Skip to content
129 changes: 129 additions & 0 deletions docs/build_cache_requirements.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,129 @@
# Build Cache Service — Requirements for Crank Integration

## Context

Crank (the .NET benchmarking tool) has been updated to support a new `buildcache` channel that downloads pre-built runtime binaries from the Build Cache Service (BCS) instead of resolving versions from VMR/NuGet feeds. This gives per-commit granularity for performance testing and regression bisection.

The crank-side changes are complete. This document describes what's needed on the BCS/dotnet-performance-infra side to make the integration work end-to-end.

---

## Requirement 1: Public Blob Access

**Status:** Already in progress (per prior discussion).

Crank's BCS client uses unauthenticated HTTP GET requests to download artifacts. The blobs in the `pvscmdupload` storage account's `$web` container need to be publicly readable.

**URLs crank will hit:**

```
GET https://pvscmdupload.z22.web.core.windows.net/builds/{repoName}/latest/{branch}/latestBuilds.json
GET https://pvscmdupload.z22.web.core.windows.net/builds/{repoName}/buildArtifacts/{commitSha}/{configKey}/{artifactFile}
```
Comment on lines +19 to +22

Copilot AI Apr 14, 2026

Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This doc’s example URLs use the blob endpoint with /$web/..., but the agent default base URL (and client URL construction) uses the static website endpoint (https://pvscmdupload.z22.web.core.windows.net) without /$web. Please align the documented URLs with what the agent actually calls (or clarify which base URL should be configured via --build-cache-base-url).

Copilot uses AI. Check for mistakes.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Addressed by 1ecc239 (round 1): the doc URLs in docs/build_cache_requirements.md now match what the agent actually calls — the static-website endpoint https://pvscmdupload.z22.web.core.windows.net/builds/... (no /$web/). The default --build-cache-base-url value is documented and matches.


Where:
- `repoName` = `runtime` or `aspnetcore` (selected per-job via the `buildCacheRepo` property; defaults to the agent's `--build-cache-repo-name`)
- `branch` = e.g., `main`, `release/10.0`
- `configKey` = e.g., `coreclr_x64_linux`, `coreclr_arm64_windows` (runtime); `aspnetcore_x64_linux`, `aspnetcore_arm64_windows` (aspnetcore)
- `artifactFile` = e.g., `BuildArtifacts_linux_x64_Release_coreclr.tar.gz` (runtime); `BuildArtifacts_linux_x64_Release_aspnetcore.nupkg` (aspnetcore — the verbatim runtime-pack nupkg)

The aspnetcore `latestBuilds.json` lives at `builds/aspnetcore/latest/{branch}/latestBuilds.json` and contains only the 5 `aspnetcore_*` config keys plus an `all` entry and `branch_name`; it does **not** carry the runtime `coreclr_*` keys. Crank's parser enumerates keys dynamically, so it tolerates either repo's file.

---

## Requirement 2: Commit Index File (Not Required)

~~Originally proposed as a per-branch `commitIndex.json` mapping commits to timestamps.~~

**Decision:** Not needed. For the default case, `latestBuilds.json` provides the latest commit. For specific-commit runs (e.g., bisection), users will already know the SHAs — either from git history, GitHub, or a local tool that queries the GitHub API for the commit list. A separate index in BCS would be redundant.

If automated bisection tooling is built in the future, it can query GitHub directly for ordered commit SHAs and then check BCS blob existence per-commit.

---

## Requirement 3: latestBuilds.json Compatibility

**Resolved.** The actual `latestBuilds.json` uses PascalCase (`CommitSha`, `CommitTime`), not snake_case. Crank's parser has been updated to accept both casings for forward compatibility.

---

## Requirement 4: Artifact Layout Stability

Crank extracts **runtime** artifacts using this path convention inside the archive:

```
microsoft.netcore.app.runtime.{rid}/Release/runtimes/{rid}/lib/net{X}.0/ → managed DLLs
microsoft.netcore.app.runtime.{rid}/Release/runtimes/{rid}/native/ → native libs
{rid}.Release/corehost/ → host binaries (dotnet, libhostfxr, libhostpolicy)
```

For **aspnetcore** artifacts the stored blob is the **verbatim runtime-pack nupkg** (a zip,
extension `.nupkg`), so the layout is the nupkg's own — `runtimes/{rid}` sits at the archive root,
with no `microsoft.aspnetcore.app.runtime.{rid}/Release` wrapper. Crucially, the verbatim nupkg
carries the host-resolvable framework metadata next to the managed assemblies:

```
runtimes/{rid}/lib/net{X}.0/Microsoft.AspNetCore.*.dll → managed assemblies
runtimes/{rid}/lib/net{X}.0/Microsoft.AspNetCore.App.deps.json → host-resolvable metadata (REQUIRED)
runtimes/{rid}/lib/net{X}.0/Microsoft.AspNetCore.App.runtimeconfig.json → host-resolvable metadata (REQUIRED)
runtimes/{rid}/native/ → native libs (optional)
```

Because the nupkg is a complete framework, crank **places `Microsoft.AspNetCore.App` directly** from
the pack into the per-job dotnet home (the whole managed set incl `deps.json`/`runtimeconfig.json`,
no feed contribution) and **fails the job** (`BuildCacheIncompleteException`) if the pack is missing
managed assemblies, `deps.json`, or `runtimeconfig.json` — for perf runs, erroring is preferable to
silently running mixed/feed bits. The base runtime + host stay feed-resolved (the aspnetcore pack
ships neither). Self-contained (SCD) publishes are the exception: the framework is co-mingled with
the app under the app's own `.deps.json`, so for SCD crank overlays only the managed `*.dll`
(+ native), not the framework metadata.

This differs from **runtime**, whose archive is raw build output (no shared-framework
`deps.json`/`runtimeconfig.json`), so the runtime flavour overlays BCS binaries onto a feed-installed
runtime (reusing the feed's metadata) rather than placing directly.

Where `{rid}` = `linux-x64`, `linux-arm64`, `win-x64`, `win-arm64`, `win-x86`. (aspnetcore v1 has no
musl/osx/arm32 configs.)

The runtime layout was confirmed by inspecting `BuildArtifacts_linux_x64_Release_coreclr.tar.gz` (no
framework metadata in the pack lib). The aspnetcore contract — a verbatim runtime-pack nupkg carrying
managed + `deps.json` + `runtimeconfig.json`, uploaded as `BuildArtifacts_{os}_{arch}_Release_aspnetcore.nupkg`
— was confirmed against a locally-built `Microsoft.AspNetCore.App.Runtime.win-x64.nupkg` and
dotnet/performance#5243's `stage-bcs-nupkg-aspnetcore.ps1`.
**If either layout changes in future builds, the crank extraction will break.** Consider treating it
as a stable contract or documenting it.

---

## Nice-to-Have: Artifact Manifest

A `manifest.json` per commit+config that describes the archive contents would make extraction more robust:

```
builds/{repoName}/buildArtifacts/{commitSha}/{configKey}/manifest.json
```

```json
{
"runtimeVersion": "10.0.0-preview.4.26120.3",
"commitSha": "abc123...",
"rid": "linux-arm64",
"managedPath": "microsoft.netcore.app.runtime.linux-arm64/Release/runtimes/linux-arm64/lib/net10.0",
"nativePath": "microsoft.netcore.app.runtime.linux-arm64/Release/runtimes/linux-arm64/native",
"corehostPath": "linux-arm64.Release/corehost"
}
```

This isn't blocking — crank currently discovers paths by convention — but it would decouple crank from the internal archive layout and make future changes safe.

---

## Summary

| # | Requirement | Priority | Blocking? |
|---|-------------|----------|-----------|
| 1 | Public blob access | High | Yes — crank can't download without it |
| 2 | ~~Commit index~~ | N/A | Dropped — users provide SHAs directly or use GitHub |
| 3 | `latestBuilds.json` field names | N/A | Resolved — crank parser updated to handle PascalCase |
| 4 | Artifact layout stability | Medium | Not now, but breaking changes would break crank |
| 5 | Artifact manifest.json | Low | Nice-to-have for robustness |
76 changes: 75 additions & 1 deletion docs/dotnet_versions.md
Original file line number Diff line number Diff line change
Expand Up @@ -55,9 +55,12 @@ When a TFM is configured, the agent will download the corresponding .NET SDK ver
- `current`: only latest public versions, this is the default
- `latest`: latest versions used by ASP.NET
- `edge`: latest nightly builds available
- `buildcache`: runtime from the Build Cache Service (per-commit builds)

The difference between `latest` and `edge` is that `latest` will pick runtimes and SDKs that are deemed compatible together. For instance a very recent .NET core runtime might be compatible with a less recent ASP.NET runtime. The `edge` is used to pick the absolute latest build for the select TFM.

The `buildcache` channel uses the Build Cache Service (BCS) from `dotnet-performance-infra` to resolve runtime versions by individual commit SHA rather than from VMR feeds. This provides much finer-grained control — every cached runtime commit is available, whereas VMR feeds may have multi-day gaps between ingested commits. SDK and ASP.NET Core versions are resolved from `latest` when using `buildcache`.

In order to benchmark and ASP.NET application using very recent runtimes of .NET 5, the `latest` channel is recommended:

```
Expand Down Expand Up @@ -115,4 +118,75 @@ The following command uses the `edge` channel but ASP.NET is fixed so it doesn't

```
> crank --config /crank/samples/hello/hello.benchmarks.yml --scenario hello --profile local --application.framework netcoreapp5.0 --application.channel edge --application.aspnetCoreVersion 5.0.0-preview.6.20279.12
```
```

## Using the Build Cache channel

The `buildcache` channel resolves pre-built binaries for individual commits from the Build Cache Service (BCS). This is useful for performance regression bisection where VMR feed gaps make it hard to pinpoint which commit caused a regression.

The channel can resolve from one of two repositories, selected per-job with the `buildCacheRepo` property:

- `runtime` (default) — **overlays** the base .NET runtime (`Microsoft.NETCore.App`) with BCS bits built from a [dotnet/runtime](https://github.com/dotnet/runtime) commit. The runtime archive is raw build output (no shared-framework metadata), so BCS binaries are overlaid onto a feed-installed runtime.
- `aspnetcore` — **places** the ASP.NET Core shared framework (`Microsoft.AspNetCore.App`) directly from a [dotnet/aspnetcore](https://github.com/dotnet/aspnetcore) commit's BCS build. The aspnetcore archive is the runtime-pack nupkg stored verbatim (carrying `deps.json` + `runtimeconfig.json`), so the framework folder is built entirely from BCS and the job **fails** if the pack is incomplete. The base runtime stays at the feed-resolved version.

When no `buildCacheRepo` is supplied the job falls back to the agent-level `--build-cache-repo-name` (which itself defaults to `runtime`), so existing runtime usage is unchanged.

### Basic usage (latest cached build on main)

```
# runtime (default)
> crank --config benchmarks.yml --scenario json --profile aspnet-perf-lin --application.channel buildcache
```

### Bisecting ASP.NET Core

Add `--application.buildCacheRepo aspnetcore` to resolve an ASP.NET Core commit instead. The `buildCacheCommitSha` / `buildCacheBranch` values are then interpreted as **dotnet/aspnetcore** commit/branch:

```
# latest cached aspnetcore build on main
> crank --config benchmarks.yml --scenario json --profile aspnet-perf-lin --application.channel buildcache --application.buildCacheRepo aspnetcore

# specific aspnetcore commit
> crank --config benchmarks.yml --scenario json --profile aspnet-perf-lin --application.channel buildcache --application.buildCacheRepo aspnetcore --application.buildCacheCommitSha a1b2c3d4e5f6...
```

> **One BCS component per job (v1).** A single job overrides EITHER the base runtime OR the ASP.NET Core shared framework from BCS, not both at once. The `buildCacheRepo` selector decides which one; the other framework comes from the normal feed. If a job somehow requests both, the selected flavour wins and the other is skipped with a log warning.

### Specific commit SHA

```
> crank --config benchmarks.yml --scenario json --profile aspnet-perf-lin --application.channel buildcache --application.buildCacheCommitSha a1b2c3d4e5f6...
```

If the commit is not found in the cache, crank will fail with an error rather than falling back.

### Different branch

```
> crank --config benchmarks.yml --scenario json --profile aspnet-perf-lin --application.channel buildcache --application.buildCacheBranch release/10.0
```

### Mixed channels (BCS runtime + pinned ASP.NET)

```
> crank --config benchmarks.yml --scenario json --profile aspnet-perf-lin --application.channel buildcache --application.aspNetCoreVersion 10.0.0-preview.3.26115.7
```

### Build Cache properties

| Property | Default | Description |
|----------|---------|-------------|
| `buildCacheRepo` | (agent `--build-cache-repo-name`, i.e. `runtime`) | Which BCS repository to resolve from: `runtime` (overlays `Microsoft.NETCore.App` from a feed install) or `aspnetcore` (places `Microsoft.AspNetCore.App` directly from the verbatim runtime-pack nupkg; fails loud if incomplete). |
| `buildCacheCommitSha` | (empty) | Specific commit SHA in the selected repo (runtime or aspnetcore). If empty, uses the latest cached build for the branch. |
| `buildCacheBranch` | `main` | Branch to query for the latest build. |
| `buildCacheConfig` | (auto-detected) | BCS configuration key (e.g., `coreclr_x64_linux` for runtime, `aspnetcore_x64_linux` for aspnetcore). Auto-detected from agent platform and the selected repo. |

### Agent configuration

The agent supports these command-line options for BCS:

| Option | Default | Description |
|--------|---------|-------------|
| `--build-cache-base-url` | `https://pvscmdupload.z22.web.core.windows.net` | Base URL for BCS blob storage. |
| `--build-cache-repo-name` | `runtime` | Repository name in BCS. |
| `--build-cache-disabled` | (not set) | Disables BCS integration on this agent. |
Loading