[
https://issues.apache.org/jira/browse/BEAM-14540?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Brian Hulette updated BEAM-14540:
---------------------------------
Description:
With https://s.apache.org/batched-dofns (BEAM-14213), we want to encourage
users to develop pipelines that process arrow data within the Python SDK, but
communicating batches of data across SDKs or from SDK to Runner is left as
future work. So when Arrow data is processed in the SDK, it must be converted
to/from Rows for transmission over the Fn API. So the current ideal Python
execution looks like:
1. read row oriented data over the Fn API, deserialize with SchemaCoder
2. Buffer rows and construct an arrow RecordBatch/Table object
3. Perform user computation(s)
4. Explode output RecordBatch/Table into rows
5. Serialize rows with SchemaCoder and write out over the Fn API
Note that (1,2) and (4,5) will exist in every stage of the user's pipeline, and
they'll also exist when Python transforms (e.g. dataframe read_csv) are used in
other SDKs. We should improve performance for this hot path by making a native
(cythonized) implementation for (1,2) and (4,5).
was:
With https://s.apache.org/batched-dofns (BEAM-14213), we want to encourage
users to develop pipelines that process arrow data within the Python SDK, but
communicating batches of data across SDKs or from SDK to Runner is left as
future work. So when Arrow data is processed in the SDK, it must be converted
to/from Rows for transmission over the Fn API. So the ideal Python execution
looks like:
1. read row oriented data over the Fn API, deserialize with SchemaCoder
2. Buffer rows and construct an arrow RecordBatch/Table object
3. Perform user computation
4. Explode output RecordBatch/Table into rows
5. Serialize rows with SchemaCoder and write out over the Fn API
We can improve performance for this type of flow by making a native
(cythonized) implementation for (1,2) and (4,5).
> Native implementation for serialized Rows to/from Arrow
> -------------------------------------------------------
>
> Key: BEAM-14540
> URL: https://issues.apache.org/jira/browse/BEAM-14540
> Project: Beam
> Issue Type: Improvement
> Components: sdk-py-core
> Reporter: Brian Hulette
> Priority: P2
>
> With https://s.apache.org/batched-dofns (BEAM-14213), we want to encourage
> users to develop pipelines that process arrow data within the Python SDK, but
> communicating batches of data across SDKs or from SDK to Runner is left as
> future work. So when Arrow data is processed in the SDK, it must be converted
> to/from Rows for transmission over the Fn API. So the current ideal Python
> execution looks like:
> 1. read row oriented data over the Fn API, deserialize with SchemaCoder
> 2. Buffer rows and construct an arrow RecordBatch/Table object
> 3. Perform user computation(s)
> 4. Explode output RecordBatch/Table into rows
> 5. Serialize rows with SchemaCoder and write out over the Fn API
> Note that (1,2) and (4,5) will exist in every stage of the user's pipeline,
> and they'll also exist when Python transforms (e.g. dataframe read_csv) are
> used in other SDKs. We should improve performance for this hot path by making
> a native (cythonized) implementation for (1,2) and (4,5).
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
