[
https://issues.apache.org/jira/browse/FLINK-39631?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18079278#comment-18079278
]
featzhang commented on FLINK-39631:
-----------------------------------
I would like to work on this sub-task under the FLINK-39625 umbrella. Could a
committer please assign it to me (Jira username: featzhang)? Thanks!
> Integrate GPU sidecar inference with Table/SQL and provide end-to-end example
> -----------------------------------------------------------------------------
>
> Key: FLINK-39631
> URL: https://issues.apache.org/jira/browse/FLINK-39631
> Project: Flink
> Issue Type: Sub-task
> Components: Documentation, Examples, Table SQL / API
> Reporter: featzhang
> Priority: Major
> Labels: gpu, model-inference
>
> h2. Background
> Streaming SQL users should be able to invoke the GPU sidecar without
> writing DataStream code. Flink's existing scalar and table-function UDF
> surfaces are the natural place to expose inference: users declare the
> function once via DDL, then reference it from arbitrary SQL statements.
> This sub-task closes the loop by delivering a user-facing Table / SQL
> integration plus an end-to-end example and documentation.
> h2. Scope of this sub-task
> * Add {{GpuInferScalarFunction}} in {{flink-table-api-java-bridge}} (or
> equivalent) that forwards inputs to the sidecar via the async client
> from the earlier sub-task. The function is marked asynchronous so that
> the planner generates an async-lookup-style execution plan.
> * Register the function via standard Flink UDF DDL; an optional SQL
> hint selects a non-default sidecar endpoint for A / B scenarios.
> * Add an end-to-end example under {{flink-examples-streaming}} that reads
> from a generator source, runs inference, and writes to a print sink.
> The example runs out of the box against the mock sidecar.
> * Add user documentation under {{docs/content/docs}} covering
> configuration, deployment topology, metrics, and a short
> troubleshooting checklist.
> h2. Out of scope
> * Planner-level optimisations such as auto-batching of SQL projections
> into a single sidecar call. That can be tackled once real workloads
> surface concrete patterns.
> h2. Acceptance criteria
> * SQL client session can create the function and run
> {{SELECT gpu_infer(x) FROM src}} end-to-end against the mock sidecar.
> * Documentation builds without warnings and the example is referenced
> from it.
> * Coverage of the new UDF on CI: at least one integration test under
> {{flink-table-planner}}.
> h2. Affected modules
> * {{flink-table-api-java-bridge}}
> * {{flink-table-planner}}
> * {{flink-examples-streaming}}
> * {{docs}}
> h2. Links
> Parent: see umbrella issue linked to this sub-task.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
