Introduction

The MergeRecord Processor allows the user to take many FlowFiles that consist of record-oriented data (any data format for which there is a Record Reader available) and combine the FlowFiles into one larger FlowFile. This may be preferable before pushing the data to a downstream system that prefers larger batches of data, such as HDFS, or in order to improve performance of a NiFi flow by reducing the number of FlowFiles that flow through the system (thereby reducing the contention placed on the FlowFile Repository, Provenance Repository, Content Repository, and FlowFile Queues).

The Processor creates several 'bins' to put the FlowFiles in. The maximum number of bins to use is set to 5 by default, but this can be changed by updating the value of the <Maximum number of Bins> property. The number of bins is bound in order to avoid running out of Java heap space. Note: while the contents of a FlowFile are stored in the Content Repository and not in the Java heap space, the Processor must hold the FlowFile objects themselves in memory. As a result, these FlowFiles with their attributes can potentially take up a great deal of heap space and cause OutOfMemoryError's to be thrown. In order to avoid this, if you expect to merge many small FlowFiles together, it is advisable to instead use a MergeContent that merges no more than say 1,000 FlowFiles into a bundle and then use a second MergeContent to merges these small bundles into larger bundles. For example, to merge 1,000,000 FlowFiles together, use MergeRecord that uses a <Maximum Number of Records> of 1,000 and route the "merged" Relationship to a second MergeRecord that also sets the <Maximum Number of Records> to 1,000. The second MergeRecord will then merge 1,000 bundles of 1,000, which in effect produces bundles of 1,000,000.

How FlowFiles are Binned

How the Processor determines which bin to place a FlowFile in depends on a few different configuration options. Firstly, the Merge Strategy is considered. The Merge Strategy can be set to one of two options: Bin Packing Algorithm, or Defragment. When the goal is to simply combine smaller FlowFiles into one larger FlowFiles, the Bin Packing Algorithm should be used. This algorithm picks a bin based on whether or not the FlowFile can fit in the bin according to its size and the <Maximum Bin Size> property and whether or not the FlowFile is 'like' the other FlowFiles in the bin. What it means for two FlowFiles to be 'like FlowFiles' is discussed at the end of this section.

The "Defragment" Merge Strategy can be used when records need to be explicitly assigned to the same bin. For example, if data is split apart using the SplitRecord Processor, each 'split' can be processed independently and later merged back together using this Processor with the Merge Strategy set to Defragment. In order for FlowFiles to be added to the same bin when using this configuration, the FlowFiles must have the same value for the "fragment.identifier" attribute. Each FlowFile with the same identifier must also have the same value for the "fragment.count" attribute (which indicates how many FlowFiles belong in the bin) and a unique value for the "fragment.index" attribute so that the FlowFiles can be ordered correctly.

In order to be added to the same bin, two FlowFiles must be 'like FlowFiles.' In order for two FlowFiles to be like FlowFiles, they must have the same schema, and if the <Correlation Attribute Name> property is set, they must have the same value for the specified attribute. For example, if the <Correlation Attribute Name> is set to "filename" then two FlowFiles must have the same value for the "filename" attribute in order to be binned together. If more than one attribute is needed in order to correlate two FlowFiles, it is recommended to use an UpdateAttribute processor before the MergeRecord processor and combine the attributes. For example, if the goal is to bin together two FlowFiles only if they have the same value for the "abc" attribute and the "xyz" attribute, then we could accomplish this by using UpdateAttribute and adding a property with name "correlation.attribute" and a value of "abc=${abc},xyz=${xyz}" and then setting MergeRecord's <Correlation Attribute Name> property to "correlation.attribute".

It is often useful to bin together only Records that have the same value for some field. For example, if we have point-of-sale data, perhaps the desire is to bin together records that belong to the same store, as identified by the 'storeId' field. This can be accomplished by making use of the PartitionRecord Processor ahead of MergeRecord. This Processor will allow one or more fields to be configured as the partitioning criteria and will create attributes for those corresponding values. An UpdateAttribute processor could then be used, if necessary, to combine multiple attributes into a single correlation attribute, as described above. See documentation for those processors for more details.

When a Bin is Merged

Above, we discussed how a bin is chosen for a given FlowFile. Once a bin has been created and FlowFiles added to it, we must have some way to determine when a bin is "full" so that we can bin those FlowFiles together into a "merged" FlowFile. There are a few criteria that are used to make a determination as to whether or not a bin should be merged.

If the <Merge Strategy> property is set to "Bin Packing Algorithm" then then the following rules will be evaluated. Firstly, in order for a bin to be full, both of the thresholds specified by the <Minimum Bin Size> and the <Minimum Number of Records> properties must be satisfied. If one of these properties is not set, then it is ignored. Secondly, if either the <Maximum Bin Size> or the <Maximum Number of Records> property is reached, then the bin is merged. That is, both of the minimum values must be reached but only one of the maximum values need be reached. Note that the <Maximum Number of Records> property is a "soft limit," meaning that all records in a given input FlowFile will be added to the same bin, and as a result the number of records may exceed the maximum configured number of records. Once this happens, though, no more Records will be added to that same bin from another FlowFile. If the <Max Bin Age> is reached for a bin, then the FlowFiles in that bin will be merged, even if the minimum bin size and minimum number of records have not yet been met. Finally, if the maximum number of bins have been created (as specified by the <Maximum number of Bins> property), and some input FlowFiles cannot fit into any of the existing bins, then the oldest bin will be merged to make room. This is done because otherwise we would not be able to add any additional FlowFiles to the existing bins and would have to wait until the Max Bin Age is reached (if ever) in order to merge any FlowFiles.

If the <Merge Strategy> property is set to "Defragment" then a bin is full only when the number of FlowFiles in the bin is equal to the number specified by the "fragment.count" attribute of one of the FlowFiles in the bin. All FlowFiles that have this attribute must have the same value for this attribute, or else they will be routed to the "failure" relationship. It is not necessary that all FlowFiles have this value, but at least one FlowFile in the bin must have this value or the bin will never be complete. If all of the necessary FlowFiles are not binned together by the point at which the bin times amount (as specified by the <Max Bin Age> property), then the FlowFiles will all be routed to the 'failure' relationship instead of being merged together.

Once a bin is merged into a single FlowFile, it can sometimes be useful to understand why exactly the bin was merged when it was. For example, if the maximum number of allowable bins is reached, a merged FlowFile may consist of far fewer records than expected. In order to help understand the behavior, the Processor will emit a JOIN Provenance Events when creating the merged FlowFile, and the JOIN event will include in it a "Details" field that explains why the bin was merged when it was. For example, the event will indicate "Records Merged due to: Bin is full" if the bin reached its minimum thresholds and no more subsequent FlowFiles were able to be added to it. Or it may indicate "Records Merged due to: Maximum number of bins has been exceeded" if the bin was merged due to the configured maximum number of bins being filled and needing to free up space for a new bin.

When a Failure Occurs

When a bin is filled, the Processor is responsible for merging together all of the records in those FlowFiles into a single FlowFile. If the Processor fails to do so for any reason (for example, a Record cannot be read from an input FlowFile), then all of the FlowFiles in that bin are routed to the 'failure' Relationship. The Processor does not skip the single problematic FlowFile and merge the others. This behavior was chosen because of two different considerations. Firstly, without those problematic records, the bin may not truly be full, as the minimum bin size may not be reached without those records. Secondly, and more importantly, if the problematic FlowFile contains 100 "good" records before the problematic ones, those 100 records would already have been written to the "merged" FlowFile. We cannot un-write those records. If we were to then send those 100 records on and route the problematic FlowFile to 'failure' then in a situation where the "failure" relationship is eventually routed back to MergeRecord, we could end up continually duplicating those 100 successfully processed records.

Examples

To better understand how this Processor works, we will lay out a few examples. For the sake of simplicity of these examples, we will use CSV-formatted data and write the merged data as CSV-formatted data, but the format of the data is not really relevant, as long as there is a Record Reader that is capable of reading the data and a Record Writer capable of writing the data in the desired format.

Example 1 - Batching Together Many Small FlowFiles

When we want to batch together many small FlowFiles in order to create one larger FlowFile, we will accomplish this by using the "Bin Packing Algorithm" Merge Strategy. The idea here is to bundle together as many FlowFiles as we can within our minimum and maximum number of records and bin size. Consider that we have the following properties set:

Property Name Property Value
Merge Strategy Bin Packing Algorithm
Minimum Number of Records 3
Maximum Number of Records 5

Also consider that we have the following data on the queue, with the schema indicating a Name and an Age field:

FlowFile ID FlowFile Contents
1 Mark, 33
2 John, 45
Jane, 43
3 Jake, 3
4 Jan, 2

In this, because we have not configured a Correlation Attribute, and because all FlowFiles have the same schema, the Processor will attempt to add all of these FlowFiles to the same bin. Because the Minimum Number of Records is 3 and the Maximum Number of Records is 5, all of the FlowFiles will be added to the same bin. The output, then, is a single FlowFile with the following content:

Mark, 33
John, 45
Jane, 43
Jake, 3
Jan, 2

When the Processor runs, it will bin all of the FlowFiles that it can get from the queue. After that, it will merge any bin that is "full enough." So if we had only 3 FlowFiles on the queue, those 3 would have been added, and a new bin would have been created in the next iteration, once the 4th FlowFile showed up. However, if we had 8 FlowFiles queued up, only 5 would have been added to the first bin. The other 3 would have been added to a second bin, and that bin would then be merged since it reached the minimum threshold of 3 also.