feat(execd): tune jupyter idle polling and sse completion wait

[skyler0513]

Summary

This PR improves execd's Jupyter execution stream handling and SSE tail latency behavior.

Changes included:

• add configurable Jupyter idle polling interval via EXECD_JUPYTER_IDLE_POLL_INTERVAL / --jupyter-idle-poll-interval
• replace the fixed 300ms post-idle polling sleep with the configured poll interval
• update RunCode and RunInSession to wait for OnExecuteComplete as a best-effort signal instead of always sleeping for a fixed graceful shutdown window

Background

This change was motivated by an execution latency regression observed on code execution nodes.

Based on local observation:

• without sandbox, execution latency was around 60ms
• with sandbox enabled, execution latency increased to around 2.4s

After tracing the slow path, the added tail latency was mainly caused by fixed waiting behavior in the execd execution flow. With this change, the observed latency was
reduced to around 100ms in the same scenario.

Motivation

Previously:

• Jupyter post-idle polling used a hard-coded 300ms interval
• SSE handlers always paid a fixed tail latency because the controller slept unconditionally after execution

This change makes the Jupyter idle polling behavior configurable and reduces unnecessary fixed response tail latency for SSE endpoints.

Testing

Executed under components/execd:

• go test ./pkg/flag
• go test ./pkg/jupyter/execute
• go test ./...
• go test -race ./...
• golangci-lint run -v ./...

Results:

• go test ./pkg/flag: pass
• go test ./pkg/jupyter/execute: pass
• go test ./...: fails on existing pkg/web/controller TestReadMetrics
• go test -race ./...: fails on existing pkg/web/controller TestReadMetrics and an existing race in pkg/runtime
• golangci-lint run -v ./...: reports existing gocognit complexity issue at pkg/jupyter/execute/execute.go:117

Notes

• The repository currently has pre-existing test/lint issues unrelated to this change.
• This PR only modifies components/execd.

[Copilot]

Pull request overview

This PR reduces execd’s SSE tail latency and makes Jupyter “idle” post-processing polling configurable, targeting an observed execution latency regression in sandboxed environments.

Changes:

• Add a configurable Jupyter idle polling interval via EXECD_JUPYTER_IDLE_POLL_INTERVAL / --jupyter-idle-poll-interval.
• Replace the fixed 300ms post-idle polling sleep with the configured interval in Jupyter streaming execution.
• Update SSE endpoints (RunCode, RunInSession) to wait for OnExecuteComplete (best-effort) instead of always sleeping a fixed graceful shutdown window.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 8 comments.

File	Description
components/execd/pkg/web/controller/codeinterpreting.go	Switch SSE tail behavior from unconditional sleep to best-effort completion wait.
components/execd/pkg/jupyter/execute/execute.go	Make post-idle polling interval configurable rather than hard-coded.
components/execd/pkg/flag/parser.go	Add env/CLI parsing for the new Jupyter idle poll interval.
components/execd/pkg/flag/flags.go	Add the new global flag variable for Jupyter idle polling interval.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Pangjiping]

Could you fix those go linter error by make golint under components/execd? 😊

[skyler0513]

I fixed two blockers: the golangci-lint failure and the C# E2E error-contract regression.
For lint, I refactored ExecuteCodeStream into smaller helpers (no behavior change) so cognitive complexity now passes. For C# E2E, I fixed Jupyter terminal-event semantics
so failed executions emit error as terminal and do not emit complete, which prevents error events from being dropped on the client side. I also added regression tests for
both paths.

[Pangjiping]

I fixed two blockers: the golangci-lint failure and the C# E2E error-contract regression. For lint, I refactored ExecuteCodeStream into smaller helpers (no behavior change) so cognitive complexity now passes. For C# E2E, I fixed Jupyter terminal-event semantics so failed executions emit error as terminal and do not emit complete, which prevents error events from being dropped on the client side. I also added regression tests for both paths.

Thanks for your GREAT submission. I will review it and merge it as soon as possible.
It looks like this solves the issue we’ve long been criticized for — overly conservative connection keep-alive by sleep 😂😂.

[Pangjiping]

LGTM

[skyler0513]

LGTM

Hello,Jiping, What is the latest available execd image version?

[Pangjiping]

LGTM

Hello,Jiping, What is the latest available execd image version?

I’ll take care of it today or tomorrow.