fix(results): split mix partitions by a committed durations file

crates/aster-srv/README.md

@@ -183,9 +183,8 @@ merged JSON.
 `GET /v1/test-results/durations?target=<address>&target=…&day=<YYYY-MM-DD>`
 returns each test file's median duration over the targets' recent runs, for
 splitting a partitioned suite (`MIX_TEST_PARTITION=K mix test --partitions N`)
-by time. Every partition of one commit asks with the same targets and the
-commit's UTC `day`, and must get the same answer, so the answer depends on
-nothing else: it covers runs received in the 7 days before `day` began (less
+by time. The answer for a set of targets and a UTC `day` depends on nothing
+else: it covers runs received in the 7 days before `day` began (less
 15 minutes for uploads still committing), and a `day` that has not begun is
 `409`. Per file, the first-attempt durations of its tests are summed in each
 of the latest 20 uncached, completely captured runs of each target, and the
@@ -201,9 +200,35 @@ empty without history.
 }
 ```
 
-Aster asks when a target runs a partition of a suite and `ASTER_SRV_URL` is
-set, then runs the files that partition gets from a longest-first split
-(`ASTER_BALANCE_PARTITIONS=off` keeps Mix's own split; so does any error).
+Partitions never ask the server while they run: one shard that could not
+reach it would split differently from the rest, and a file would run on two
+shards or none. Instead `aster results durations` asks for every partitioned
+suite in the workspace and writes `aster-durations.json` at its root, to be
+committed. A target that runs a partition of a suite reads its suite there
+and runs the files it gets from a longest-first split, so every shard of a
+commit splits the same way. Without the file or the suite in it, or with
+`ASTER_BALANCE_PARTITIONS=off`, Mix splits as usual. Files new since the
+refresh count as the suite's median. Refresh it on a schedule, for example:
+
+```yaml
+on:
+  schedule: [{ cron: "0 6 * * 1" }]
+permissions: { id-token: write, contents: write, pull-requests: write }
+jobs:
+  durations:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - run: aster results durations
+        env: { ASTER_SRV_URL: https://aster-srv.example.com }
+      - run: |
+          git diff --quiet aster-durations.json && exit 0
+          git switch -c aster-durations-$(date +%F)
+          git commit -am "chore: refresh aster-durations.json"
+          git push -u origin HEAD && gh pr create --fill
+        env: { GH_TOKEN: "${{ github.token }}" }
+```
+
 The shards' tests are recorded under the project whose suite it is (the
 directory the command `cd`s to), so a test's id does not depend on the shard
 that ran it.

docs/specs/test-tracking.html

@@ -455,8 +455,8 @@ <h1>Aster Test Tracking</h1>
       acc: ["Payload signed as X-Aster-Signature: sha256=<hmac>", "A failing endpoint never blocks ingest or classification", "Delivery log kept 30 days", "No endpoints configured means nothing is sent"],
       files: ["crates/aster-srv/src/webhooks.rs", "crates/aster-srv/src/config.rs"], deps: ["T22"] },
     { id: "T24", p: "P4", area: "client", title: "Duration-balanced mix partitions", sub: "p50 file durations, longest first",
-      obj: "A MIX_TEST_PARTITION=K mix test --partitions N target (the firstlanding shard projects) runs an explicit file list: Mix's file set, split longest-first by each file's median duration from GET /v1/test-results/durations, as of the commit's UTC day so every shard gets the same answer. Lanes stay configured selectors. Shard tests are recorded under the project that owns them (the cd ../platform target), so a test keeps its id on any shard.",
-      acc: ["Every shard of one commit gets the same split: no file runs on no shard or on two", "Falls back to Mix's own split, unchanged, with no ASTER_SRV_URL, no history, an error, or ASTER_BALANCE_PARTITIONS=off", "Retry and result capture work on the rewritten command", "Slowest partition's time drops on firstlanding (measured after rollout; logs show up to ~20% headroom)"],
+      obj: "A MIX_TEST_PARTITION=K mix test --partitions N target (the firstlanding shard projects) runs an explicit file list: Mix's file set, split longest-first by each file's median duration from aster-durations.json, committed at the workspace root and refreshed from GET /v1/test-results/durations by aster results durations (on a schedule). Shards never ask the server while running, so every shard of a commit reads the same durations. Lanes stay configured selectors. Shard tests are recorded under the project that owns them (the cd ../platform target), so a test keeps its id on any shard.",
+      acc: ["Every shard of one commit gets the same split: no file runs on no shard or on two", "Falls back to Mix's own split, unchanged, on every shard alike: no aster-durations.json, the suite missing from it, a malformed file, or ASTER_BALANCE_PARTITIONS=off", "Retry and result capture work on the rewritten command", "Slowest partition's time drops on firstlanding (measured after rollout; logs show up to ~20% headroom)"],
       files: ["src/plugins/mix_partitions.rs", "src/test_results/balance.rs", "src/executor/runner/partitions.rs", "crates/aster-srv/src/test_results/durations.rs"], deps: ["T17"] },
     { id: "T25", p: "P4", area: "docs", title: "Docs", sub: "README, GETTING_STARTED, service runbook",
       obj: "Document capture per runner, the env vars (ASTER_SRV_URL, ASTER_QUARANTINE), the quarantine rules and self-hosting.",

src/cli/commands.rs

@@ -626,6 +626,20 @@ pub enum ResultsCommands {
         #[arg(long)]
         url: Option<String>,
     },
+    /// Write aster-durations.json from aster-srv, for balanced partitions
+    ///
+    /// Finds every `MIX_TEST_PARTITION=K mix test --partitions N` suite in
+    /// the workspace and writes each test file's median duration, from the
+    /// partitions' recent runs, to aster-durations.json at the workspace
+    /// root. Commit it: each partition then splits the suite by those
+    /// durations, from the checkout, so all partitions of a commit agree.
+    /// Uses the credentials `upload` uses. Leaves the file unchanged and
+    /// exits 1 when any suite cannot be fetched.
+    Durations {
+        /// The aster-srv URL [default: $ASTER_SRV_URL]
+        #[arg(long)]
+        url: Option<String>,
+    },
 }
 
 /// Project-level subcommands

src/executor/runner.rs

@@ -37,6 +37,7 @@ use crate::test_results::{
 use crate::ui::ProgressDisplay;
 
 mod partitions;
+pub use partitions::partitioned_suites;
 mod retry;
 
 use super::command::parse_command;
@@ -802,7 +803,6 @@ impl<'a> Executor<'a> {
                 &commands,
                 &working_dir,
                 self.workspace_root,
-                project_map,
                 &|name| std::env::var(name).ok(),
             ) {
                 Some(balanced) => vec![balanced],