[ci] Retry .NET SDK provisioning with exponential backoff

[simonrozsival]

Problem

The install .NET SDK to bin//dotnet pipeline step intermittently fails with transient CDN network errors, e.g.:

curl: (56) Recv failure: Connection reset by peer
##[error]Bash exited with code '56'.

Example: https://dev.azure.com/dnceng-public/public/_build/results?buildId=1508469

There was no retry/backoff around the download or install, so a single dropped connection to the .NET CDN failed the whole job.

Fix

Add retry-with-exponential-backoff (5 attempts, 2s → 4s → 8s → 16s) around both network operations in the provisioning scripts — the dotnet-install script download and the SDK install itself (which also downloads from the CDN):

• eng/install-dotnet.sh: new retry() helper wraps the dotnet-install.sh download and the SDK install invocation. curl also gains --retry 5 --retry-delay 1 --retry-all-errors (so it retries connection resets on its own) and an EXIT trap to clean up the temp file.
• eng/install-dotnet.ps1: new Invoke-WithRetry helper wraps Invoke-WebRequest and the install-script invocation, which now throws on a non-zero exit code so retries actually trigger.

Testing

• bash -n and pwsh syntax checks pass on both scripts.
• Unit-tested the bash retry helper: succeeds when the command eventually passes, gives up after 5 attempts with a non-zero exit otherwise.

[Copilot]

Pull request overview

This PR aims to harden CI/local provisioning of the repo-local .NET SDK (bin/<Configuration>/dotnet) against transient network/CDN failures by adding retry logic with exponential backoff around both the dotnet-install script download and the SDK install step.

Changes:

• Add an exponential-backoff retry helper to the Bash provisioning script and wrap both network operations.
• Add an exponential-backoff retry helper to the PowerShell provisioning script and ensure non-zero exit codes trigger retries.

Show a summary per file

File	Description
eng/install-dotnet.sh	Adds retry() helper and wraps curl download + dotnet-install.sh invocation with exponential backoff.
eng/install-dotnet.ps1	Adds Invoke-WithRetry helper and wraps Invoke-WebRequest + dotnet-install.ps1 invocation with exponential backoff + explicit non-zero exit handling.

Copilot's findings

• Files reviewed: 2/2 changed files
• Comments generated: 3

[simonrozsival]

Addressed the review findings in a follow-up commit:

1. curl --retry-all-errors portability (install-dotnet.sh): Removed curl's own --retry/--retry-delay/--retry-all-errors flags. --retry-all-errors isn't supported by the older system curl on some macOS images. The outer retry() helper already backs off on any failure, including transport errors like curl: (56) Recv failure, so nothing is lost.

2. Invoke-WebRequest -UseBasicParsing on PowerShell 7+ (install-dotnet.ps1): Now edition-gated — -UseBasicParsing is only passed on Windows PowerShell (Desktop, 5.1) where it's needed to avoid the IE engine; it's omitted on PowerShell 7+ (Core) where basic parsing is the only mode.

3. Large unrelated deletions: That was a bad diff from the initial push (a --no-checkout worktree produced a commit whose tree contained only the 2 files). The branch was rebuilt on origin/main and force-pushed before this review round — the PR now correctly shows only the 2 intended files changed.

[jonathanpeppers]

/review

[github-actions]

✅ Android PR Reviewer completed successfully!

[github-actions]

🤖 Code Review — ✅ LGTM (with minor suggestions)

Wraps both network operations in eng/install-dotnet.sh and eng/install-dotnet.ps1 with a 5-attempt exponential-backoff retry to absorb transient .NET CDN failures. The implementation is correct and well-commented.

Verified:

• Bash — retry() composes safely with set -euo pipefail (the if "$@" guard suppresses errexit per-attempt), the curl -o tmp + atomic mv cache pattern is preserved, and the EXIT trap is correctly scoped and cleared after success. Removing curl's own --retry-all-errors (unsupported on older macOS curl) in favor of the outer retry() is reasonable.
• PowerShell — Invoke-WithRetry rethrows after $MaxAttempts and the install-script wrapper converts a non-zero $LASTEXITCODE into a throw so retries actually trigger. Edition-gating -UseBasicParsing (Desktop only) is correct. Backoff schedule (2s→4s→8s→16s) matches the comments.
• CI — All 44 dotnet-android checks are green.

Suggestions (non-blocking): 2 💡 inline comments — retry scope for non-transient failures, and PowerShell try-block indentation.

Nice, focused reliability fix. 👍

Issues: ❌ 0 · ⚠️ 0 · 💡 2

Generated by Android PR Reviewer for #12120 · 77.3 AIC · ⌖ 13 AIC · ⊞ 6.8K
_{Comment /review to run again}

[github-actions]

🤖 💡 Patterns — retry() retries on any non-zero exit, so a genuinely deterministic failure (bad SDK version → curl -f HTTP 404/exit 22, or dotnet-install.sh rejecting an invalid --version) now burns all 5 attempts and ~30s of backoff before surfacing. That's an acceptable tradeoff for a provisioning script, but consider narrowing retries to transient transport errors (or at least noting in the header comment that non-transient failures are also retried) so future readers aren't surprised by the delay on a hard failure.

(Rule: Retry scope — non-transient failures)

[github-actions]

🤖 💡 Formatting — The body inside this try { ... } block (lines 67–79) is indented one level shallower than the surrounding try/finally (compare Move-Item on the next block, which is at 4 spaces). Re-indenting the $iwrArgs hashtable and Invoke-WithRetry block to match keeps the try body visually nested. Cosmetic only.

(Rule: Consistent indentation)