@Rocky Can we keep a connection open for doing HTTP Stream Load on a FE?

In StarRocks, the behavior for keeping connections open during HTTP Stream Load depends on whether you are talking about the TCP connection (Keep-Alive) or the data stream itself. 1. HTTP Keep-Alive (Persistent TCP Connections) Yes, you can keep a TCP connection open to the FE using standard HTTP Keep-Alive. However, there are important architectural details to consider: * Redirect (307): By default, when you send a Stream Load request to the FE, it returns an HTTP 307 Temporary Redirect to a specific BE (the coordinator for that job). * Client Behavior: Most clients (like

curl --location-trusted

) will then close the FE connection or return it to a pool and open a new connection to the redirected BE to transmit the actual data. * Efficiency: Keeping a connection open to the FE is useful for the initial "handshake" or for sending many separate small load requests. However, since the actual data transfer happens with the BE, the connection to the FE is only used for the redirect instruction. 2. FE Proxy (Data Proxying) If your BE nodes are in a private network and only the FE is accessible, you might want the FE to proxy the data instead of redirecting the client. * Kubernetes/Helm: StarRocks provides an FE Proxy component (often an Nginx-based sidecar or service) specifically for this. It allows you to send all data to the proxy endpoint, which handles the internal routing to the BEs. * Arrow Flight SQL: For Arrow Flight-based ingestion (not standard Stream Load), there is a global variable

arrow_flight_proxy_enabled

that allows FE to proxy data. 3. Continuous Streaming (Single PUT) A single HTTP Stream Load request (PUT) is atomic. * Job Completion: The job is committed only when the HTTP stream is closed and the server receives the end of the request. You cannot keep a single PUT request "open forever" to continuously shove data in like a message queue; the system needs to finalize the version and commit the transaction. * Alternative (Transaction Interface): If you need to send data in multiple parts, use the Stream Load Transaction Interface (

/api/transaction/begin

). This allows you to call

/api/transaction/load

multiple times. You can use persistent connections for these subsequent API calls to reduce overhead. 4. Merge Commit (v3.4+) If your goal is to handle many high-frequency small loads without the overhead of opening/closing many connections, you should look into Merge Commit (introduced in v3.4). It allows StarRocks to merge multiple small Stream Load requests into a single internal commit, significantly improving throughput for "streaming-like" workloads. Recommendation * To reduce latency: Connect directly to a BE (or use a Load Balancer in front of BEs) to avoid the 307 Redirect overhead. * For high frequency: Use Merge Commit or the Transaction Interface. * For network isolation: Use the FE Proxy component provided in the Kubernetes Operator or Helm chart. References • quick_start/helm.md • administration/management/FE_configuration.md • administration/management/BE_configuration.md