When using Postgres decodebuf plugin, tt seems a lot of time is consumed in TypeMetadataImpl() constructor, which is called from PostgresChangeRecordEmitter.schemaChanged(). TypeMetadataImpl() parses column modifiers using regular expression and as it's called for almost every message, lots to time is spend in this part of the code during processing of the records.

This is a flame graph for load generated by pgbench tool with highlighted part consumed by checking for schema change: