The replicator operates by extracting (or applying) an entire transaction. This means that when extracting data from the binary log, and writing that to THL, or extracting from the THL in preparation for applying to the target, the entire transaction, or an entire statement within a multi-statement transaction, must be held in memory.

In the event of a very large transaction having to be extracted, this can cause a problem with the memory configuration. The actual configuration of how much memory is used is determined through a combination of the number of fragments, the size of the internal buffer used to store those fragments, and the overall fragment size.

Although you can increase the overall memory allocated to the replicator, changing the internal sizes used can also improve the performance and ability to extract data.

First, try reducing the size of the buffer ( replicator.global.buffer.size ) used to hold the transaction fragments. The default for this value is 10, but reducing this to 5 or less will ease the required memory:

replicator.global.buffer.size=10

Altering the size of each fragment can also help, as it reduces the memory required to hold the data before it is written to disk and sent out over the network to Replica replicators. Reducing the fragment size will reduce the memory footprint. The size is controlled by the replicator.extractor.dbms.transaction_frag_size parameter:

replicator.extractor.dbms.transaction_frag_size=1000000

Note that if you change the fragment size, you may need to reset the service on the extractor so that the binary log is parsed again. You can reset the service by using the trepctl reset command.