Batch size best practices

When processing data for Text Analytics with Auto Importers, Medallia Experience Cloud splits the data into smaller chunks, each of which is processed as one unit of work.

Several factors influence the system's behavior:

Signal size (record size) is a significant factor. Bigger signals require more time and memory to process, making database updates slower.
Signal count also has an impact on the behavior of Auto Importers, because operations like de-duplication, and data access are run simultaneously when batch size is too big.
The number of concurrent Auto Importers is also a factor. When there are simultaneous operations in progress, the same records can be accessed from different sources. This causes the entire batch to be retried and processing time is lost. A smaller batch size makes the batch process faster, reducing the chances of running a concurrent operation on a record. Additionally, retrying smaller batches is less expensive.

Manage these factors via the size of batches on the farm parameter. For information about this parameter and other processing options, see Text Analytics processing and initialization.


Batch size	Recommended use cases	Notes
Around 500 records per batch (small)	Use for clients loading records with big signal size, or clients that are using a lot of Auto Importers at the same time, which could cause concurrency issues. This batch size is also well suited for clients that do not ingest too many records, preferable under the batch size threshold. Use for big signal clients, were a lot of data is affected by updates, when loading one signal can require MBs of memory. Best concurrency support	Smaller batches enhance concurrency, because in cases when one record is accessed by two different web feeds, the system will roll one of the batches back: if the batch is small enough, reprocessing the whole batch does not have an impact.
Between 1,000 and 2,000 records per batch (medium)	Use for clients with moderate size number of Auto Importers or medium size signals. Ideal for clients with medium signal size or when number of web feeds is moderate.	When the number of records to ingest is high, work with bigger batches to accelerate record ingestion.
5,000 records per batch (default)	Use for new clients or scenarios when the workload is not clearly defined. Use for clients that do not have too many Auto Importers running concurrently. Good option when the signal size is small.	Most performant batch size.

Note: Larger batch sizes can be implemented and may prove beneficial in specific scenarios that warrant individual analysis.