-
Notifications
You must be signed in to change notification settings - Fork 4k
Description
Currently if we want to read data from ORC data into Arrow Record Batch in Java runtime, we needs to first use the ORC Java reader to load data into memory then convert it to Arrow RecordBatch.
However, since ORC Java Reader only read orc data into on heap memory, while Arrow Record only support off heap memory on Java, memory copy in unavoidable in conversion process. In our internal benchmark, the conversion time can take up to 25% E2E latency when running tpch q1.
To workaround this overhead, a method is to enable the Java runtime directly reading data from native ORC c++ reader, which will load data directly into off heap memory and only pointer manipulation and schema ser/de would be involved in the conversion process.
Reporter: Yurui Zhou / @yuruiz
Assignee: Yurui Zhou / @yuruiz
Subtasks:
PRs and other links:
Note: This issue was originally created as ARROW-4714. Please see the migration documentation for further details.