored model and then scores a stream of tuples using the model. The tuples read by the classifier must contain a text field that can be used for classification. The classify function uses a Lucene analyzer to extract the features from the text so the model can be applied. By default the classify function looks for the analyzer using the name of text field in the tuple. If the Solr schema on the worker node does not contain this field, the analyzer can be looked up in another field by specifying the analyzerField parameter.

Each tuple that is classified is assigned two scores:

The commit function wraps a single stream (A) and given a collection and batch size will send commit messages to the collection when the batch size is fulfilled or the end of stream is reached. A commit stream is used most frequently with an update stream and as such the commit will take into account possible summary tuples coming from the update stream. All tuples coming into the commit stream will be returned out of the commit stream - no tuples will be dropped and no tuples will be added.

The complement function wraps two streams (A and B) and emits tuples from A which do not exist in B. The tuples are emitted in the order in which they appear in stream A. Both streams must be sorted by the fields being used to determine equality (using the on parameter).

The daemon function wraps another function and runs it at intervals using an internal thread. The daemon function can be used to provide both continuous push and pull streaming.

The delete function wraps other functions and uses the id and _version_ values found to send the tuples to a SolrCloud collection as Delete By Id commands.

This is similar to the update() function described below.

The eval function allows for use cases where new streaming expressions are generated on the fly and then evaluated. The eval function wraps a streaming expression and reads a single tuple from the underlying stream. The eval function then retrieves a string streaming expression from the expr_s field of the tuple. The eval function then compiles the string streaming expression and emits the tuples.

The executor function wraps a stream source that contains streaming expressions, and executes the expressions in parallel. The executor function looks for the expression in the expr_s field in each tuple. The executor function has an internal thread pool that runs tasks that compile and run expressions in parallel on the same worker node. This function can also be parallelized across worker nodes by wrapping it in the parallel function to provide parallel execution of expressions across a cluster.

The executor function does not do anything specific with the output of the expressions that it runs. Therefore the expressions that are executed must contain the logic for pushing tuples to their destination. The update function can be included in the expression being executed to send the tuples to a SolrCloud collection for storage.

This model allows for asynchronous execution of jobs where the output is stored in a SolrCloud collection where it can be accessed as the job progresses.

The fetch function iterates a stream and fetches additional fields and adds them to the tuples. The fetch function fetches in batches to limit the number of calls back to Solr. Tuples streamed from the fetch function will contain the original fields and the additional fields that were fetched. The fetch function supports one-to-one fetches. Many-to-one fetches, where the stream source contains duplicate keys, will also work, but one-to-many fetches are currently not supported by this function.

The having expression wraps a stream and applies a boolean operation to each tuple. It emits only tuples for which the boolean operation returns true.

The leftOuterJoin function wraps two streams, Left and Right, and emits tuples from Left. If there is a tuple in Right equal (as defined by on) then the values in that tuple will be included in the emitted tuple. An equal tuple in Right need not exist for the Left tuple to be emitted. This supports one-to-one, one-to-many, many-to-one, and many-to-many left outer join scenarios. The tuples are emitted in the order in which they appear in the Left stream. Both streams must be sorted by the fields being used to determine equality (using the on parameter). If both tuples contain a field of the same name then the value from the Right stream will be used in the emitted tuple.

You can wrap the incoming streams with a select function to be specific about which field values are included in the emitted tuple.

The hashJoin function wraps two streams, Left and Right, and for every tuple in Left which exists in Right will emit a tuple containing the fields of both tuples. This supports one-to-one, one-to-many, many-to-one, and many-to-many inner join scenarios. The tuples are emitted in the order in which they appear in the Left stream. The order of the streams does not matter. If both tuples contain a field of the same name then the value from the Right stream will be used in the emitted tuple.

You can wrap the incoming streams with a select function to be specific about which field values are included in the emitted tuple.

The hashJoin function can be used when the tuples of Left and Right cannot be put in the same order. Because the tuples are out of order this stream functions by reading all values from the Right stream during the open operation and will store all tuples in memory. The result of this is a memory footprint equal to the size of the Right stream.

Wraps two streams, Left and Right. For every tuple in Left which exists in Right a tuple containing the fields of both tuples will be emitted. This supports one-to-one, one-to-many, many-to-one, and many-to-many inner join scenarios. The tuples are emitted in the order in which they appear in the Left stream. Both streams must be sorted by the fields being used to determine equality (the 'on' parameter). If both tuples contain a field of the same name then the value from the Right stream will be used in the emitted tuple. You can wrap the incoming streams with a select(…​) expression to be specific about which field values are included in the emitted tuple.

The intersect function wraps two streams, A and B, and emits tuples from A which DO exist in B. The tuples are emitted in the order in which they appear in stream A. Both streams must be sorted by the fields being used to determine equality (the on parameter). Only tuples from A are emitted.

The list function wraps N Stream Expressions and opens and iterates each stream sequentially. This has the effect of concatenating the results of multiple Streaming Expressions.

The merge function merges two or more streaming expressions and maintains the ordering of the underlying streams. Because the order is maintained, the sorts of the underlying streams must line up with the on parameter provided to the merge function.

The null expression is a useful utility function for understanding bottlenecks when performing parallel relational algebra (joins, intersections, rollups etc.). The null function reads all the tuples from an underlying stream and returns a single tuple with the count and processing time. Because the null stream adds minimal overhead of its own, it can be used to isolate the performance of Solr’s /export handler. If the /export handlers performance is not the bottleneck, then the bottleneck is likely occurring in the workers where the stream decorators are running.

The null expression can be wrapped by the parallel function and sent to worker nodes. In this scenario each worker will return one tuple with the count of tuples processed on the worker and the timing information for that worker. This gives valuable information such as:

The outerHashJoin function wraps two streams, Left and Right, and emits tuples from Left. If there is a tuple in Right equal (as defined by the on parameter) then the values in that tuple will be included in the emitted tuple. An equal tuple in Right need not exist for the Left tuple to be emitted. This supports one-to-one, one-to-many, many-to-one, and many-to-many left outer join scenarios. The tuples are emitted in the order in which they appear in the Left stream. The order of the streams does not matter. If both tuples contain a field of the same name then the value from the Right stream will be used in the emitted tuple.

You can wrap the incoming streams with a select function to be specific about which field values are included in the emitted tuple.

The outerHashJoin stream can be used when the tuples of Left and Right cannot be put in the same order. Because the tuples are out of order, this stream functions by reading all values from the Right stream during the open operation and will store all tuples in memory. The result of this is a memory footprint equal to the size of the Right stream.

The parallel function wraps a streaming expression and sends it to N worker nodes to be processed in parallel.

The parallel function requires that the partitionKeys parameter be provided to the underlying searches. The partitionKeys parameter will partition the search results (tuples) across the worker nodes. Tuples with the same values as partitionKeys will be shuffled to the same worker nodes.

The parallel function maintains the sort order of the tuples returned by the worker nodes, so the sort criteria must incorporate the sort order of the tuples returned by the workers.

For example if you sort on year, month and day you could partition on year only as long as there are enough different years to spread the tuples around the worker nodes.

Solr allows sorting on more than 4 fields, but you cannot specify more than 4 partitionKeys for speed considerations. Also it’s overkill to specify many partitionKeys when we one or two keys could be enough to spread the tuples.

Parallel stream was designed when the underlying search stream will emit a lot of tuples from the collection. If the search stream only emits a small subset of the data from the collection using parallel could potentially be slower.

The plist function wraps N Stream Expressions and opens the streams in parallel and iterates each stream sequentially. The difference between the list and plist is that the streams are opened in parallel. Since many streams such as facet, stats and significantTerms push down heavy operations to Solr when they are opened, the plist function can dramatically improve performance by doing these operations in parallel.

The priority function is a simple priority scheduler for the executor function. The executor function doesn’t directly have a concept of task prioritization; instead it simply executes tasks in the order that they are read from its underlying stream. The priority function provides the ability to schedule a higher priority task ahead of lower priority tasks that were submitted earlier.

The priority function wraps two topic functions that are both emitting tuples that contain streaming expressions to execute. The first topic is considered the higher priority task queue.

Each time the priority function is called, it checks the higher priority task queue to see if there are any tasks to execute. If tasks are waiting in the higher priority queue then the priority function will emit the higher priority tasks. If there are no high priority tasks to run, the lower priority queue tasks are emitted.

The priority function will only emit a batch of tasks from one of the queues each time it is called. This ensures that no lower priority tasks are executed until the higher priority queue has no tasks to run.

The reduce function wraps an internal stream and groups tuples by common fields.

Each tuple group is operated on as a single block by a pluggable reduce operation. The group operation provided with Solr implements distributed grouping functionality. The group operation also serves as an example reduce operation that can be referred to when building custom reduce operations.

The rollup function wraps another stream function and rolls up aggregates over bucket fields. The rollup function relies on the sort order of the underlying stream to rollup aggregates one grouping at a time. Accordingly, the sort order of the underlying stream must match the fields in the over parameter of the rollup function.

The rollup function also needs to process entire result sets in order to perform its aggregations. When the underlying stream is the search function, the /export handler can be used to provide full sorted result sets to the rollup function. This sorted approach allows the rollup function to perform aggregations over very high cardinality fields. The disadvantage of this approach is that the tuples must be sorted and streamed across the network to a worker node to be aggregated. For faster aggregation over low to moderate cardinality fields, the facet function can be used.

See section in graph traversal.

The select function wraps a streaming expression and outputs tuples containing a subset or modified set of fields from the incoming tuples. The list of fields included in the output tuple can contain aliases to effectively rename fields. The select stream supports both operations and evaluators. One can provide a list of operations and evaluators to perform on any fields, such as replace, add, if, etc.

The sort function wraps a streaming expression and re-orders the tuples. The sort function emits all incoming tuples in the new sort order. The sort function reads all tuples from the incoming stream, re-orders them using an algorithm with O(nlog(n)) performance characteristics, where n is the total number of tuples in the incoming stream, and then outputs the tuples in the new sort order. Because all tuples are read into memory, the memory consumption of this function grows linearly with the number of tuples in the incoming stream.

The top function wraps a streaming expression and re-orders the tuples. The top function emits only the top N tuples in the new sort order. The top function re-orders the underlying stream so the sort criteria does not have to match up with the underlying stream.

The unique function wraps a streaming expression and emits a unique stream of tuples based on the over parameter. The unique function relies on the sort order of the underlying stream. The over parameter must match up with the sort order of the underlying stream.

The unique function implements a non-co-located unique algorithm. This means that records with the same unique over field do not need to be co-located on the same shard. When executed in the parallel, the partitionKeys parameter must be the same as the unique over field so that records with the same keys will be shuffled to the same worker.

The update function wraps another functions and sends the tuples to a SolrCloud collection for indexing as Documents.