read-batch-size.md

Understanding readBatchSize

At its core, readBatchSize is the maximum number of messages a single vertex replica (pod) will try to read per cycle.

Default Behavior (Single Partition)

By default, the edge connecting vertices has one partition. In this common setup, a single replica reads from that single partition.

Note: Monovertex always follows this default behavior, as it does not have the concept of partitions.

• Example: If readBatchSize is 500 and you have 1 replica, that replica will read up to 500 messages in each processing cycle. Simple!

Behavior with Multi-Partitioned Edges

When you configure a vertex to have multiple partitions (not applicable for Source Vertices), you increase its ability to process data in parallel. The key thing to understand is that the readBatchSize limit applies to the read operation on each partition that a replica is responsible for.

This means the total number of messages a single replica can process per cycle becomes: $$\text{Total Messages per Replica} = \text{readBatchSize} \times \text{Number of Partitions}$$

• Example: Let's say you configure a vertex to have 3 partitions and it's running with 1 replica. If readBatchSize is 500:
  • The single replica will read up to 500 messages from Partition 1.
  • It will also read up to 500 messages from Partition 2 and 3.
  • In total, that single replica can now process up to 1,500 messages per cycle (500 * 3).